Wi-Fi Range Is a Two-Way Conversation

The first idea to fix in your head: a wireless link only works if both ends can hear each other. An access point can be built with strong amplifiers and high-gain antennas and push its signal a long way. Your phone can’t push back. It’s running a tiny radio off a battery, held at hip height, sometimes buried in a pocket.

That’s why “long-range” routers disappoint so reliably. Out at the edge of coverage your phone still shows a bar or two, because it hears the AP fine. But pages crawl and calls drop, because the AP can’t hear the phone. The usable range of any wireless network is set by its weakest transmitter, and that’s almost always the client device. When I plan a hotel floor, I don’t design around the AP’s capabilities. I design around the oldest, cheapest phone a guest might pull out at the far end of the hallway.

The Building Is the Spec Sheet

Here’s something that surprises people: every Wi-Fi vendor works from roughly the same power budget. In the US, the FCC’s Part 15 rules for unlicensed devices cap how loud a radio is allowed to be, so nobody is selling you a legally louder transmitter. They’re selling better antennas, better software, and better marketing. Which means the variable that decides your coverage isn’t the AP you bought. It’s the structure you bolt it to.

Some rough numbers from my own installs, all with the same class of gear. Outdoors with clear line of sight, a decent AP is still usable at 300 feet or more. In a typical wood-and-drywall house, one well-placed AP covers 1,500 to 2,000 square feet without drama. In a 1960s hotel poured from concrete and block, I plan on an AP being useful about two guest rooms in either direction down the hall, and I assume nothing survives crossing two block walls. Same radio, three wildly different answers. The building is the spec sheet.

What Actually Eats Your Signal

Indoors, coverage is really a question of what sits between the AP and the device:

• Drywall and wood framing shave a little signal off per wall. Two or three interior walls are usually fine.
• Brick, block, and old plaster hit much harder. The plaster-and-lath walls in older western PA farmhouses often hide a layer of wire mesh, which behaves like a screen door for radio waves.
• Poured concrete with rebar is the hotel killer. Floors and shear walls in mid-century buildings will flatten a strong signal in one hop, which is why hospitality Wi-Fi is dense by design.
• Metal is a wall, full stop. A steel pole barn is effectively a Faraday cage. I learned early that an AP mounted outside a barn, aimed at the door, covers a sliver of the interior at best. The fix is always an AP inside the building.
• Water absorbs 2.4 and 5 GHz well. People, aquariums, and trees are mostly water. I’ve watched a farm wireless link that ran flawlessly all winter degrade every spring when the treeline leafed out.

If you want the physics underneath all of this (why signals bend, bounce, and fade the way they do), I wrote a deeper explainer on how Wi-Fi signals actually propagate.

The Band Tradeoff: Reach vs. Speed

Frequency is the other half of the story. As a rule, lower frequencies travel farther and punch through walls better; higher frequencies carry more data but give up sooner. In practice: 2.4 GHz will reach the far corner of the house and a good way across the yard, but it’s slow and crowded. 5 GHz is fast and clean but loses noticeably more to every wall. 6 GHz is fantastic in the same room as the AP and mediocre one wall away.

On the farm, that tradeoff is very visible. A 2.4 GHz signal from the house will still move data at an outbuilding 200 feet away through some trees; 5 GHz at that distance is a coin flip. If you’re not sure which band your devices should live on, I’ve covered how to pick between 2.4, 5, and 6 GHz in its own post.

More Transmit Power Isn’t More Range

The most common mistake I see in home setups (and inherit in hotel takeovers) is every radio cranked to maximum transmit power. Remember the two-way conversation: turning the AP up doesn’t make your phone any louder, so past a point you’re just extending the zone where the connection looks good and performs badly.

In multi-AP buildings, high power actively hurts. Phones cling to the distant AP they first joined instead of roaming to the closer one, and every guest at the far end of that hallway suffers for it. In hotel deployments I almost always turn 2.4 GHz power down to force clean roaming. If you run UniFi gear at home, transmit power, minimum data rates, and band steering are all covered in my guide to advanced UniFi Wi-Fi settings.

How I’d Plan Your Coverage

Here’s the practical version of everything above:

• Stop thinking in range; think in AP count. One AP per floor or per 1,500 to 2,000 square feet of framed house, placed centrally and up high, beats any single “long-range” unit shouting from a corner.
• Wire the backhaul where you can. An Ethernet run to a second AP is worth more than any antenna upgrade you’ll ever buy.
• Walk-test with a phone, not the router’s app. Aim for roughly -65 dBm in the spots that matter; by -75 you’re in dropped-video territory.
• For outbuildings, use a point-to-point bridge. A pair of $100 directional radios put solid Wi-Fi in my barn 400 feet from the house. No consumer router will manage that through 400 feet of air and metal siding.
• Keep APs out of closets, basements, and metal racks. You’d be surprised how many “weak Wi-Fi” complaints end with a router sitting behind a furnace.

