AMD's "Market Opportunity"

Last year AMD executives were talking up plans to compete in everything from servers to cell phones. Processors and GPUs for servers, desktops and notebooks would still be the bulk of the business, but the acquisition of ATI gave the company the portfolio to sell more chips for digital TVs and handhelds. These new markets would increase AMD's TAM, or total addressable market, by a combined $6.4 billion in 2007. The company has since been forced to retrench.

Last month, AMD completed the sale of its digital TV business to Broadcom, and it is seeking a buyer for its handheld division. AMD is now focused strictly on chips for servers, desktops and notebooks--including GPUs--though the total market has grown to $46.5 billion by 2009 according to yesterday's presentation. (This implies a compound annual growth rate of about 10%--the market for commercial and consumer IT was $38.3 billion in 2007--which seems plausible.) Though it makes for a less interesting product portfolio, this is a smart strategy--AMD simply doesn't have the resources to compete in all of these areas.

Before (late 2007)

After (now)

The Notebook Roadmap

The biggest changes are in AMD's notebook roadmap, where executives announced six new processors slated to appear between 2009 and 2011. Last year, the big news was Shrike, the first platform that would include a processor, code-named Swift, with both a CPU and a GPU on the same silicon die. This was set to appear sometime in 2009 on the new 45nm process technology. Now these APUs (application processor units)--for both notebooks and desktops--have been pushed all the way back to 2011 and will debut at 32nm.

AMD's current notebook platform Puma, which uses Turion X2 Ultra dual-core processors (Griffin) and ATI Radeon Mobility 3000 series graphics, is a competitive product. But it is designed strictly for mainstream notebooks, the relatively inexpensive models with 14- or 15-inch displays that you see on Best Buy shelves. AMD doesn't have a solution for ultraportables, and at the opposite extreme, it doesn't have quad-core processors for desktop replacements and mobile workstations with 17-, and now 18-inch displays. The new lineup isn't as exciting from a technology standpoint--at least not until 2011--but it will address the entire notebook market.

In the first half of 2009, AMD will release the Yukon platform, which will initially include a dual-core processor code-named Conesus with 1MB of cache. This has been billed as AMD's netbook platform, but it doesn't compete with Intel Atom at all--Conesus is a 65nm chip that will use less than 25W--and it isn't designed even for the larger netbooks with 10- to 12-inch displays. AMD has been very clear that this is a chip for ultraportable notebooks, and it will most likely show up in ultraportables with 13.3-inch widescreen displays. Later in the year, AMD will add Geneva, a dual-core CPU manufactured using the new 45nm process with twice the cache and support DDR3 memory.

If all goes smoothly, AMD will probably announce its new mainstream notebook platform, Tigris, at the Computex tradeshow next June. The first CPU (Caspian) will be a dual-core with 2MB of cache and DDR2. By 2010 AMD plans to offer its first quad-core mobile processor (Champlain) with 2MB of cache and support for DDR3 memory. Both of these CPUs will be manufactured at 45nm.

The following year the APUs will finally show up on 32nm. These will include Ontario (dual-core, 1MB cache, DDR3) for ultraportables and Liano (quad-core, 4MB cache, DDR3) for both mainstream notebooks and desktops. Liano promises to be a technological marvel. It will be interesting to see of if AMD can pack all of this and a decent GPU on a single chip--all while using IBM/The Foundry Company's new 32nm process technology and new materials (high-k dielectrics and metal gates) for the first time.

Before (late 2007)

After (now)

The Desktop Roadmap

AMD's desktop roadmap isn't quite as complicated. The big news yesterday was the announcement of the Phenom II X4. Previously known as Deneb, this AMD's first 45nm quad-core desktop chip, and it will have 8MB of cache and support DDR2 and later DDR3. It is part of the new Dragon platform for enthusiast PCs that will ship in the first quarter of next year, replacing the current high-end Spider platform with the original, star-crossed Phenom X4. In 2011, AMD is planning a 32nm high-end desktop processor, Orochi, with 'more than' four cores and 8MB of cache. This will use the Bulldozer cores that AMD has previously talked about.

There are two mainstream desktop platforms--Pisces for consumers and Kodiak for businesses. Both are slated for the second half of next year. They will use a different 45nm chip, Propos, also with four cores but with only 2MB of cache. This looks pretty similar to the old roadmap, especially for business PCs, but there's no longer any mention of triple-core (Heka or Regor) processors and the "Cartwheel" refresh for consumer desktops has disappeared. Sometime in 2011, AMD will shift to the 32nm Liano APUs for these mainstream markets. Finally, AMD is still planning to release a new platform for home theater PCs, code-named Maui, before the end of this year.

Before (late 2007)

After (now)

The Server Roadmap

AMD met the schedule it set for Shanghai, its first 45nm Opteron, which is now shipping. According to last year's roadmap, AMD planned to follow-up this chip in the second half of 2009 with a new platform that supports DDR3 memory with Montreal four- and eight-core CPUs. The new roadmap instead shows a six-core server processor, code-named Istanbul, in 2009 that runs on the current server platform as well a new Fiorano platform, which will offer more virtualization features. The first server platform that supports DDR3 memory (Maranello) shows up in 2010, along with COUs with more cores--Magny-Cours and Sao Paolo--though the exact number isn't clear yet. These code names have appeared in news reports for several months so it is likely AMD adjusted its server roadmap some time ago.

Before (late 2007)

After (now)

Foundry Co. and Process Technology

At yesterday's meeting, Doug Grose, who will serve as CEO of the Foundry Company when AMD completes the spin-off, gave some of the first details on the future foundry's plans. Foundry Co. will operate two AMD fabs in Germany, Fab 36, which currently produces AMD's CPUs, and Fab 38, which would start production sometime in 2010. AMD manufactures its 65nm, and now 45nm, CPUs using process technology jointly developed with IBM and known as SOI, for silicon-on-insulator. AMD outsources production of its GPUs, however, to several foundries including TSMC and UMC in Taiwan, and Chartered in Singapore. The Foundry Co. also plans to manufacture some of AMD's GPUs, so it has licensed a different technology, 45nm and 32nm bulk CMOS, from IBM as well. And by the end of 2011 the Foundry Co. hopes to have a new fab in upstate New York up and running, though it would most likely start commercial production at 22nm sometime in 2012.

This plan is ambitious, especially given the limited capital budget, so it's not too surprising that when you compare the new plan to last year's process technology roadmap, it looks like things have slipped. The Foundry Co. plans to produce the first test chips on 32nm SOI in early 2010, and production would start sometime in second half of the year. Intel is almost exactly a full year ahead on process technology--it introduced the first 45nm chips in November 2007 and will start production at 32nm sometime late next year--so AMD and the Foundry Co. can't afford to let 32nm production slip much further.

Before (late 2007)

After (now)

The new AMD Opterons range from 2.3- to 2.7GHz with an ACP, or Average CPU Power, rating of 75 watts (AMD uses different power ratings so you can't directly compare this with the TDP rating on Intel Xeons). In the first quarter of 2009, AMD will release additional server chips with both lower and higher power ratings, as well as a new high-end desktop platform, code-named Dragon, which will include 45nm Phenom II X4 processors paired with the existing 700 series chipset and ATI Radeon 4000 series graphics.

