GB300 Links
A reading list. The web is finally publishing deskside GB300 coverage that is not a partner datasheet. Four pieces are worth your time now: ServeTheHome's Gigabyte W775 deep dive (the best public read on Station networking topology), StorageReview's full MSI XpertStation WS300 review, Level1Techs’ week-long hands-on with an MSI WS300 loaner, and Ahmad Osman sitting next to a DGX Station on MTS. New on October 7: Alex Ziskind and Wendell put a Station and an 8× RTX PRO 6000 box on the same agent job. Everything below is their numbers, labeled. Ours live on the GB300 topic page.
Desk is not rack. DGX Station is one Grace + one Blackwell Ultra, published as 252 GB HBM3e + 496 GB LPDDR5X = 748 GB coherent. GB300 NVL72 is a 72-GPU liquid rack. Aggregator pages that list “GB300, 288 GB HBM” are talking about the rack chip, not the box under a desk. ServeTheHome's Computex look at the Super AI Station is the cleanest public statement of why: Station SKUs enable 7 of 8 HBM3e stacks.
Start here
StorageReview WS300 is the independent review. Level1Techs’ WS300 hands-on is the serving-and-training log. MTS + Ahmad Osman at 18:19 is the Station in a room. ServeTheHome’s W775 topology page is where the CX8 wiring finally gets measured.
Spec sheet
NVIDIA's own Station page: 252 GB HBM3e at 7.1 TB/s, 496 GB LPDDR5X at 396 GB/s, 900 GB/s C2C, 1,600 W, ConnectX-8 up to 800 Gb/s.
The 748 GB trick
Coherent address space, not 748 GB of HBM. Weights that miss HBM still cross C2C and pay Grace's ~396 GB/s. StorageReview measured that path.
Our lab
Recipes and dated measurements: DGX Station GB300. Different engines and quants. Do not paste their tok/s next to ours.
Reviews and hands-on coverage
ServeTheHome: Gigabyte W775-V10-L01, video and 8-page article
OVER 2B Tokens Per Day MASSIVE NVIDIA DGX Station from Gigabyte — ServeTheHome (Patrick Kennedy), October 6, 2026, 24:39, with the companion ~10K-word article. Gigabyte loaned the unit and STH labels it sponsored. Read the article, not just the video: the video's topology walk-through (around 10:50) draws the ConnectX-8 hanging off Grace, and the article is where they show it is more interesting than that.
The finding worth bookmarking, their measurements:
- The ConnectX-8 has two hosts. Grace reaches it over what behaves like PCIe Gen5 x8 (~252 Gb/s on paper, 228.9 Gb/s measured, and they found NVIDIA documentation saying 229). The Blackwell Ultra GPU has its own PCIe Gen6 x16 link to the same CX8 (968 Gb/s). No OEM block diagram they found shows the GPU link.
- So Station networking depends on who owns the buffer. GPU-to-GPU RDMA between two Stations: 392–400 Gb/s on one 400G rail and 784.3 Gb/s on two. Grace-to-Grace: still 229.0 Gb/s with both rails up.
- The CX8 is also a PCIe Gen6 switch. Both Gen6 x4 M.2 slots, the AQC113C 10GbE, and the top USB-C hang off it. That is why the GPU can read those two M.2s at roughly full Gen6 x4 speed without crossing Grace.
- C2C 2×2 matrix (Grace or GPU compute, LPDDR5X or HBM weights) on Qwen3.8-27B, GLM-5.3-Flash, DeepSeek-V4-Flash-0731 and Qwen3.8-Flash-Next: GPU compute reading weights from LPDDR5X over C2C beat Grace compute on its own local LPDDR5X across the board. HBM still wins decode. That matches how this lab ended up using Grace: as an offload tier the GPU reads, not a second inference engine.
- Serving: 26,000+ output tok/s on Qwen3.6-35B-A3B NVFP4 at c=512 with ~24–30-token prompts and 1,024-token outputs (their ~2.25 B tokens/day headline). Patrick calls it a short-context stunt on camera, and they say they kept 10K+ tok/s with longer contexts. Nemotron3-Super-120B-A12B ~7,000 tok/s at c=512. GLM-5.3 Flash and DeepSeek V4 Flash 0731 throughput/TTFT sweeps are in the article as charts.