None of this is exotic. Coverage is a plan, not a number you buy. Figure out what your walls are made of, count the rooms that actually have to work, and design for the cheapest phone that will ever connect. That phone is the one setting the rules.

The post Wi-Fi Range in Real Buildings: How I Plan Coverage appeared first on Blog of Cody Deluisio.

Getting there was the easy part

Anyone who has crawled out of a big municipal fireworks lot at 10:30 pm knows the tax you usually pay for a good show. Leechburg charged none of it. There was no traffic coming into town and parking was easy to find; the show started right on time and ran a solid 25 minutes. We had pizza in town earlier in the evening, so the whole night happened within a few blocks.

Room to actually watch

The part that surprised me most was the space. I stood on the bike trail near Hicks Avenue and the nearest group of people was 50 to 100 feet away. I could have moved closer to the launch area if I wanted; I was already about as close as made sense for the best view, so I stayed put. If you wanted to hang back instead, all of Market Street was open, and the bridge would make a fine seat too. There is no fighting for a patch of grass here. Families had blankets and chairs spread out with room for kids to run, and nobody was packed shoulder to shoulder.

The show itself

Twenty five minutes is longer than most small town displays run, and it did not pad the time with long gaps. Because you can stand so close, the shells break more or less overhead; that changes the experience in a way that a distant skyline show cannot match, no matter how large the budget behind it. Standing on the trail with the river behind you and the hills boxing in the sound, every report echoes off the valley walls. You feel this show as much as you see it.

If you go next year

Come into town early enough for dinner, park once, and walk. The trail near Hicks Avenue is where I would stand again; Market Street and the bridge work if you want to stay farther back. Bring chairs for the kids and do not stress about the crowd, because there is not one. Taken together, easy parking, an on time start, room to spread out, and 25 minutes of shells breaking overhead make this the show I will plan around next July.

– Cody Deluisio

The post Leechburg’s July 3rd Fireworks: The Best Show in the Valley Nobody’s Fighting Traffic For appeared first on Blog of Cody Deluisio.

The shop

Leechburg Pizza Company sits at 101 Market Street, next to the laundromat, in one of the old brick storefronts that make this end of the Kiski Valley worth walking. It is family owned, the dough is made in house, and the whole menu is chalked on boards above a wall of framed pizza jokes. Hoagies get top billing; I have not worked through that side of the board yet.

The pizza

The loaded meat Sicilian is pepperoni, sausage, ham, and bacon on a thick square crust. The bottom browned properly, the edges crisped where the cheese ran onto the pan, and the slices held their weight; the cheese pull in the last photo was not staged. The pie and the drink ran about $35, and an 18 inch square is a lot of pizza; there were leftovers for morning chores.

It was good pizza. Nothing about it is trying to be clever; it is a thick square pie with a lot of meat on it, made by people who clearly do this every day, and it tastes like it. That is the whole review.

The details

Leechburg Pizza Company
101 Market Street, Leechburg, PA 15656
(724) 236-0841
Tuesday through Sunday, 11:00 am to 8:00 pm; closed Mondays
Online ordering at leechburgpizzacompany.com

I usually write about other things, however Leechburg is a good town and this place earns its corner of it. Next trip is for the hoagies.

The post Leechburg Pizza Company Review By Cody Deluisio appeared first on Blog of Cody Deluisio.

The rule: the model explains, the engines compute

Here is the decision everything else hangs on. Every number my system produces, every EPD, every dollar value, every projected calf, every coefficient of inbreeding, comes out of a deterministic math engine. Never the language model. The model’s only two jobs are to pick which engine to run and to explain the result in plain English.

That sounds like a limitation. It is actually the feature. Because the model only ever sees numbers the engines already calculated, it physically cannot hallucinate a $C or invent a percentile. When it tells me a mating projects to a certain marbling EPD, that figure was computed, not generated. I can act on it. If I let the model do the arithmetic, I would be one confident-sounding mistake away from breeding the wrong bull to twenty cows. Wiring it so the model narrates and the math engines compute is the only version of this I trust in a chute.

Plan, execute, narrate

Every question I ask runs through the same pipeline. The model maps my request to one analysis skill plus its arguments, the deterministic engine runs, and then the model explains what came back. Ask it to rank sires for a cow and it calls the ranking engine. Ask it something that needs two steps and it chains them on its own.

The example that sold me: I asked it to rank one cow’s bull options and factor in her pedigree. Without being told how, it ran the pedigree engine first, then fed that into the sire-ranking engine, then synthesized both into one answer. That is the agentic part. I described the goal and it assembled the steps.

The heavy reasoning runs on a top model, Opus 4.8, with high effort and extended thinking, so it actually works through the straw trade-offs before it writes. Routine briefs drop down to a lighter local model. If the big model has a transient failure it retries, and can fall back rather than dying mid-plan. It reconstructs prior turns from a durable log too, so a follow-up like “now score the runner-up” resolves against the real earlier numbers instead of guessing.