The key improvements in the new Opterons include higher frequencies without using more power, a 4X increase in L3 cache to 6MB and support for 800MHz DDR2 memory. AMD says this will result in performance improvements of as much as 40% over the current Opterons (aka Barcelona) on certain applications such as virtualization, databases, and Java. The new chips also include several new power management features, and AMD has posted some performance test results which it claims show that clock-for-clock the new Opterons deliver better performance per watt than Intel Xeons.

Today's announcement was expected. AMD executives have repeatedly said Shanghai was running ahead and schedule and the company is holding its annual analysts day today. In addition to its server and desktop roadmaps, AMD is expected to discuss the pending spin-off of its manufacturing operations, the Fusion processor with on-die graphics, and its strategy to address the growing netbook market.

[Read the AMD 45nm press release.]

These first Core i7s are Bloomfield processors designed for high-end desktops. There are three versions corresponding to different market segments: the 3.2GHz Core i7 965 Extreme Edition ($999), 2.93GHz Core i7 940 ($562) performance part, and the more mainstream 2.66GHz Core i7 920 ($284).

The new Core i7s are all quad-core processors (four cores on a single die) and each core can handle two threads. They also have an integrated memory controller and a new three-level cache, including an 8MB shared L3 cache. All of that adds up to a relatively hefty chip with 731 million transistors and a die size of 263 square millimeters. By comparison, a dual-core Penryn Core 2 Duo, which uses the same 45nm process technology, has 410 million transistors and measures 107 square millimeters. AMD's current quad-core Phenoms--a more accurate comparison since they have four cores and a similar architecture--are even larger at 283 square millimeters, but it is still manufactured at 65nm. AMD's 45nm Shanghai server processors--also expected later this month--will reportedly be nearly identical in size to Nehalem.

Though it is based on the same process as Penryn, the list of new features is long. Core i7s have three memory channels that support 1,066MHz DDR3; a new high-speed system bus, Quick Path Interconnect (QPI), that connects the processors to other system components (and to one another in multi-socket systems); a new socket; a new supporting chipset (the X58); more sophisticated power management features; a Turbo Mode that transparently pushes more power to the active core when running single-threaded applications; and new SSE instructions.

The results look very good. The Core i7 is the fastest desktop processor across the board on mainstream applications, multimedia tests, games and workstation applications. It looks like the second time's the charm for Hyper Threading since the performance is especially impressive on applications that can take advantage of simultaneous multi-threading such as video encoding, 3D graphics rendering, and file compression and encryption. On some tests, the sub-$300 Core i7 920 outperformed the current Penryn 3.2GHz Core 2 Extreme QX9770, which sells for $1,399. The current Penryns were already faster than AMD's fastest quad-core, the 2.6GHz Phenom X4 9950 BE, and Core i7 pads Intel's lead. Shanghai should close the gap a bit, but at this point it's hard to see how AMD can catch up anytime soon.

For now, the Core i7 and X58 are strictly for high-priced desktops, but that will soon change. Next year Intel will release server versions, currently known as Nehalem-EX, and eventually more quad-core and dual-core versions for both desktops and laptops. By late 2009, some desktop and mobile packages will also include integrated graphics processing units.

Intel Core i7 reviews coverage:

• Intel Core i7-965 Extreme Edition [CNET Reviews]
• Intel's Core i7 processors: Nehalem arrives with a splash [The Tech Report]
• Intel Core i7 Review: Nehalem Gets Real [ExtremeTech]
• Intel's Core i7 920, 945 & 965 processors [bit-tech.net]
• Intel Core i7 'Nehalem' processor and X58 chipset [The Register Hardware]
• First Look: Intel's Nehalem Smashes Performance Thresholds [ChannelWeb]
• Falcon Northwest Mach V [PC Magazine]

In a letter to SanDisk's board, Samsung Electronics CEO Yoon Woo Lee wrote that the recent quarterly results heightened concerns about SanDisk's ability to weather a protracted downturn in the memory industry and the broader economy. In addition, SanDisk had just announced a preliminary agreement with its manufacturing partner, Toshiba, to in effect sell some of its share of the NAND flash back to Toshiba in exchange for cash and relief from some other financial obligations. EE Times has posted the full text of Samsung's letter.

For its part, SanDisk said it remained open to discussions with Samsung, but only at a price that it felt reflected the value of the company's intellectual property. And it questioned whether Samsung really wanted to buy the company in the first place. (The SanDisk press release is here.) The negotiations (or lack thereof) were also complicated by a key licensing agreement between Samsung and SanDisk that is set to expire in less than a year. SanDisk earns hundreds of millions of dollars in royalties from this agreement each year.

SanDisk's quarterly results were disappointing, but it's hard to believe that they were a surprise to Samsung. The entire memory industry has been suffering through a severe downturn for some time, and the math is simple. Memory suppliers invested heavily in expanded production, which resulted in sharp oversupply that is expected to extend from 2007 through the first half of next year. As a result memory prices have fallen below production costs, even for suppliers with the most advanced technology. Initially this mainly affected the DRAM suppliers, but eventually it caught up with the NAND flash players as well.

Ultimately it seems SanDisk and Samsung were too far apart on price to even begin to tackle other formidable issues. Here is a link to my previous post on the technology that made SanDisk an attractive target for Samsung.

Nehalem is manufactured using the same 45nm process as current desktop and mobile chips, but it has a new design or microarchitecture. I've written previously about some of the new features in Nehalem.

Next year Intel will release mainstream desktop processors, as well as mobile and server versions of Nehalem. Like the Core i7, the Lynnfield desktop and Clarksfield mobile processors will use a separate GPU; the Havendale desktop and Auburndale mobile version will integrate a GPU in the processor package, thought it will also work with a separate GPU. The Lynnfield and Clarksfield processors will have four cores, while Havendale and Auburndale will be dual-core chips. The server processors are currently referred to as Nehalem-EX.

Desktops PCs with these first Core i7 chips will also have new, high-end motherboards based on the Intel X58 (or Tylersburg) chipset. Even though it has several new features such as integrated memory controller and Quick Path Interconnect, the X58 platform will still use a chipset that consists of two separate chips--like current systems that have a Northbridge and Southbridge. Next year's Ibex Peak (for Lynnfield and Havendale desktop chips) and Ibex Peak-M (for the Clarksfield and Auburndale laptop chips) chipset will in fact be a single chip--or a two-chip solution if you count the processor.

The details on the X58 platform have been slowly leaking out since the Computex tradeshow last June where motherboard makers show off their products. Several enthusiast sites such as Anandtech and Tom's Hardware have posted previews, but we've yet to see any real performance results.

AMD is set to launch its first 45nm processors (code-named Shanghai) sometime this quarter. As with Barcelona, the first versions will be server processors followed by desktop parts most likely in early 2009. The new 45nm process should help AMD reach high frequencies (and cut manufacturing costs), but there's a big gap between the current 2.6GHz Phenom X4 9550 and a possible 3.2GHz Intel Core i7-965XE, so it seems unlikely that AMD will wrest the performance crown from Intel with Shanghai.