- Agentic coding A/B: same browser kart-game prompt. DeepSeek V4 decoded faster (~120 vs ~100 tok/s), but GLM-5.3 finished the main coding work in ~2.1 h vs ~4.7 h. Time-to-done is not tokens/s. On camera they say a 2-node GB10 cluster took ~18–20 h on comparable runs and an M3 Ultra Mac Studio 12–16 h. Those are anecdotes, not a matched harness.
- Grace CPU: 300 W cap, >3.3 GHz, ~42 ns core-to-core. On their estimated SPEC CPU2026 -O3 int rate, it scores ~2.71 per core vs ~2.58 for an EPYC 8635P.
- Power and noise: GPU idles at ~148 W with a 1.3 kW cap. Grace module idles at 77–82 W. They never saw wall idle under ~250 W. 36–38 dBA light load, 44–46 dBA full. Patrick says he tripped 15 A breakers during testing.
- Hardware oddities: leak-detection strips with a shutdown controller (first time STH has seen that in a workstation), six 12V-2×6 board power connectors, an unexplained Altera Cyclone V FPGA, and SOCAMMs under the cold plate, so effectively not serviceable.
Read with three corrections. (1) The article says the GPU–CX8 link "rivals" NVLink-C2C. It does not. Gen6 x16 is ~121 GB/s each way, and C2C is 900 GB/s total (~450 GB/s each way), so the CX8 link is about a quarter of C2C. The point that it is the fastest PCIe link in the box still holds. (2) Page 1 says the box runs on most 120 V circuits, but their own PSU section (and Patrick's tripped breakers) says it needs a dedicated 20 A circuit. Plan for the 20 A. (3) The video's topology diagram routes the CX8 through Grace only. Use the article's measured diagram.
Why it matters here. The 229 Gb/s cap applies to anything RDMA'd to or from Grace host memory. That covers host-staged NCCL, bulk weight streaming from /models, and Station→Spark expert shipping. Our remote-expert hop moves ~14 KiB activations and is latency-bound, so this cap is not its limiter. Any bandwidth-bound Station path should register GPU memory for RDMA rather than staging through Grace.
Transcript via YouTube auto-captions (English, generated), checked against the 8-page article. All numbers are ServeTheHome's, not ours, and they are on different engines and quants from our recipes.
Alex Ziskind + Wendell: a Station and an 8× RTX PRO 6000 box on the same job
I Gave Two AI Supercomputers a Real Job — Alex Ziskind, October 7, 2026, 18:35. ASUS loaned the Station. The other box is a Comino Grando with eight RTX PRO 6000s. Wendell drives Turnstone, a self-hosted, multi-node agent orchestrator with personas (orchestrator, engineer, auditor, executive, scribe) and pluggable model endpoints. The video opens with an Abacus AI ad read (~1:47–3:00).
- Same task, same model, both boxes: a Windows-on-ARM port branch of Alex's dev-bootstrap repo. Station finished in ~21 min on ~26 M input tokens; the Grando took ~37 min on ~23 M. Alex says up front that LLM runs are not one-to-one. That is one run each, with timing pulled from Turnstone's database after the fact.
- DeepSeek V4.1 Flash FP8 (~614 GB): fits entirely in the Grando's 8×96 GB of GDDR7. On the Station it has to split HBM + Grace. Single-user decode: Grando with DSpark ~307 tok/s, Station "a little" slower. Wendell's point: prefill is where HBM is untouchable.
- Multi-user: ~3,800 tok/s aggregate on one model and just over 2,000 on Nemotron 3 Super (on-screen charts, not a harness we can see). Alex says the full numbers are linked under the video.
- Pattern they endorse: a frontier cloud model as manager, eight to ten local workers doing tests and routine dev. Cost math for the one ARM job at API prices: ~$5–$36 with prompt caching, ~$24–$240 without.
- Side notes: the Station ran headless off a Jackery battery, idling ~319 W, and Alex powered it on through the BMC KVM. H3 video generation ran faster on the Station than the Grando, ~2 min per 5-second clip.