What the engines actually do

The math underneath is where the real work happens:

Mating projection builds midparent EPDs and dollar values for any sire crossed on any dam. Confidence banding puts accuracy-based error bars around those projections and flags the shaky ones. And the one I lean on hardest, an inbreeding engine built on Wright’s path coefficients, computes the true coefficient of inbreeding for any pairing.

That inbreeding engine has already saved me from myself. It flagged one cross at 6.35 percent, caught another sire that shared six ancestors with the cow I was considering, and outright disqualified Coleman Easy Decision on one mating at 13.2 percent. None of those straws got ordered. There is also carrier-by-carrier recessive screening waiting on genetic-condition data, and a set of advanced scoring engines, covariance-aware scoring that accounts for correlated-trait drag, risk-adjusted scoring with confidence bands, a constraint solver, and Pareto multi-objective optimization, all now reachable from chat.

The breeding-specific smarts

The generic genetics is only half of it. The parts that make it feel like it understands my operation:

Goal tilt. I told it “mothers and bull-test growth, ignore $EN,” and it re-ranked every cow on a balanced Milk, $M, CEM, WW, YW, RADG weighting instead of just sorting by $C. It is not a single-trait maximizer. It weights toward what I actually want that year.

Tank-aware allocation. It knows my straw inventory, including which are sexed-female and how many remain, and it does a greedy allocation so my best cows claim the scarce sires first. That is how it cracked my five-straw Grand Slam problem: spend four on the top matches, hold one in reserve, send everyone else to Discipline. My handwritten chute sheet came out matching it bull for bull.

Pedigree value the dollar index ignores. It measures seedstock value that $C never captures, high-influence ancestors by descendant footprint, Pathfinder count, cow-family depth. That is what reframed 238B for me. On the dollar index she read as a low-ceiling fallback. On pedigree she read as a maternal-family cow worth breeding for a keeper daughter, so the call flipped to sexed Breakthrough. Most spreadsheets never surface that.

How a real session went

I asked for balanced mothers and growth. It ranked all eight cows on that tilt. It allocated my five scarce Grand Slam straws, four plus a reserve. The pedigree engine reframed 238B toward a sexed keeper daughter. Opus 4.8 wrote the whole plan up and emailed it to me. And along the way it independently caught two inbreeding landmines I had not spotted. When I walked out to the chute with my own handwritten sheet, the plan matched it, animal for animal.

Why I built it this way

I could have pointed a chatbot at my herd data and asked it to “recommend some matings.” Plenty of people are doing exactly that. The problem is you can never fully trust the answer, because the same system that reasons about the numbers is also free to make them up. Splitting those jobs, deterministic engines for every figure, the language model only for choosing and explaining, is what turns this from a toy into something I actually breed cows off of. The computer does the math. I make the call.

The post AI for AI: How I Run My Cattle Breeding on a Language Model That Never Touches the Numbers appeared first on Blog of Cody Deluisio.

A new spot on Market Street

You can’t miss it. The storefront is painted a soft sage green, with “Friendship Donut Co.” spelled out across the top and a little donut standing in for the O. A donut-shaped neon “Open” sign hangs in the window, there are string lights and trailing ivy inside, and a “Donuts” flag sits out by the curb in case you somehow walked right past. A couple of wrought-iron bistro tables round out the sidewalk.

A little about Leechburg

If you’re not from around here, Leechburg is a small borough in Armstrong County, sitting right along the Kiskiminetas River. Market Street is the heart of it: a walkable stretch of old brick buildings, local shops, and the kind of downtown that people have been slowly bringing back to life. An independent donut shop opening on that street is a good sign, and it fits the block. Friendship Donut Co. took an older storefront and made it feel like it had always been there. That matters in a town like Leechburg, where the good spots tend to spread by word of mouth rather than advertising.

Get there early

The practical part: they’re open Wednesday through Friday from 7 a.m. to noon, and Saturday and Sunday from 8 a.m. to 1 p.m. A dozen is $16, which is fair for what you get, and the early crowd tends to clear the case. If you want a full range of flavors, show up near opening. I normally come in at 11 and they still have a decent selection.

The flavors worth knowing

The menu runs from plain glazed and classic cake donuts up through the more creative iced ones, and it moves around depending on the day. The blueberry is really good, made with real blueberries rather than just blue coloring, so it actually tastes like the fruit instead of generic sweet. If you like a fruit donut without it turning into candy, that’s the one to try. But the one I keep going back for is the lemonade.

I almost skipped the lemonade. It’s the vanilla cake ring under bright pink frosting and rainbow sprinkles, so it looks like a birthday donut and I figured it would taste like one. The frosting is the surprise: a tart, citrusy hit that’s closer to real lemonade than I expected, and it balances the sweetness of the cake instead of piling more sugar on top. The look undersells it. If you only get one thing on your first visit, make it this.