After eight years at the helm of AMD as CEO, and more recently as chairman, Hector Ruiz will become chairman of The Foundry Company. Doug Grose, currently AMD's senior vice president of manufacturing operations, will become CEO of The Foundry Company, which will be based in Silicon Valley. Most of the 3,000 employees at the new company will come from the fabs in Dresden, though about 300 process engineers from AMD offices in Sunnyvale, Austin and East Fishkill will also join The Foundry Company. The Foundry Company will join the IBM alliance for developing both silicon-on-insulator (SOI) and bulk silicon process technology through the 22nm generation.

The investment company, Advanced Technology Investment Company (ATIC) will invest $2.1 billion to purchase its stake in The Foundry Company, including $1.4 billion invested directly in the new company and $700 million paid to AMD to purchase shares in the new company. ATIC also assumed some AMD's formidable debt and committed significant funding--up to $5.7 billion over five years--to expanding manufacturing operations. This includes completing Fab 38, an advanced 300mm wafer fab in Dresden, in 2009, followed by construction of the Luther Forest facility in Malta, New York. Later The Foundry Company will consider a new fab in Abu Dhabi depending on market conditions.

"Today's announcement significantly alters the global semiconductor industry," said ATIC Chairman Waleed Al Mokarrab. "This is great news for those communities, but most importantly it is great news for the semiconductor industry."

When the deal closes around the end of this year, AMD will own 44.4% and ATIC will own 55.6% of The Foundry Company.

As part of the announcement, Mubadala Development Company, a sovereign wealth fund of the government of the Emirate of Abu Dhabi, that currently owns 8.1% of AMD increased its stake to 19.3% through the purchase for $314 million of 58 million newly-issued AMD shares with warrants to purchase 30 million additional shares. Mubadala also gets a seat on AMD's board.

AMD had been promising this "asset-smart" plan for so long that it had become the butt of jokes among industry analysts. The announcement turns the page on a difficult period for AMD, marked by the disappointment of its Barcelona native quad-core processor, and begins a new chapter in which AMD becomes a fabless semiconductor company. AMD's President and CEO, Dirk Meyer, promised the combination of the manufacturing spin-off and Mubadala increased investment would "result in a stronger and more tightly focused AMD."

SanDisk doesn't directly manufacture the NAND flash chips that go into cards. Instead it works with Toshiba through several joint technology and manufacturing agreements. SanDisk supplies some of the capital and its flash know-how, and Toshiba operates the factories, or fabs. In exchange, SanDisk gets a steady supply of chips from Toshiba's fabs, though when business is booming it also purchases a small amount of chips from other "non-captive" sources. SanDisk then puts these chips in products such as microSD cards for cell phones, SD cards for digital cameras, and USB drives.

The main technique for reducing costs and increasing the density of flash memory is to build chips using increasingly finer features. Today advanced NAND flash is manufactured using 40nm process technology. A single chip manufactured at this node can store up to 16Gb (or 2GB) of data. Micron is working on a 34nm 32Gb chip, and the rest of the industry won't be far behind. By using an advanced lithography technique called double-patterning to print the circuits, the industry can push it down to 20nm, but that's probably the end of the road for NAND flash, at least in its current incarnation.

If you can't shrink the cells anymore, then you need to increase the amount of data that can be crammed in each cell. In 2005, SanDisk and Toshiba developed the first 8Gb chip by using MLC (multi-level cell) technology, which stores two bits of information per cell. Today the vast majority of NAND flash is MLC, though costlier SLC (single-level cell) flash is still used for applications that require higher levels of performance or better endurance. SanDisk holds the key patents on MLC-based flash, which is why Samsung already pays them hundreds of millions in royalties each year.

The next logical step is X3, or three bits per cell, and eventually X4. As it turns out SanDisk has cornered this as well. Earlier this year SanDisk announced that it had begun mass producing the industry's first X3 NAND flash, a 16Gb chip jointly developed with Toshiba and manufactured at 56nm. The X4 technology comes from SanDisk's November 2006 acquisition of M-systems. X3 isn't as efficient as physical scaling because it requires some extra peripheral circuitry to make it all work--that's why SanDisk's current 43nm 16Gb MLC is still more cost-effective. If all else is equal (process node, density and wafer size), an X3 chip costs about 20% less to produce. That might sound like a lot, but in an industry where the average selling price of the chips has been dropping 60% each year, X3 is really only a stopgap measure.

Eventually the industry will need an entirely new technology, and SanDisk believes it has the solution in 3D memory, the technology it gained through the January 2006 acquisition of Matrix Semiconductor. The concept behind 3D memory is that you can simply stack arrays of cells vertically to increase storage density rather than increasing the size (and cost) of the chip. It's like building a skyscraper rather than expanding the footprint of a building. SanDisk says it has more than 200 patents around 3D memory, and it already sells a basic form of 3D memory known as OTP (one-time programmable) that is used in applications such as Nintendo game cartridges. TSMC, a semiconductor foundry in Taiwan, manufactures 3D OTP at 80nm, and is currently working on a version using advanced 45nm process technology.

The trick is figuring how to make a version of 3D memory to which you can write, erase and rewrite data (3D R/W), like with NAND flash. Last quarter SanDisk signed a long-term agreement with Toshiba to jointly develop 3D R/W, which it says will replace NAND flash "within a decade." There is no shortage of other contenders. Toshiba, Samsung, and Spansion/Saifun are all working on a technology known as charge-trap flash (CTF). And Numonyx, the spin-off of Intel and STMicroelectronics, has developed a phase-change memory (PCM). None of these seem close to commercialization.

There is one other major area where SanDisk has technology that may be attractive to Samsung. As flash memory becomes more complex, it requires increasingly sophisticated microcontrollers to keep track of all the data being written, read and erased to the storage device. The controllers perform critical chores such as error-correction and wear-leveling to compensate for some of the physical limitations of flash. There are companies that specialize in flash controllers such as Phison and Silicon Motion, but SanDisk has its own expertise in controllers for cards, USB drives and SSDs. This expertise should be especially valuable as more SSDs based on cheaper MLC flash start to hit the market.

AMD plans to release new versions of its ATI Radeon 4000-series graphics processing units (GPUs) for entry-level graphics cards, according to the site TGDaily. The new cards are a response to Nvdia's new 9400 GT, which was announced last week, and is available in graphics cards starting at $60. These entry-level cards are the first step up from systems using chipsets with integrated graphics from Intel, as well as AMD and Nvidia.

The competition to offer better graphics at lower prices seems to affecting the business at all levels. In its recent quarterly call, Nvidia CEO Jen-Hsun Huang admitted that the company had underestimated the price-performance of AMD's high-end Radeon 4800 series, which forced Nvidia to cut prices on its competing GTX 260 and GTX 280 GPUs shortly after launching them. While these high-end GPUs get most of the glory, graphics cards costing less than $100 account for the bulk of sales, and a head-to-head battle at $60 could push average prices even lower.

Several sites have reported that AMD will release two new GPUs on September 10: the Radeon HD 4650 and Radeon HD 4670 with 512MB and 1GB GDDR3 memory, respectively. Tom's Hardware reported that these GPUs will compete directly with Nvidia 9500 GT, which is available in cards costing $70 to $100. AMD either thinks that it will offer better performance to justify a slightly higher price, or it will push the Radeon HD 4650--or a similar GPU--down to about $60 to match the 9400 GT.