Chatting with Alex Ziskind about the Future with AI — Level1Techs, October 7, 2026, 30:05. The companion conversation, mostly not hardware. The part worth your time is ~21:40–26:30 on benchmarking agents and harnesses. Wendell argues that pass/fail hides quant damage. A 4-bit model can finish the Breakout clone, but the token count, wall time and retry loops show it struggling, and KV-cache quantization can be worse than weight quantization. A strong harness (scripted tests, linting) rescues weaker models by spending more tokens. Elsewhere: Devember this year is "offline home lab on potato hardware," AI sycophancy on the L1 forum, a SaaS-moat argument using Salesforce, the trusting-trust paper as a security frame, and open weights vs regulatory capture.
Two corrections, one tie-in. (1) Around 8:10 Wendell puts Grace LPDDR5X at "like 600 GB/s." The spec is 396 GB/s, and his own September WS300 video measured ~300–320 (above). (2) In the conversation, IBM Granite is described as the only model trained on licensed data with legal indemnity. Granite's indemnity is real, but IBM is not the only vendor offering IP indemnity. Treat the Navier–Stokes credit story and the China/US sentiment percentages as on-camera recollection, not sourced claims. Tie-in: his remark that Engram-style memory "is fine running from slower memory" is how this lab's DeepSeek V4.1 Flash lane is built: Engram tables on Grace, the hot path in HBM.
Transcripts via YouTube auto-captions (English, generated). All numbers are theirs. The Station-vs-Grando timing is one task, one run per box.
Ahmad Osman on MTS, with a Station in the studio
Why Frontier AI Labs Are Terrified of Open-Source — MTS, September 12, 2026, 39:15. Guest is Ahmad Osman (Osmant). They physically brought a DGX Station into the studio and then could not plug it in: 20 A circuit, too hot for the set. Jump to 18:19 for the hardware block.
Auto-captions, not a signed bench. Osman describes the box as a GB300 / B300 machine: 252 GB HBM3e at 7.1 TB/s plus 496 GB LPDDR5X at ~396 GB/s. He says he is running GLM-5.3 Flash and Qwen 3.8 Next Flash, and claims on-box serving on the order of thousands of tokens per second with 10–12 concurrent requests at about 90–100 tok/s per request. Treat that as a practitioner talking on camera, not a harness. The rest of the interview is the sovereignty argument: you need enough local inference to notice when a cloud API has been quantized out from under you.
Transcript via YouTube auto-captions (English, generated). Timestamp 18:19 is the hardware turn James sent.
Level1Techs: a week with the WS300
An AI Supercomputer On Your Desktop: Checking Out The MSI XpertStation WS300 — Level1Techs (Wendell), September 18, 2026, 34:50. This is the one that is closest to our lab in intent: the box on loan for about a week, running real workloads instead of a spec-sheet sweep, and saying out loud which numbers are canned and which are not. He also says the second video — point-blank NVFP4-to-BF16 kernel quality on this hardware — is still coming, so treat this as part one.
Headline numbers, his measurements:
- GLM-5.3 (753 B, frontier class), out of the box: ~1 tok/s. After the community fix — PyTorch
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, weight offload off, pinned memory, the GLM-5.2 nightly container path — ~25–28 tok/s at 32K generation, roughly 25×. That config left about 69 GB of HBM for KV, good for ~1.5 M tokens of context. He names the trade explicitly: more experts in HBM means less room for KV. - GPT-OSS-120B: ~380 tok/s single stream, over 1,000 tok/s aggregate at concurrency 4 — barely degrades, because the model is tiny next to 252 GB.
- DeepSeek V4.1 Flash: the release landed while the box was in his lab. ~230 tok/s aggregate, ~100 tok/s per stream, and he ran it with three or four concurrent users at up to 1 M context. GLM-5.3 went 1 → 35 tok/s, this one ~270 tok/s aggregate. His words: with only 252 GB of HBM3e he did not think this was possible.
- 1 M-context recall on GPT-OSS-120B: needle placed randomly, 100% recall, 40 ms to first answer, ~827 tok/s aggregate prompt processing from 750K to the full 1 M input. The honest caveat: wall-clock is dominated by prefill, so give it 740K tokens and you wait before the first token.