Tips for your first visit

A few things I’ve picked up from going more than once. Go early, and go earlier on weekends, because that’s when the case empties fastest. If you’re feeding a group, the $16 dozen is the move, and it gives you room to mix flavors instead of committing to one. Grab a seat at the sidewalk tables if the weather’s decent; it’s a nice spot to sit with a coffee and watch Market Street. And if you’re bringing kids, the sprinkle donuts basically sell themselves.

Friendship Donut Co. FAQ

Where is Friendship Donut Co. located?

On Market Street in downtown Leechburg, PA, in a sage-green storefront you’ll spot by the donut-shaped neon sign and the sidewalk seating.

What are the hours?

Wednesday through Friday, 7 a.m. to noon, and Saturday and Sunday, 8 a.m. to 1 p.m. They tend to sell out before closing, so treat the early end of that window as the real one.

How much is a dozen?

A dozen is $16, which is a solid deal for freshly made cake donuts.

What’s the best donut to order?

My pick is the lemonade, with the blueberry a close second. You can’t go too wrong, but the lemonade is the one I recommend to first-timers.

Do they have anything besides donuts?

They serve coffee, drinks, breakfast sandwhiches.

Worth the early start

Everything I’ve had there has been good, from the plain glazed on up. But the lemonade is the one I think about later. If you’re anywhere near Leechburg, stop in, grab a dozen, and put a lemonade in the box. Just don’t leave it until 11.

— Cody Deluisio

The post Cody Deluisio Tackles Leechburg’s Best-Kept Secret appeared first on Blog of Cody Deluisio.

I hold Aruba CX certifications and I run AOS-CX on much smaller switches, so a 9300S with its cover off was the thing I stopped longest at on my walk through HPE Networking. The 9300S is far bigger than anything I deploy, but it runs the same operating system I already know, and getting to see what is under the lid is not something you usually get on a show floor.

Key Takeaways

• The CX 9300S is a 16 Tbps, 1U data-center switch: 32x 100GbE on QSFP28 plus 8x 400GbE on QSFP-DD, running AOS-CX.
• The 400G ports break out to 2x200G or 4x100G, and the 100G ports to 4x25G, so one box spans a 25, 100, and 400G migration.
• Inside, it is one large switching ASIC under an HPE-branded lid, plus an x86 control plane with DDR5 ECC memory, an NVMe boot SSD, and an FPGA.
• The unit on the stand was a labeled pre-production prototype marked “Not For Sale,” and the main board carried a 2023 date, so this is a mature design by show time.

Specs at a Glance

• Product: HPE Aruba Networking CX 9300S (9300S-32C8D class)
• Switching capacity: 16 Tbps in 1U
• Ports: 32x 100GbE QSFP28, plus 8x 400GbE QSFP-DD
• Breakouts: 400G to 2x200G or 4x100G; 100G to 4x25G
• Operating system: AOS-CX, with the Network Analytics Engine and Fabric Composer
• Fabric and features: VXLAN BGP-EVPN, BGP, OSPF, VRF-Lite, IPv6, RoCEv2, NVMe-oF, lossless QoS, 9K jumbo frames
• Timing: front-panel 10MHz, 1PPS, and ToD ports for precision timing distribution
• Power: dual hot-swap PSUs, 1600W 80 PLUS Platinum on the unit shown, with AC or DC options
• Cooling: six hot-swap fans, front-to-back or back-to-front airflow

The Front: Ports and Timing

The faceplate tells you what this switch is for before you ever open it.

Along with the QSFP cages, look at the small round connectors labeled 10MHz, 1PPS, and ToD. Those are precision timing ports. A switch that exposes hardware timing like this is built for fabrics where clock accuracy matters, and it is not something you see on an ordinary top-of-rack box.

Lid Off: Mostly One Giant ASIC

Open it up and the layout is honest about where the work happens.

The plate stamped with the product name is not the chip. It is a branded lid sitting on top of the main switching ASIC, with a large finned heatsink on either side. One honest limit of a show-floor look: that lid hides whose silicon is underneath, so I cannot tell you the exact ASIC from a photo. What the layout does tell you is the split. The center is the data plane, and everything around it, the fans, the power, and the small board up top, exists to feed and cool that one chip.

A Switch With ECC Memory

The control plane is where this stops looking like a switch and starts looking like a server.

That is an SK Hynix 32GB DDR5-5600 ECC SODIMM in a socketed slot, with a second slot beside it. ECC memory on the control plane is the right call for something meant to run for years without a reboot. Note the board date too. The main board is stamped December 2023, which on a 2026 show floor tells you this design has been baking for a while.