It will be interesting to see how these stack up to one another, as well as to the best integrated graphics solutions such as AMD's 780G.

[Update 9/2/2008 9:00AM PST: Expreview has posted what appear to be the first performance test results for the Radeon HD 4650 and Radeon HD 4670.]

Notebooks have been coming on strong, but the Atom processor for smaller devices is also a key part of this strategy. In a session on Atom, Intel executives discussed the performance of the current Atom processor and its Menlow platform, and gave us a peek at its successor, Moorestown, scheduled to ship next year.

Menlow consists of the Atom processor and a system controller that combines the functions of the north and south bridges including a low-power 3D graphics core. Ticky Thakkar, an Intel fellow and chief architect of the ultra mobility group, showed tests results in which the 1.1GHz Atom Z510 and 1.6GHz Atom Z530 delivered more than twice the performance of the latest ARM processors, which are used in many smartphones. But Menlow is still too big, and uses too much power, for many of the gadgets that use ARM-based processors today. That's where Moorestown comes in.

A sneak peek at Moorestown For the next platform, Intel "repartitioned" the architecture to make it more of a System-on-Chip (SoC) design. The Lincroft CPU will still be manufactured using a 45nm process, but it will include the graphics processing, memory controller and video decoding and encoding on a single chip. The Langwell system controller hub will include power management, a full SSD controller and a number of new I/O features "which I'm not going to talk about today," Thakar said. Finally Moorestown includes a new class of chip, a Power Management IC or PMIC, that handles the power delivery and battery charging circuitry.

At first glance, this would seem to undermine Intel's goal of a more-integrated design. But the PMIC replaces many of the active and passive elements--what Intel calls "jelly beans"--which sit on a typical system board. This effort started with Menlow, which has about 600 of these jelly beans, compared with 1,100 in a typical laptop, according to Thakar. Reducing these jelly beans cuts cost, and makes it possible to build smaller devices.

One other note on Moorestown: The Echo Peak wireless chipset in Menlow will be replaced with an Evans Peak chipset that adds GPS to a menu that already includes WiFi, WiMax and Bluetooth. "Additionally we will also have a 3G solution in this timeframe," Thakar said, though he declined to give any details.

Moorestown is on track for 2009-2010. To prove it, during his keynote, Anand Chandrasekher, the senior vice president and general manager of Intel's ultra mobility group, showed a wafer containing the "first silicon" for Moorestown.

The rise of netbooks Currently, the Menlow platform is used three different types of gadgets: Mobile Internet Devices or MIDs, netbooks and nettops. Intel continues to show off MIDs from Lenovo, Fujitsu, Clarion and others, but the rollout of a Mobile WiMax wireless broadband--a key enabling technology--is behind schedule and MIDs, like UMPCs before them, still seem like a solution in search of a problem. Instead Atom has found a home in notebooks which have morphed from a "One Laptop Per Child" sort of product to a much broader market--a development that seems to have caught Intel by surprise.

"All of these [netbooks] are wonderful. When we envisioned the netbook, we thought it was going to be predominantly for emerging markets," said Dadi Perlmutter, Intel's executive vice president and general manger of mobile platforms, in his keynote. "But we are surprised, happily surprised, to see that this technology is also being desired in mature markets."

The dual-core Atom processor arrives Finally, Intel also announced the first dual-core Atom processor, the Atom 330, paired with a new mini-ITX format motherboard--both of which are designed specifically for a third category of Menlow products: nettops. Like their mobile counterparts, these are basic desktops designed for both emerging markets and as a second or third PC in households in mature markets. The Asus Eee Box and MSI Wind PC are two examples. (Intel says the mini-ITX board and Atom chips may also be used in Internet kiosks, thin clients and point-of-sale systems.) One of the problems with these systems is that some mini-ITX components are actually more costly than standard desktop components, so it is tough to build nettops that are much cheaper than full-fledged budget PCs, though if the nettop volumes increase, that will change.

Still it's clear that the primary market for Atom is mobile devices. During his keynote, Perlmutter showed Intel's estimates for worldwide computer shipments through 2012, which illustrate just how fast notebooks, and increasingly Atom-based mobile devices, are expected to grow--and how important they've become to Intel's strategy.

The new information on Nehalem, discussed in keynotes and presentations at the Intel Developer Forum this week, was expected. Nehalem isn't a single product, but a family of products all based on a new microarchitecture. The first products, due to ship in the fourth quarter, will be performance and "extreme" desktop chips bearing the name Core i7. Nehalem will also be used in processors with four cores for servers and workstations (Nehalem-EP), followed by mobile processors for laptops. In his keynote, senior vice president Pat Gelsinger first showed a silicon wafer of Nehalem-EX processors each with eight cores scheduled for the second half of 2009.

Intel didn't provide much in the way of performance comparisons for Nehalem, partly because it is not a single product, but also, I suspect, because it is saving the numbers for the launch of the Core i7 chips. In general terms, the goal of Nehalem was to increase system performance, while using about the same or less power than the current Core 2 and Xeon platforms. "High performance and energy efficiency are not mutually exclusive. They may sound mutually exclusive but if you innovate enough, it is possible," said Rajesh Kumar, an Intel fellow in a separate Nehalem presentation. "We want to give you a racecar that gives you the miles per gallon of an economy car."

The changes to the memory architecture in Nehalem are well-known at this point. It has a memory controller integrated on the die and uses a new interface, Quick Path Interconnect or QPI, which links the cores to one another and to the I/O hub, or Southbridge. (The Northbridge, a memory controller that was part of a separate chipset, and the Front Side Bus are headed for the Computer History Museum.) Nehalem also uses DDR3 memory, which operates at higher frequencies (1,066MHz initially) and at lower voltage, which saves power. All of these changes boost system bandwidth (by 3.4 times, according to an Intel presentation) and decrease latency, the time it takes for a processor to retrieve information from memory. AMD is quick to point out that it has had many of these features in its processors for years.

The cache also has a big impact on system latency, and this is becoming a bigger problem in systems with four or eight cores, each of which in the case of Nehalem can process two instructions simultaneously (multi-threading). To address this, Nehalem adds a third level of cache. Intel also made some lower level changes to the cache--new error correction algorithms, independent power control and the use of low-voltage 8-T SRAM in place of the usual 6-T SRAM--so that the entire system can operate at lower voltages without running into memory errors.

The power management features of Nehalem were not as well known until now. Nehalem's turbo mode can automatically turn cores on and off depending on the workload. This happens at a low level so it is invisible to the operating system and user. Intel says this will provide optimal performance with both single-threaded applications--by shifting all the power to a single core--and on highly multi-threaded applications which can take advantage of four or even eight cores with two threads each.

To make turbo mode work, Intel said it designed "new transistors and silicon technology" so that a power gate at each core can shut down power completely, and added a Power Control Unit--a separate microcontroller with more than 1 million transistors of its own--to control the gates. In a not-so-subtle knock at competitor AMD and its "asset-smart" plans, Kumar said the new power management features showed why close cooperation between process technology engineers and chip designers was so important in advanced microprocessors.