- Fine-tuning, not just serving: Qwen 3.8 27B dense, all 26.9 B parameters, one full epoch over ~350 M input tokens (conversations up to 142K long), NeMo Megatron Bridge/Core, BF16 weights and gradients with FP32 master weights and optimizer state, AdamW state parked in Grace LPDDR5, flash attention + per-layer activation checkpointing and CPU offload of long-context activations. ~26.5 training hours excluding checkpoint saves, ~3,673 input tok/s, all 2,291 updates with no skips. 50 GB of weights becomes ~451 GB once gradients, master weights and optimizer state are added — which is exactly what the 496 GB LPDDR5 tier is for.
Hardware notes that matter after you buy:
- Power: 1,600 W PSU, C19, 20 A circuit minimum. On 120 V it delivers ~1,300 W; a 220/240 V circuit gets you the full 1,600 W and a few percent more performance. Real numbers: prompts over 120K tokens struggled to push past ~1,100 W at the wall and stayed under 1,000 W in nvidia-smi. The training epoch is the workload that actually approaches the rated peak — instantaneous module peaks ~1.78 kW, one-second averages ~1.5 kW module / ~1.32 kW GPU.
- Add-in GPU is a budget swap, not free compute. Only Blackwell add-in cards are officially supported. When the add-in is busy it takes priority and Blackwell Ultra loses headroom; at the 1,000–1,300 W band that is about a 7% worst-case, ~3% real-world inference hit. A 300 W RTX PRO 6000 Max-Q was barely noticeable off a 240 V circuit. Also remember the GB300 has no display encoder/decoder — that is what the add-in card is for.
- C2C reality: the 900 GB/s NVLink-C2C rating is real but not the limiter. LPDDR5 tops out just under 400 GB/s and measured ~300–320 GB/s in Linux, with latency around 200 ns — higher than he expected. Same floor StorageReview measured from the other direction.
- Storage topology: two of the four M.2 slots hang off the ConnectX-8 NIC (the software view presents them as Gen 6) and two go to Grace via a RAID-1 boot pair. His framing: flash close to the GPU is a real lever for model load/cache time — GLM takes ~20 minutes from cold — but the software to exploit it is still bleeding edge.
- Video out: the Mini DisplayPort is ASPEED BMC video at VGA-equivalent, and his adapter truth table was ugly — worked with mini-DP to VGA, refused two of three mini-DP to DP cables and an active mini-DP to HDMI.
- MIG: seven instances, which he cares about for robotics and lab-automation work.
Economics, in his framing: GB300 NVL72 street price is roughly $4 M a rack, and one Station is on the order of 1/72nd of that horsepower for roughly half the per-unit cost of the rack at scale — so you are paying a premium for deskside and offline. Coding workloads only pencil out at 70–90% sustained utilization; otherwise cloud tokens are cheaper. His back-of-envelope payback is about 18 months, and he still recommended buying two and clustering over the 800 Gb/s ConnectX-8 link if one pencils out. His highest-and-best-use list: robotics (sim-to-real, low-latency loop closure), lab automation and bioinformatics where patient data cannot leave the room, and yak-shaving the Git/DevOps plumbing for people whose time is worth more than a token.
Transcript via YouTube auto-captions (English, generated). He labels his own mix of real-world and canned benchmarking, and flags the GLM numbers as fix-dependent, so read the fix as much as the tok/s. His DeepSeek V4.1 Flash caveats are worth having in mind against our own numbers: “thinking disabled” not actually short-circuiting reasoning (his 128-token cap returned nothing), counting tasks failing roughly 1 in 3, and variable tracking across a large codebase going either perfect or catastrophic with no middle ground.
StorageReview: MSI XpertStation WS300
MSI XpertStation WS300 Review: 748GB of Coherent Memory and 20 PetaFLOPS on a Desk — Divyansh Jain, August 24, 2026. This is the review to send someone. NVIDIA supplies the Grace Blackwell Ultra baseboard; MSI supplies chassis, loop, PSU, storage layout, I/O, and support. They tested remotely on an MSI-hosted unit with no RTX PRO installed, so the Superchip had the full accelerator budget.
Two caveats they printed themselves, which most recaps drop:
- Speculative-decode runs set forced acceptance equal to the draft length. Those tok/s numbers are a best-case ceiling, not live serving.