NVMe Boot, a Coin Cell, and a “Not For Sale” Sticker

A few inches over, the rest of the computer comes into view.

An ADATA industrial M.2 SSD handles boot and storage for the operating system. Next to it sits a Panasonic BR2032 coin cell, the same kind of real-time-clock battery you would find in a desktop. And the yellow label is the giveaway that this was a validation unit: it reads “Not For Sale.” This was a prototype HPE chose to open up, which is exactly why we get to see any of this.

The FPGA Doing the Quiet Work

One more chip is worth pointing out.

That is an Altera Cyclone V FPGA. On a switch like this, a part like that typically handles management-plane housekeeping: board control, sensors, sequencing, and often the timing logic that feeds those front-panel clock ports. It is the kind of chip you never think about until you are the one tracing why a board will not come up.

Power and Cooling

The back half is all about moving watts and air.

Six hot-swap fans sit in a row, each labeled for orientation, feeding the front-to-back airflow. The power supply on the unit was rated 1600W and 80 PLUS Platinum, and there are two of them for redundancy. Everything here is field-replaceable from outside the chassis, which is the part I care about most as the person who would have to service it at 2 a.m.

What I Could Not Confirm

The main switching ASIC is capped with an HPE-branded lid, so I could not read the silicon vendor or model from the unit. The control-plane CPU sits under its own heatsink and was not legible either. I saw one 32GB DDR5 ECC SODIMM and a second slot, but could not confirm the total populated memory. And while the front-panel 10MHz, 1PPS, and ToD ports are clearly there, I did not verify the exact PTP or SyncE feature support tied to them on this specific SKU. Check HPE’s QuickSpecs before relying on any of those.

Final Words

The thing that stuck with me is how ordinary the computer part is. This is a 16 Tbps switch, and the brains running it are a small x86 board with a DDR5 ECC SODIMM, an M.2 SSD, a coin-cell battery, and an FPGA. All the exotic performance lives in one big ASIC under that branded lid. Everything else is a PC that boots Linux and tells the chip what to do.

I run AOS-CX on much smaller switches, so the 9300S is bigger than anything I rack. But seeing the same operating system I know on a 16 Tbps box, and seeing that the guts are this serviceable, with hot-swap fans, hot-swap supplies, a socketed SODIMM, and an M.2, makes it feel less like a sealed appliance and more like something I could actually live with.

It was also honest about itself. A 2023 board date and a “Not For Sale” validation sticker mean this is a mature design being shown, not a mockup dressed up for the booth. That date is its own quiet signal about how long this silicon has been in the works.

So the question the 9300S left me with is the one I keep circling back to with all this gear. When the hard part is one ASIC and the rest is a serviceable little computer, how much of a modern switch is really the hardware, and how much is the software driving it? Where would you put your money?

The post Inside the HPE Aruba Networking CX 9300S: A 16 Tbps Switch With the Lid Off appeared first on Blog of Cody Deluisio.

I caught the E835 cards at the Intel Ethernet booth, a shelf apart. Up top, the E835-CQDA2 with two QSFP cages and a heavy black heatsink. Below it, the little E835-XXVDA2, two SFP28 ports and not much else. I had just photographed Intel’s E830-XXVDA4F one stand over, so the first thing I wanted to know was why Intel now sells both an E830 and an E835. More on that below.

Key Takeaways

• The E835 lineup spans a configurable 200GbE card (E835-CQDA2) and a low-cost dual-port 25GbE card (E835-XXVDA2), both on Intel’s new E835 controller.
• The CQDA2 reconfigures: one 200G port, two 100G ports, or breakouts down to 50 and 25G, set with Intel’s Ethernet Port Configuration Tool.
• Intel pitches the E835 at dense virtualized servers, claiming up to 1.9x the performance per watt of NVIDIA’s ConnectX-6 Dx and 1.4x of a comparable Broadcom card.
• Even Intel’s fastest here tops out at 200GbE, so this is mainstream server networking, not the 400 and 800G AI-fabric tier.

Specs at a Glance

E835-CQDA2

• Speeds: up to 200GbE. 1x200G on QSFP56, 2x100G on QSFP28, with breakouts to 50 and 25G
• Ports: 2x QSFP
• Controller: Intel Ethernet Controller E835
• RDMA: concurrent iWARP and RoCEv2
• Host interface: PCIe 4.0
• Design: low-power, configured with the Intel Ethernet Port Configuration Tool
• Intel claim: up to 1.9x the performance per watt of NVIDIA ConnectX-6 Dx, 1.4x of Broadcom BCM957508

E835-XXVDA2

• Speeds: 25/10GbE, 2-port SFP28
• Features: RDMA, DPDK, Dynamic Device Personalization
• Price: ServeTheHome put the recommended price of the dual 25G part at roughly $216 or under

The 200G Card: E835-CQDA2

The CQDA2 is the one Intel wants you to look at, and the heatsink shows why.