Intel also dropped some hints about how the technology in Nehalem will be used in other products, most notably processors that have both general-purpose x86 cores and GPUs on the same silicon die. For example, the QPI can be used to connect the CPUs to the GPUs and turbo mode can be extended to control the GPUs as well, so that you can have any combination of x86 cores and GPUs running depending on the applications you are using. There will be both desktop (Havendale) and mobile (Auburndale) processors with GPUs onboard; the desktop and laptop versions without integrated graphics are code-named Lynnfield and Clarksfield, respectively. These are all due in the second half of 2009.

AMD has its own version of this CPU-GPU tandem, which it refers to as Fusion, though initial versions will put the CPU and GPU in the same overall package, but not on the same physical piece of silicon.

AMD executives said the processor innovations you'll hear about next week are imitations of technology they introduced as much as five years ago, and Intel's Larrabee graphics architecture remains little more than a PowerPoint slide. For its part, AMD did not announce any new products or changes to its roadmap, but they said the company had gotten its "swagger" back, releasing a string of competitive PC platforms and graphics products this year.

This week Intel will release more details on its Nehalem architecture, the first iteration of which is now known as Core i7. Randy Allen, the senior vice president of AMD's computing division, said many innovations in Core i7 are really imitations of features such as the integrated memory controller and HyperTransport already in AMD processors.

"I guess on one level it is sort of gratifying. Imitation is the sincerest form of flattery," Allen said. "But on another level it is somewhat annoying . . . [Nehalem is] not rewriting the book, but rather imitating or photocopying our innovations."

AMD showed a series of benchmark results to demonstrate that Opteron is still the one to beat on applications such as Web servers, virtualization, databases and high-performance computing. The upcoming 45nm processor, code-named Shanghai, is already in production, and Allen said it was meeting or exceeding the company's targets for frequency, yield and schedule. AMD will ship 45nm server processors to OEMs in the fourth quarter of this year, "closely followed" by client chips.

Intel already sells 45nm processors, and it is expected to release a handful of new 45nm server and high-end desktop chips, based on the Core i7 microarchitecture, by the end of the year. But AMD thinks it will be a long time before Intel gets 2P and 4P server versions of Nehalem out the door. That could be significant because features such as the integrated memory controller in Opteron and Shanghai, as well Core i7, really shine on those types of servers. Intel declined to comment on AMD's characterization of its server plans.

Allen conceded the current Phenom desktop and Turion Ultra laptop processors can't match Intel's fastest chips, but he said AMD's PC platforms were better-positioned in the broader market. "The market has shifted away from that performance crown. It's a nice halo to have, and we certainly have plans there," he said. "But in terms of what really matters, we exceed the requirements of the value and mainstream segments."

Rick Bergman, the senior vice president of AMD's graphics group, had the easier job, essentially a victory lap after the company's new 4800 series GPUs--which he called the "first teraflops semiconductor device, period"--beat out Nvidia's GT200. AMD is also on a roll with integrated graphics having released both the M780G (Puma) for laptops and 790FX for desktops--both of which use Radeon HD 3000 series IGPs. "Our visual experience is two to three times better than our competitor," Bergman said, referring to Intel. "We have a clear performance lead over the G35 and G45."

The performance of Intel's latest integrated graphics was the subject of a dust-up last week after a video called "AMD Intel Mobile Challenge" compared the ATI Radeon HD 3200 and Intel Graphics Media Accelerator X4500 on similarly-priced HP Pavilion dv5 mainstream notebooks, and appeared to show flaws in Intel's Blu-ray HD video playback. An Intel blogger fired back, but ended up admitting that the G35/G45 has some issues that will hopefully be addressed by future driver updates. AMD has also posted YouTube video comparing game-play on the two platforms (Intel's G35, in this case).

There's no doubt that AMD's product portfolio is in much better shape than it was a year ago. But Intel continues to gain market share in microprocessors, according to IDC's figures. In the second quarter, Intel's processor shipments grew more than 4%, and nearly 21% compared with the same quarter a year ago, while AMD's shipments were flat. That pushed Intel's share of the overall market to just shy of 80%, while AMD lost more than a percentage point and slipped below 20% overall.

The Radeon 4870 X2 combines two of AMD's fastest GPU, the RV770, each clocked at 750MHz on a single graphics card for a total of 1,600 stream processors. It is also the first card to offer 1GB of GDDR5 graphics memory, which supports higher data rates than GDDR3. Asus, Diamond Multimedia, Gigabyte, MSI, Palit, Sapphire and several other board makers will offer Radeon 4870 X2 graphics cards for around $550. These should be available immediately. AMD also announced a dual-GPU version of the Radeon 4850 (625MHz GPUs, 1,600 stream processors and 2GB GDDR3 memory), which will be available in September at around $400.

The Radeon 4870 X2 competes directly with Nvidia's GeForce GTX 280. When Nvidia launched the GTX 280 and GTX 260 in June, it said the high-end cards would cost about $650, but you can now find them for significantly less.

Given the strong performance of the Radeon 4850 and 4870 (also launched in June), it is little surprise that dual-GPU versions would post impressive scores (several enthusiast sites jumped the gun and posted previews last month). The final reviews confirm that AMD has grabbed the lead in performance--at least for now. The difference is clear, especially on demanding games such as Crysis at high resolutions with high-quality settings.

The Radeon 4870 X2 and 4850 X2 can be used in CrossFireX mode (two cards), but the performance here did not scale as you would expect. In the press release, AMD states CrossFireX will deliver about three times the performance of a single Radeon 4870 graphics card on "some games," but few games really take advantage of the extra GPUs, and in some cases, a single Radeon 4870 X2 graphics card outperforms three GPUs. Eventually drivers and games will be better-optimized for CrossFireX and Nvidia's SLI, but since the Radeon 4870 X2 provides plenty of horsepower for nearly all current games, it's hard to complain too much about CrossFireX.

AMD ATI Radeon 4870 X2 reviews:

• CustomPC
• DriverHeaven.net
• ExtremeTech
• HotHardware
• Neoseeker
• PC Perspective
• The Tech Report
• TechGage
• The Guru of 3D
• Tom's Hardware
• TweakTown

After months of rumors, Taiwan-based Via Technologies has confirmed that it will stop producing chipsets for PCs that use Intel and AMD processors. A company spokesperson told the hardware site Custom PC that the third-party chipset business was on its way out, and said Via would instead focus on developing entire PC platforms.

Via is best-known for producing the chipsets and low-power processors found in small-form factor desktops. More recently, it has received a lot of press for its work on netbooks. HP uses Via's C-7 chip in its Mini-Note 2133, and at the Computex tradeshow in June, Via announced a new chip, the Nano, which competes head-to-head with Intel's Atom.

With chipsets, however, the writing was on the wall. In the second quarter, Via and its S3 Graphics joint venture combined shipped 1.1 million GPUs, or 1% of the total market; in the same quarter last year it shipped 6.26 million, representing nearly 8% of the market, according to Jon Peddie Research. Meanwhile Intel has grown its share to more than 47% of the total GPU market, followed by Nvidia and AMD.