- 748 GB is one coherent address space. HBM stays the hot tier. Grace is the overflow tier. Crossing C2C is a real cost, and they measured it.
| What they measured | Their number | Why it matters |
|---|---|---|
| MAMF, dense NVFP4 / FP8 / BF16 | 6,134 / 3,909 / 1,967 TFLOPS | 4–5× a 600 W RTX PRO 6000; 17–19× a DGX Spark, stable across precisions. |
| HBM device-local read | 6,875 GB/s (~6.9 TB/s) | Within a few percent of the 7.1 TB/s rating. |
| Grace → HBM / HBM → Grace | 390 / 382 GB/s | C2C is rated 900 GB/s; the limiter they measured is Grace LPDDR at 396 GB/s. |
| Thermals under their load | GPU 71 C, CPU 65 C, GPU 1,292 W | MSI rates the loop 1,400 W across CPU+GPU. One 1,600 W PSU, C19, dedicated 20 A circuit. |
| DeepSeek V4 Flash (0731), native FP8, 512/512 | 149 tok/s C1; 1,766 tok/s C32 | Fits in HBM. Their earlier teaser tweet used the C32 figure. |
| MiniMax M2.7 NVFP4, 512/512 | 193 tok/s C1; 4,801 tok/s C128 | 125 GB in HBM. Comfortable fit. |
| MiniMax M3 NVFP4 | 223 GB HBM + 21 GB experts on Grace | “Barely fits.” Prefill-heavy falls off after C2. |
| GLM-5.2 NVFP4 (433 GB ckpt) | 218 GB HBM + 216 GB on Grace; 36→139 tok/s C1→C32 | Does not fit. Both prompt lengths track, so they call expert migration the limiter. Sweep dies at C32: no HBM left for more KV. |
| Nemotron-3-Ultra 550B NVFP4 | 114 GB on Grace; 43→168 tok/s C1→C32 | Largest model they loaded. Same Grace-bound pattern. |
| GPT-OSS-20B vs RTX PRO 6000 vs Spark, 512/512 @ C128 | 22,161 / 9,000 / 1,469 output tok/s | Shared-model comparison: Station 2.3–3× the card and 14–19× Spark at full concurrency; gap opens toward 6× the card on long prompts. |
They also note the optional RTX PRO is not free performance on top. vsloshd gives the discrete card priority; when both are busy, GB300 clocks come down. Mini DisplayPort is BMC video at 1024×768. If you want a real display, you add an RTX PRO. That is the same story ServeTheHome told from the other side of the plexiglass.
Conclusion, in their words: one or two of these next to a desk, as a development node, against cluster-reservation cost. Four of them and you should be shopping an eight-way B300 server instead.
Official
- NVIDIA DGX Station product page — 748 GB coherent, “up to 20 petaFLOPS,” models “up to 1 trillion parameters,” ConnectX-8, link two Stations. Specs table: 252 GB HBM3e | 7.1 TB/s; 496 GB LPDDR5X | 396 GB/s; NVLink-C2C 900 GB/s; FP4 Tensor Core 20 | 15 PFLOPS (sparse | dense); 1,600 W; 2× QSFP112 400 Gb/s; 4× M.2 Gen 5; MIG 7. There is now a “DGX Station for Windows” announcement on the same page.
- DGX Station Development Guide — System Overview — the short architecture note: 72-core Grace Neoverse V2, 114 MB L3, GB300 GPU “up to 252 GB HBM3e,” Ubuntu 24.04 with NVIDIA AI Developer Tools.
- DGX Station datasheet (PDF) — same numbers in vendor form.
- GTC 2025 announcement — the original Station reveal. Early copy said 784 GB coherent; shipping pages say 748 GB. If you still see 784, it is leftover launch language.
- GB300 NVL72 and DGX GB300 — the rack, not the desk. 72 Blackwell Ultra GPUs, 36 Grace CPUs, 20 TB GPU memory. Keep these tabs closed when you are shopping a Station.