Two QSFP cages, a long black fin stack over the controller, and a card that does not commit to one port layout. With the Ethernet Port Configuration Tool, the same hardware becomes a single 200G pipe, a pair of 100G links, or a handful of 25 and 50G ports off breakout cables. For anyone who stocks server NICs, one SKU covering all of that is the actual selling point.

The Rest of the Shelf

One shelf down sat the other end of the line.

The E835-XXVDA2 is the plain one: two SFP28 ports, 25GbE, with RDMA, DPDK, and Dynamic Device Personalization to keep CPU overhead down. At a recommended price in the low couple-hundred-dollar range, it is the volume card, the one that actually ends up in racks by the pallet. The CQDA2 gets the booth lighting. The XXVDA2 gets the purchase orders.

So Why Both E830 and E835?

This is the question I walked up with, because I had just shot the E830-XXVDA4F at the next stand. Both families top out at 200GbE. The cleanest read I can give from the booth is that the E835 is the newer launch, tuned for dense virtualized servers and led with a performance-per-watt pitch, while the E830 leans harder into precision timing. Even up close, the line between the two is not obvious, and that is worth saying plainly rather than papering over.

What I Could Not Confirm

The placards gave the models and port configurations but not the host interface. I am listing PCIe 4.0 for the CQDA2 based on the E835 family and the coverage so far, so confirm the exact generation and lane width against Intel’s spec sheet. I also could not confirm whether the E835 carries the same precision timing follower features as the timing-optimized E830 parts, or pin down the precise functional split between the two families.

Final Words

The CQDA2 is the interesting one for anyone who actually orders server NICs. One card that can be a single 200G link, two 100G links, or a set of 25 and 50G breakouts means fewer part numbers to stock and a NIC that adapts to whatever the top-of-rack switch wants. That flexibility is worth more in the field than any spec-sheet record.

Intel’s headline is performance per watt, with a claim of up to 1.9x NVIDIA’s ConnectX-6 Dx. It is worth reading that carefully. The ConnectX-6 Dx is a prior-generation NVIDIA part, and NVIDIA is now shipping 400 and 800G with ConnectX-7 and 8. So the claim is real, but it is a mainstream-versus-mainstream comparison, not Intel catching the high end. For the servers most of us actually buy, mainstream is the right fight to pick.

I deploy this class of card, so the E835 lands as a practical lineup rather than an exciting one, and that is meant as a compliment. A flexible 200G card on top, a cheap 25G card underneath, and a controller with both flavors of RDMA covers the large middle of the market that never needed an AI fabric in the first place.

So the real question the E835 asks is not whether Intel can beat NVIDIA at 800G. It is whether one flexible 200G card, with a cheap 25G one beneath it, is all most of your servers ever needed. Is it all yours need?

The post Intel E835 Up Close: 200GbE, 25GbE, and One Card That Reconfigures Its Ports appeared first on Blog of Cody Deluisio.

I found it under glass at the Intel Ethernet booth, a plain green board with a big silver heatsink. After a day of 800-gigabit switch trays and liquid-cooled racks, this is the card I actually recognize from the field. Four ports of 25GbE, which is 100 gigabits total, on Intel’s E830 controller, the successor to the E810 that already sits in a lot of servers I have racked.

Key Takeaways

• The E830-XXVDA4F is a 4-port 25GbE SFP28 PCIe adapter on Intel’s E830 controller, the successor to the widely deployed E810.
• Four 25GbE ports is 100 Gbps of aggregate bandwidth, the mainstream server tier rather than the AI-fabric high end.
• Its real draw is precision timing (IEEE 1588 PTP and SyncE), RoCEv2 RDMA, and modern security with signed firmware, secure boot, and a hardware root of trust.
• This is an HPE-channel option card, sold for use in HPE ProLiant servers.

Specs at a Glance

• Product: Intel Ethernet Network Adapter E830-XXVDA4F
• Controller: Intel Ethernet Controller E830, the successor to the E810
• Ports: 4x SFP28, 25/10GbE
• Aggregate bandwidth: 100 Gbps
• Host interface: PCIe 4.0 x16
• RDMA: RoCEv2
• Timing: IEEE 1588 PTP, SyncE, precision timing follower
• Security: signed firmware, secure boot, hardware root of trust, CNSA 1.0
• Virtualization: SR-IOV, VMDq, VT-c, flexible port partitioning
• Family ceiling: the E830 controller scales up to 200GbE
• HPE-channel list price: around $7,000

The Card

The display put the board face-on, so the heatsink is the first thing you see.

That large finned heatsink sits over the E830 controller, with a second cooler over the port side of the board. For a 100-gigabit card this is a lot of passive metal, which is a fair reminder that a modern multi-port controller is a real chip doing real work, not just a set of cages on a bracket.