The introduction of new graphics technology, such as SLI/CrossFire and hybrid graphics, as well as the new memory interface in Intel's upcoming Core i7, have made it more challenging for third parties to license the necessary technology and develop chipsets.

Via said it will continue to develop chipsets for its own Nano chip, following the trend set by Intel, and more recently AMD, of selling entire platforms, rather than individual PC components, which are becoming increasingly commoditized. Earlier this month the newswire DigiTimes reported that Nvidia had told motherboard makers it was preparing to exit the chipset business as well, but the company quickly denied it.

Core i7 is the name for chips that use a new microarchitecture known as Nehalem. In Intel's parlance, this is a "tock," meaning the i7 uses the same 45nm manufacturing process as the current Penryn processors (a "tick"), but it is the first major design change since the Core 2 architecture. New features include two threads per core (the return of Hyper Threading), an on-die memory controller--a feature AMD offers in Phenom desktop and Opteron server chips--and a new cache subsystem. Some enthusiast sites have already posted promising numbers on early, pre-production systems using i7 processors.

The new microarchitecure will be used in desktop processors with up eight processing cores, and eventually in mobile processors as well. Though Intel hasn't discussed specific products yet, many sites have reported that the initial Core i7 processors in production in the fourth quarter will be at 2.66GHz ($284), 2.93GHz ($562) and 3.2GHz ($999). The latter is an Extreme Edition processor; these will now be distinguished by a black Core i7 logo.

Intel said Core i7 it is the first of several new names that it will roll out over the next year.

The latest presentation has a lot of technical detail on the architecture--which is very different from the typical massively-parallel GPU--but it raises as many questions as it answers. The exact number of cores (8 to 32?), the size of the chip, how much power it will consume, and of course how it will actually perform on 3D games all remain big question marks. In his Speeds and Feeds blog, Peter N.Glaskowsky notes that a Larrabee chip with 32 1GHz cores could theoretically exceed a teraflop--around the performance of today's fastest GPUs, Nvidia's GX280 and AMD's ATI Radeon HD 4870--but it would probably be commercially impractical even for Intel. Curiously there's little mention of ray-tracing anymore. Larrabee uses the same DirectX and OpenGL APIs as ordinary GPUs to run games, but it goes about it in a very different way; Ars Technica has a nice analysis of how Larrabee renders 3D frames in software, rather than in hardware as in a GPU.

Since Larrabee-based add-in boards won't be available until early 2010, it is little surprise that product details are still sketchy, and it is too early to tell how it will really stack up to true discrete GPUs , a market completely dominated by AMD and Nvidia. Intel has failed here before, but given its vast resources, you can't count it out. In a research note, industry analyst Jon Peddie predicted that Intel will ship 46 million Larrabee "GPUs" in 2010. The total market for discrete GPUs was 350 million units last year.

More coverage of Larrabee:

• Larrabee: Intel's biggest leap since the Pentium Pro [Ars Technica]
• Intel's Larrabee--more and less than meets the eye [Speeds and Feeds]
• Intel details future 'Larrabee' graphics chip [Nanotech: The circuits blog]
• Intel's Larrabee Graphics To Take On Nvidia, ATI ... In 2010 [ChannelWeb]

These chips are not based on the Atom processor already used in netbooks and Mobile Internet Devices (MIDs)--those versions won't arrive until sometime next year. Rather the EP80579 family consists of eight chips all based on the Pentium M core, running at speeds from 600MHz to 1.2GHz. These are System-On-Chip designs (SoCs), meaning that in addition to the core processor a single chip also integrates system-level functions. In this case, that includes the memory controller, I/O controller hub and, on some chips, RISC-based accelerators for specialized tasks such as data encryption.

Initially these SoCs will be used primarily as microcontrollers for industrial applications. Intel is seeking to capitalize on what it sees as a trend toward making everything from ATM machines to VoIP controllers more like PCs with higher performance and more powerful networking capabilities. Though it gave few details, Intel said it has more than 50 customers working on 15 different designs based on its "Smart SoCs." Eventually Intel hopes to push its SoCs into a wide range of consumer electronics including set-top boxes, TVs and in-car entertainment systems.

By integrating the functions of up to four different chips on a single piece of silicon, Intel said its SoCs are 45 percent smaller and use 34 percent less power. They are still, however, too power hungry at 11 watts to 21 watts to be used in many portable devices. By comparison, the ARM-based chips found in many smartphones typically require less than 2 watts.

That is why future SoCs will be based on the 45nm Atom chip, which will not only use less power but should also cost less. The EP80579 SoCs range from $40 to $95, while the simplest Atom-based SoCs could start around $3. Atom-based SoCs will reportedly include Canmore (later this year) and Sodaville (2009) SoCs for consumer electronics, new embedded processors (2009), and Moorestown for MIDs (2009-2010). Eventually these SoCs will have multi-core processors and contain hundreds of millions of transistors. All told, the versatile Atom will cover three markets: MIDs, netbooks and embedded processing.

Intel is renewing its effort as its main competitor is moving away from consumer electronics. Last week AMD took an $876 million charge to exit the handheld and digital TV businesses, which were part of the company's $5.4 billion acquisition of ATI in 2006. AMD will release a new processor, code-named Bulldozer, for low-cost PCs, an area where higher-end, Atom-based netbooks and low-end, Celeron M-based laptops are beginning to overlap, but this is a different market.

That's not to say there aren't direct competitors in this space; there are plenty of them cranking out ARM-, MIPs- and PowerPC-based embedded processors including Broadcom, Freescale, Samsung, STMicro and Texas Instruments. And Marvell designs and sells processors for storage and networking products using the same ARM-based XScale technology from Intel's last foray in this area. But Intel believes the time is right, and that an x86-based design that is compatible with existing software and relatively easy to program will finally crack these markets.

More coverage of the EP80579 SoC family announcement:

• Intel marches once again into microcontroller market [EDN]
• Intel starts foray into SoC market [EE Times]
• Intel's Embedded Processors Destined for New Gadgets [eWeek]
• Intel Revamps Its System-on-Chip Designs [Information Week]
• Intel Welcomes Home Integrated Chips [PC Magazine]
• Intel Takes on Embedded Market With Atom Chip [PC World]
• Intel Smart SoC -- The Move to Purpose-Built Chips [Technology@Intel]
• Intel launches chips that will bring Internet to everyday devices [Venture Beat]
• Intel Brings Out Multifunction Chips In Bid to Diversify [WSJ]
• Intel chip launch targets robots, cars and TV [ZDNet UK]

Samsung has been shipping first-generation PC drives based on SLC, or single-level cell, NAND flash for some time. Dell and Alienware were the first OEMs to begin offering these 64GB drives in some laptops starting last September. But they've been slow to catch on, in part because they are so costly relative to conventional hard disk drives. That is why the shift to MLC is significant. MLC, or multi-level cell, NAND flash currently stores two bits of information in each cell, which effectively cuts the cost of the chips nearly in half. Samsung is shipping 1.8- and 2.5-inch drives in both capacities.