OEM deskside boxes
NVIDIA sells the Superchip and the software stack. Partners sell the tower. Internals rhyme. Differentiation is cooling, storage population, display-card option, rack-kit, and who answers the phone.
| OEM | What the public web currently says | Read |
|---|---|---|
| MSI XpertStation WS300 | VideoCardz (August 31, 2026): shipping through ASI, D&H, and Newegg; Newegg list $99,999, marked out of stock when they wrote it. Guru3D had $99,900 two days earlier. StorageReview still had no published MSI list price in the August 24 review. | VideoCardz |
| ASUS ExpertCenter Pro ET900N G3 | VideoCardz (June 15, 2026): same 748 GB split, ConnectX-8, “future Windows support.” UK list £117,599.99. An earlier ASUS listing used 784 GB; the updated spec is 748. | VideoCardz |
| HP ZGX Fury AI Station | Notebookcheck (September 10, 2026): same GB300 Superchip, 496 GB LPDDR5X + 252 GB HBM3e, Red Hat collaboration, optional discrete RTX PRO in the x16 slot. They put the system “somewhere in the range of $100,000.” | Notebookcheck |
| Dell Pro Max with GB300 | TechRadar sponsored comparison (September 10, 2026): marketing numbers — 120B to 1T parameters, “up to 150 concurrent agents,” MaxCool liquid cooling, a Signal65/Futurum TCO claim versus cloud APIs. Not a hardware review. | TechRadar |
| Supermicro Super AI Station | ServeTheHome (June 27, 2026): plexiglass Computex unit. ~$125k Newegg affiliate tag on DGX Station as a class. 7-of-8 HBM stacks, 252 GB / 7.1 TB/s, four SOCAMMs, GB300 is not graphics-capable, optional RTX PRO for a real display, 5U rack kit as a differentiator, 1.6 kW not everyone wants at a desk. | ServeTheHome |
| Gigabyte W775-V10-L01 | ServeTheHome hands-on (October 6, 2026, Gigabyte loan): tower chassis, 29 kg two-person lift, CoolIT-labelled pump/reservoir, two 360 mm radiators, leak-detection trays with shutdown, top-panel I/O for floor placement. STH calls it a ~$100k system. | ServeTheHome |
| Exxact (MSI chassis) | That is the box in this lab. Hardware notes we actually ran: filling the empty CX8 M.2s, day one. | Ours |
Other coverage worth keeping
- ServeTheHome, MSI GTC 2026 booth — open-chassis photos. Linked from our M.2 post because MSI still does not publish a hardware install manual, only a BIOS PDF.
- LinuxGizmos WS300 internals — labeled internals and the user-serviceable vs not list. The origin 520'd while this page was being written; keep the URL, do not treat a Cloudflare error as a takedown.
- StorageReview Computex short — CoolIT-looking loop, copper cold plates, MSI's slot-load RTX PRO 6000 platform with QD fittings. One minute, useful before you open a side panel.
- StorageReview's own Reddit recap — same review, compressed. Useful if you only want the HBM-boundary experiment in one paragraph.
- Guru3D, August 29, 2026 — $99,900 retail note, “two units supported,” not a bench.
- MSI WS300T60L product page and BIOS user guide PDF — the official paper. No hardware install section.
Skip, or read with a hard filter. Cloud GB300 price aggregators (288 GB HBM, NVL72, $/GPU-hr) are a different product. Vendor TCO pages that convert a Station into “150 concurrent agents” or “87% lower token spend” are not measurements. Recap sites that reprint NVIDIA's 20 PFLOPS without saying sparse FP4 are doing brochure math. NationalPC-style pages that rate the WS300 9.2/10 from the spec sheet are not reviews.
How to read any of this against our lab
StorageReview's GLM-5.2 offload numbers and Osman's on-camera GLM-5.3 Flash claims are not substitutes for a recipe. Different checkpoints, quants, engines, acceptance cheats, and concurrency. If you want what one Station in this house actually served, with the configs that lost: DGX Station GB300.
The useful public fact, independent of whose tok/s you trust: the deskside SKU is a 252 GB HBM machine with a large, slower Grace tier attached by C2C. Everything interesting about serving on it is which side of that 252 GB line you are on.
Related here
- DGX Station GB300 — recipes and dated measurements from this lab.
- Filling the empty M.2s — CX8 slots, RAID0
/models, nevermdadmbynvmeN. - Day one — arrival, kernel trap, first token.
- Local LLM stack — where the Station sits relative to the Spark pair.