What I Could Not Confirm

The placard listed the model and the 4-port 25GbE configuration, but not the host interface or the power draw. I am listing PCIe 4.0 x16 based on the E830 family and the prior-generation 4-port card, so confirm the exact lane width and TDP against Intel’s spec sheet for this part before leaning on them. From this angle I also could not see the four SFP28 cages directly.

Final Words

This is the least flashy thing I photographed all show, and it is also the card I am most likely to actually install. Most servers do not need 800 gigabits. They need a few ports of 25GbE that work, and the E810 this replaces is already in a huge number of machines. Four ports of 25GbE in one slot is a clean way to give a server real bandwidth and some port redundancy without paying for a 100G or 200G part it will never saturate.

The spec that actually caught my eye is the timing. The E830 carries IEEE 1588 PTP and SyncE with a precision timing follower built in. I come from the voice and telecom side, where timing is not a footnote, it is the whole game, and seeing hardware timing this capable on a mainstream server NIC is the part of this card I find genuinely interesting.

It is worth being honest about where this sits. Intel’s networking tops out at 200GbE on the E830 while NVIDIA is shipping 400 and 800G parts, so nobody is building an AI fabric out of these. That is fine, because this is the card you buy for the unglamorous job: 25GbE, strong timing, solid security, and a name your server vendor already supports.

So while everyone photographs the 800G trays, the question worth asking is the one most buyers actually face. Does your next server need an AI fabric, or does it need four 25GbE ports that just work? For most of the racks I touch, it is the second one. Which is yours?

The post A 4-Port 25GbE NIC in an 800G World: The Intel E830-XXVDA4F Up Close appeared first on Blog of Cody Deluisio.

A few feet from the little 165-watt RTX PRO 4500 I wrote about, HPE had its big sibling on the same stand. The RTX PRO 6000 Blackwell Server Edition is the same Server Edition idea scaled all the way up. It is still passive and still display-less, fed by a single 16-pin connector. It just does all of that at 600 watts and 96GB. Behind it sat the server it is built for, the DL380a Gen12, where the “a” means accelerator-optimized.

Key Takeaways

• The RTX PRO 6000 Blackwell Server Edition is a dual-slot, passively cooled, 600W GPU with 96GB of GDDR7 ECC, the top of NVIDIA’s RTX PRO Blackwell line.
• It runs 24,064 CUDA cores with 96GB at 1.6 TB/s and supports up to four MIG partitions.
• HPE pairs it with the ProLiant Compute DL380a Gen12, a 4U server that holds up to eight of these 600W cards.
• It is the same Server Edition formula as the 165W RTX PRO 4500: no onboard fan, no display outputs, one 16-pin, and the chassis does the cooling. This one just runs it at 600 watts.

Specs at a Glance

• Product: NVIDIA RTX PRO 6000 Blackwell Server Edition
• GPU: Blackwell, 24,064 CUDA cores
• Memory: 96 GB GDDR7 ECC, 512-bit, 1.6 TB/s
• Power: up to 600 W, single 16-pin connector
• Form factor: dual-slot, full height full length, 10.5 in
• Cooling: passive, requires server airflow
• Interconnect: PCIe Gen5 x16, no NVLink
• Multi-Instance GPU: up to 4 isolated instances
• Precision and media: 5th-gen Tensor Cores with FP4, 9th-gen NVENC, 6th-gen NVDEC
• Display outputs: none, this is a server card
• Shown with: HPE ProLiant Compute DL380a Gen12, a 4U server that holds up to 8 of these GPUs

What HPE Put on the Stand

The card sat on a stand in front of an open DL380a Gen12, under an NVIDIA AI Computing by HPE banner.

Note the open server behind it. That is not decoration. A 600-watt passive card has no way to cool itself, so the host is the cooling system, and HPE built the DL380a around exactly that job.

The Card, and the Single 16-Pin

Up close, the design is the same language as the smaller Server Edition cards, just thicker.

Dual-slot, a full passive fin stack, and one 16-pin connector carrying the entire 600 watts. One cable. The connector standard exists precisely so a 600-watt card does not need a fistful of 8-pins behind it.

Built for a Specific Server

Stood on end next to its host, the pairing is the whole story.

The DL380a Gen12 is a 4U box with two of its three power domains dedicated to GPUs. It holds up to eight of these cards, which is the point. This is not a GPU you add to a server. It is a GPU you build a server around.

What I Could Not Confirm

The card was on a display stand, not seated in the DL380a, so I could not see the airflow path or how HPE staggers eight 600-watt cards across the chassis. I also could not confirm the exact power supply configuration on the unit on the floor.

Final Words

Do the math on a full server. Eight of these at 600 watts is 4.8 kilowatts of GPU in a single 4U box, before you add the two 144-core Xeons and the memory. It also adds up to 768GB of GPU memory in one server, which is the number that matters once your models get large.