Samsung isn't the first to mass produce MLC-based SSDs. Toshiba announced in March that it was shipping 128GB SSDs using MLC NAND flash, initially for use in Toshiba-branded laptops. Earlier this month, OCZ announced a Core line of 32GB ($169), 64GB ($259), and 128GB ($479) SSDs built on Samsung's MLC flash chips. (Samsung hasn't announced pricing for its MLC-based drives yet.) And SanDisk and Intel have announced lower density SSDs with a PATA (parallel ATA) interface using both SLC and MLC flash and targeted at handhelds and ultra low-cost PCs.

The catch with MLC is that it has slower performance and a shorter lifespan (significantly fewer read/write cycles). There are techniques to compensate for this, but they require sophisticated controllers that are still in the early stages of development. Samsung said its latest drives can write data at 70MB per second and read data at 90MB per second, which is close to the performance of current SLC drives for PCs (SLC drives for enterprise storage applications from companies such as STEC are much faster). Samsung also claims its MLC-based SSDs will last about 20 times longer than the typical laptop.

While SSDs offer many theoretical advantages, it seems the honeymoon is over, not only because of the sky-high cost and lower density, but also because the first crop of drives have failed to live up to performance expectations. The enthusiast site Tom's Hardware recently caused a stir when it published test results showing laptop battery life with SSDs was actually shorter--not longer--than with a conventional 25-inch, 5,400rpmHDD. (The four SSDs it tested, however, were all early 32GB models with a SATA 1.5-gigabits-per-second interface that won't see any widespread adoption. Most newer PC drives--including Samsung's--use a SATA-II 3.0Gbps interface.) In an interview with ComputerWorld this week, a Fujitsu executive said SSDs and hybrid drives would remain a niche market for the next couple years. "Two years ago, a lot of our competitors were hyping these drives, Fujitsu elected not to release the product we developed in the laboratory after benchmarking it because there's no value proposition for this drive," said Joel Hagberg , Fujitsu's VP of business development.

Still Samsung remains bullish on SSDs. It expects SSD sales to grow 800% in the next couple of years and says the drives will be the flash industry's fastest growing segment. Gartner predicts unit sales will increase from 635,000 units last year to 33 million units in 2012, a compound growth rate of 117%. Even if the market grows at a slower pace, it's easy to see why SSDs are so critical given the density compared with other flash devices such as cards and USB drives. A single 128GB Samsung SSD consists of 64 16Gb NAND flash chips. And in an industry that cranks out 2.5 times more NAND flash every year, it's going to take more than a few million 16GB iPhones to use up all those bits.

[Samsung press release]

Since the launch of Barcelona last September, AMD has been promising faster clock speeds, but getting there has clearly been a challenge. To reach 2.6GHz, AMD has boosted the TDP, a measure of the maximum power draw, to 140 watts. By comparison, the 2.5GHz 9850BE has a TDP of 125 watts. HotHardware has posted some of the first numbers on the 9950BE, and the results are what you'd expect. It is slightly faster than the 9850BE, and offers similar performance to Intel's 2.4GHz Core 2 Quad Q6600, but it still doesn't match the performance of Intel's faster Core 2 Quad and Core 2 Extreme chips, which extend to 3.2GHz. Like all BE versions, the 9950 is designed for easy overclocking using AMD's OverDrive utility. HotHardware said it was able to crank the 9550BE to 3.1GHz using standard cooling.

Right now both the new 9550BE and 2.5GHz 9850BE are priced at $235, but next week AMD will cut the price of the 9850BE to $205. Eventually AMD will remove the BE designation and features from the 9850 (in fact, the processor price list already reflects this change), but for a short time you'll be able get an easily overclockable 2.5GHz quad-core for about $200. Though AMD hasn't announced it yet, the other Phenom X4s such as the 2.4GHz 9750 ($215) and 2.3GHz 9650 ($195), are likely to get price cuts as well. Still Intel's Q6600 is stiff competition at $199 with a TDP of 95 watts.

AMD announced its first Energy Efficient 65-watt Phenom X4, the 9100e, in March. The 9350e and 9150e, which are based on the newer B3 stepping of the processor, replace the 9100e. Ars Technica tested the 9350e against the 2.5GHz 9850 and, once again, the 2.4GHz Core 2 Quad Q6600. Given its relatively low clock speed, it's no surprise that the 9350e can't keep up, but it does make AMD's Spider platform (9350e, 780G chipset, and Radeon HD 4850) "quite attractive to end-users who want quad-core scaling without a higher monthly bill."

Since there are few mainstream workloads that truly take advantage of multiple cores, however, the customer for PCs based on these slower, but more power-efficient, quad-cores is less clear. AMD says it is designed for two markets: small form-factor PCs, especially for home theaters, and for relatively inexpensive business desktops. By pairing the Energy Efficient X4s with the 780G chipset, OEMs should be able to offer quad-core systems with decent integrated graphics for around $500. HP already sells some desktops with the 9100e, but no major OEMs have announced systems with the 9350e or 9150e yet.

The 9150e and 9350e are priced a $175 and $195, respectively. To make room for them, AMD lowered the prices on the triple-core processors, which previously topped out at $195. The prices now range from $125 for the 2.1GHz Phenom X3 8450 to $175 for the 2.4GHz Phenom X3 8750.

The 9950BE is likely to be the end of the line for the Phenom X4. Given its 140-watt TDP, the current Phenom X4 probably won't reach frequencies beyond 2.6GHz at acceptable manufacturing yields. And most of the company's engineers have already moved on to the 45nm Deneb desktop and Shanghai server processors, which should ship by the end of this year. This straightforward processor "shrink" will not only lower manufacturing costs, but it should allow AMD to reach higher frequencies with its quad-core chips and close the performance gap with Intel at the high-end.

• AMD Turion X2 Ultra dual-core processors Mobile 7-series chipset: AMD M780G and AMD SB700 with Radeon HD 3200 integrated graphics
• Mobile Radeon HD 3800 series GPUs
• 802.11n controllers from Atheros, Broadcom, and Ralink

Designed for mainstream laptops for both consumers and small- and medium-size businesses (SMB), Puma includes more sophisticated power management features, better 3D graphics and HD video performance, and hybrid graphics--a feature AMD was the first to introduce with the 780 desktop chipset back in March (Nvidia's new nForce780a desktop chipset also support what it calls Hybrid SLI with GeForce 8800 series GPUs).

On a broader scale, AMD says Puma is a big step toward its vision of an APU that combine the CPU and GPU on a single piece of silicon, a project that would first come to full fruition with a processor code-named Shrike sometime in 2010. It would also validate AMD's acquisition of ATI.

"There's only two companies that can generate high-performance, leading-edge microprocessors in volume. And there are only two companies in the industry that can generate leading-edge GPUs in volume," AMD President and COO Dirk Meyer said. "And we happen to be the only company that can do both."

The Turion X2 Ultra (code-named Griffin) is based on the same K8 core as existing mobile processors and is manufactured using a 65nm process. The key change is the addition of three independent "power planes," which essentially means the system can separately control the power going to each of the two cores, as well as the North Bridge, the AMD M780G, including the memory controller (DDR2-800MHz) and HyperTransport 3 bus. There are at least three Turion X2 Ultras at launch: the 2.4GHz ZM-86 with 2MB L2 cache, the 2.2GHz ZM-82 with 1MB L2 cache, and the 2.1GHz ZM-80 with 1MB L2 cache.

The Radeon 3200 is AMD's first mobile IGP that supports DirectX10 graphics and the company says it delivers 3x the performance of laptops with Intel GMA X3100 graphics. Chris Cloran, VP of the notebook division, said even if Intel's upcoming Centrino 2 chipset doubles graphics performance, AMD would still have a comfortable lead. The Radeon 3200 also includes AMD's HD video decoding technology that reduces the CPU utilization and improves video quality. During the press conference, Microsoft demonstrated a "mainstream 15.4-inch laptop" with Puma and Windows Vista easily running three HD video streams simultaneously without any hiccups.

Puma can also be used with AMD's discrete GPUs, and in addition to the previously announced HD 3400 and HD 3600 GPUs, the company introduced a new high-end mobile GPU, the Radeon HD 3800. Features of the 55nm GPU include:

• DirectX10.1
• PCI Express 2.0
• 256-bit GDDR3 memory interface
• HDMI, DVI and Display Port
• Native support for up to four monitors

The Puma platform and Radeon HD 3000 series also bring hybrid graphics to laptops. That means you can use the IGP in tandem with a discrete GPU to boost performance. AMD's tests on a Puma reference notebook showed a 1.7x performance improvement on 3DMark06 with the addition of a Radeon 3450 GPU. The platform also supports CrossFireX, which means you can use multiple discrete GPUs, though the number of customers who actually choose this option is likely to be very small.

What is likely to be very attractive to both OEMs and end-users, however, is the PowerExpress feature that lets you manually or automatically switch back and forth between discrete GPU and the IGP to conserve battery life when unplugged. AMD claims it adds 90 minutes to battery life. Fujitsu-Siemens will be the first to use PowerExpress in its AMILO Xa notebooks.

AMD also introduced one other option for getting "desktop-like" gaming on laptops without draining the battery while on the go: an external GPU-in-a-box with a PCI Express 2.0 connector. The ATI XGP uses the top-of-the-line Radeon 3800, supports CrossFireX and has HDMI output. Fujitsu will be offering it in tandem with its ultraportable AMILO Sa 3650.

The Puma platform is already shipping to OEMs, and AMD says it has more than 100 design wins from Acer, Asus, Clevo, Fujitsu, HP, MSI, NEC and Toshiba. These laptops should be available later this month. Meanwhile Intel's new mobile platform, Centrino 2, has been delayed a few weeks because of an issue with FCC certification of the 802.11n chipsets and what the company described as a relatively minor "miscorrelation" between the test environment and the actual silicon that Intel sent to its customers.

There has been a lot of grumbling on blogs and in the trade press that, graphics aside, the Puma platform won't close the overall performance gap with Santa Rosa or Centrino 2--which is very likely true even though we don't have benchmarks on real systems yet--nor does it push AMD down into ultraportable notebooks. And AMD is effectively sitting on the sidelines when it comes to even smaller systems, the netbooks that are a big theme at Computex this year and one of fastest-growing segments. But this misses the point. The goal of Puma is to bring a higher level of 3D graphics and HD video performance to affordable, mainstream and desktop replacement laptops with 13- to 17-inch displays. On that basis, AMD seems to have delivered.

More Puma coverage:

• AMD's new notebook platform [PCMag.com]
• AMD's Puma prepares to pounce [Tech Report]
• AMD lets the cat of the bag with Puma launch [Info World]
• AMD's Puma platform set to kick Centrino into touch? [Hexus.net]
• AMD's Puma ready to pounce [News.com]

Via has spent years working on Nano, which represents a major leap forward from the C7. By now, the technical details are well known. Specifically, Nano has many of the microarchitectural features--64-bit x86 instruction set, superscalar pipeline, speculative and out-of-order execution of program instructions--found in advanced Intel or AMD chips.

Manufactured by Fujitsu using 65nm process technology, Nano has 94 million transistors--roughly twice as many as the C7--and measures 63.3 square millimeters. That translates to about 1,000 chips on a single, 300mm silicon wafer. Nano is significantly larger than both the C7-M (30 square millimeters) and Intel's Atom (25 square millimeters), which both have a much simpler microarchitecture. Nano is pin-compatible with C7, which means OEMs that already sell small form-factor desktops, ultraportables or UMPCs with Via processors should be able to easily upgrade.

Because Nano offers significantly better performance than the C7, Via is going after a broader market. The L version (for Low Voltage) is designed for mainstream desktops and notebooks, and it will compete directly with Intel's low-end mobile processors such as Celeron-M and Core 2 Solo. The U version (Ultra Low Voltage) version is designed for mini-notebooks and UMPCs, and it will compete with Intel's new Atom, aka Silverthorne. All Nano processors have 1MB of L2 cache.

Model	Frequency	FSB	Max power (TDP)	Idle power
Nano U2300	1.0GHz	800MHz	5 watts	100mW
Nano U2500	1.2GHz	800MHz	6.8 watts	100mW
Nano U2400	1.3GHz	800MHz	8 watts	100mW
Nano L2200	1.6GHz	800MHz	17 watts	100mW
Nano L2100	1.8GHz	800MHz	25 watts	500mW

All of the current Nano processors are single-core, though Via has hinted that a dual-core version would be a logical next step. There has also been some speculation that Via will switch to foundry TSMC's advanced 45nm process technology for the next version of Nano.

Via's competitive advantage has been in power efficiency, and the specs for Nano look impressive. The high-end processors consume a maximum of 17 to 25 watts, and the ultra low-voltage versions top out at 5 to 8 watts, depending on the clock speed. When the chip is idle, it consumes anywhere from 0.1 watts for the 1.0GHz U2300 to 0.5 watts for the fastest 1.8GHz L2800. For comparison, here are the specs for the initial Intel Atom processors, announced last month.

Model	Frequency	FSB	Max power (TDP)	Idle power	Price
Atom Z500	800MHz	400MHz	0.65 watts	80mW	$45
Atom Z510	1.10GHz	400MHz	2 watts	100mW	$45
Atom Z520	1.33GHz	533MHz	2 watts	100mW	$65
Atom Z530	1.60GHz	533MHz	2 watts	100mW	$95
Atom Z540	1.86GHz	533MHz	2.4 watts	100mW	$160

How the Nano will actually perform against the Core 2 Solo and Atom is still unclear. Via published a white paper that shows some performance numbers, but they mainly focus on improvements over C7 on synthetic benchmarks. Via told me to expect performance that comes close to the Core 2 Solo with much lower power requirements. Perhaps we'll get a better idea at Computex next week.

Via says it has already shipped Nano processors to OEMs, and they will start to show up in systems sometime in the third quarter. The pricing hasn't been announced, though it will almost certainly be very aggressive, as Via is promising the lowest-priced systems capable of handling Blu-ray high-definition movies. Also, watch for some news from Nvidia on its work to develop GPUs, and eventually chipsets, for Nano that should result in some of the lowest-cost DirectX 10-capable systems. There are many mini-notebook and UMPC design based on the C7--including Via's own OpenBook and the HP 2133 Mini-note--so it's no surprise that Via also plans to keep it around for a while.