I work with GPUs for AI and ML, and 96GB is the number I keep wishing I had. It is the difference between quantizing a model down to fit and just loading it. A single 96GB card runs model sizes that would otherwise need two or three smaller cards, and with MIG you can still slice it into four isolated instances when the job is smaller. This is the card I would want.

It is also the card I would have to build a room around. The 4500 was a card you slip into servers you already own. This one you buy a server for. Passive cooling at 600 watts means the DL380a’s fans and its GPU power domains are doing the heavy lifting, and eight of them in 4U is a power and heat density an ordinary server room actually feels.

So the question the 6000 asks is not whether you want 96GB. Everyone wants 96GB. It is whether you have somewhere that can power and cool a 4U box pulling close to 5 kilowatts. Do you?

The post 96GB and 600 Watts: The NVIDIA RTX PRO 6000 Blackwell Server Edition Up Close appeared first on Blog of Cody Deluisio.

I found it on a riser on the HPE stand, a plain dark slab of a card sitting in front of an open ProLiant Compute DL380 Gen12. Most of the GPUs people photograph at a show like this are 600-watt-plus parts with their own cooling towers. This one is the opposite. It is single-slot, it has no fan of its own, and it is the declared successor to the NVIDIA L4. As someone who runs GPUs for AI and ML work, this is the card on the table I would actually take home.

Key Takeaways

• The RTX PRO 4500 Blackwell Server Edition is a single-slot, passively cooled, 165W GPU with 32GB of GDDR7 ECC, positioned as the successor to the NVIDIA L4.
• It runs the GB203 Blackwell chip with 10,496 CUDA cores and supports up to two 16GB MIG partitions.
• HPE showed it alongside the ProLiant Compute DL380 Gen12, the kind of standard 2U server it is meant to drop into.
• The whole pitch is density: a card you can rack without redesigning power and cooling, not a 600W flagship.

Specs at a Glance

• Product: NVIDIA RTX PRO 4500 Blackwell Server Edition
• GPU: GB203 Blackwell, 10,496 CUDA cores, 82 RT cores
• Memory: 32 GB GDDR7 ECC, 256-bit, 800 GB/s
• Power: 165 W, single PCIe 16-pin connector
• Form factor: single-slot, full height full length, 4.4 in by 10.5 in
• Cooling: passive, requires server airflow
• Interconnect: PCIe Gen5 x16
• Multi-Instance GPU: up to 2 partitions at 16 GB
• Media and precision: 3x NVENC, 3x NVDEC, FP4 support, 5th-gen Tensor Cores
• Display outputs: none, this is a server card
• Shown with: HPE ProLiant Compute DL380 Gen12

What HPE Put on the Stand

The card sat in front of an open ProLiant DL380 Gen12, which is the detail that matters.

A passive GPU brings no cooling of its own. The fan row visible in that server is what keeps this card alive, so the GPU and the host are really one thermal decision, not two.

One 16-Pin Connector

Stand it on end and the power story is refreshingly dull.

A single 16-pin PCIe connector feeds the card, because at 165 watts there is nothing exotic to feed. No bank of 8-pins, no 600-watt cable to find room for behind the card.

Single Slot, No Display

The bracket tells you who this card is for.

Full height, single slot, vented, and no display outputs at all. The workstation version of this card uses a dual-slot blower and keeps its display ports. This one drops both so it can pack tighter and lean on the chassis for air.

What I Could Not Confirm

The card was on a display riser, not seated in the DL380, so I could not confirm how many of these HPE qualifies per DL380 Gen12, or how they sit relative to the drive bays and the fan wall once installed. I also could not see the host airflow path with a card actually in the slot.

Final Words

The number that decides everything here is 32GB. That is comfortable for inference, computer vision, data science, and fine-tuning of small and mid-size models. A 70B-class model is not fitting in 32GB without quantization or splitting it across cards. What you do get is MIG, so you can carve the card into two 16GB instances and run two models or two tenants at once. This is an inference and data-science card, and it is honest about that.

I work with GPUs for AI and ML, so this is one of the few things at this show I have a real opinion about from use rather than from a spec sheet. The spec I keep coming back to is not the core count. It is the 165 watts and the single slot. That combination is what lets you put several of these in a server you already buy, instead of redesigning a rack around one large accelerator.

That is the whole pitch. The card carries no cooling of its own, so the host’s fans do the work, which means deployability is a question about the server, not the GPU. The single 16-pin keeps the power cabling trivial. NVIDIA built a card that is deliberately boring to deploy, and for anyone who has to rack and power inference nodes, boring is the highest compliment.

So the question this card puts to you is not how fast one GPU can go. It is simpler: if your workload fits in 32GB, why pay the power and cooling bill for anything bigger?

The post 32GB and 165 Watts in a Single Slot: The NVIDIA RTX PRO 4500 Blackwell Server Edition Up Close appeared first on Blog of Cody Deluisio.