Rack-scale systems · B300 measured first-party · Updated July 2026
What Is the GB300 NVL72?
The NVIDIA GB300 NVL72 is a single liquid-cooled rack containing 72 B300 Blackwell Ultra GPUs and 36 Grace CPUs, wired into one NVLink domain so the whole thing behaves like one enormous GPU with 20.7TB of unified memory. It draws about 120kW and costs $3-4 million. Every figure NVIDIA publishes describes all 72 GPUs at once. We rented a single B300 and measured it across 21 workloads, so this page can tell you what one seventy-second of an NVL72 is really doing.
NVIDIA B300
288GB of HBM3e at 8 TB/s. We ran it through 21 workloads on 2026-07-12: 47.97 tok/s on Llama 3.3 70B, 9.71 it/s on FLUX.1-dev, 720p Wan video at 2.94 frames/s. The only card on our board of 102 where nothing in the core suite fails to fit.
- 47.97 tok/s on Llama 3.3 70B (measured, single stream)
- 288GB HBM3e, 12/12 core workloads fit, zero offload
- 8 TB/s bandwidth, the highest on our board
- AI Score 93.8/100, the highest score on our board
- Roughly $53,000 per GPU; no retail channel exists
- 1,400W rating, but we never measured above 1057.7W
- Single-stream LLM work runs it at 12.2-40.3% utilisation
- Measured on a CUDA 13 stack, unlike the rest of our fleet
Best for: Understanding what one B300 actually does before you spec, rent or buy anything built from them.
The GB300 NVL72 is the machine NVIDIA built when it stopped thinking of a rack as a place to put servers and started thinking of it as a single computer. Inside one 48U liquid-cooled enclosure: 18 compute trays, each carrying 4 B300 GPUs and 2 Grace CPUs, for 72 GPUs and 36 CPUs total. Nine NVLink switch trays wire every GPU to every other with 130 TB/s of aggregate bandwidth, each GPU carrying 18 fifth-generation NVLink links out over a copper backplane. The result is 20.7TB of HBM3e presented as one coherent address space. Add the Grace CPUs' LPDDR5X and NVIDIA counts up to 40TB of fast memory. It draws around 120kW and costs $3-4 million. NVIDIA's claims are rack-shaped: up to 50× the AI factory output of Hopper platforms, 1.5× dense FP4 FLOPS, 2× attention performance, real-time trillion-parameter inference. All true, all measured across 72 cooperating GPUs running batched FP4 workloads. And all useless if what you want to know is whether your model will run fast enough. So we rented one B300, the same silicon that fills those 18 trays, and measured it alone.
One of the 72. Measured LLM throughput, single stream, Q4_K_M
Peak VRAM across all 21 workloads was only 60.5GB, less than a quarter of the card's 288GB.
Why the rack exists, power draw by workload, on one B300 (1,400W rated)
1,400W rating, never reached in 21 workloads
Diffusion saturates the chip; LLM inference doesn't come close. The NVL72's fabric, pooled memory and 120kW exist to keep 72 of these fed with enough concurrent work that they stop being bandwidth-bound.
That's the finding we think matters most, and only measurement surfaces it. A B300 rated at 1,400W never drew more than 1057.7W in any of our 21 workloads. On language-model inference it averaged 286.2W to 391.6W, under 30% of its rating, while sitting at 12.2% to 40.3% utilisation. Token generation one stream at a time is bound by memory bandwidth and leaves the overwhelming majority of Blackwell Ultra's compute idle. That's not a criticism of the chip. It's the explanation for the whole rack. The NVL72 isn't fast because the GPUs are fast. It's fast because it's the only way to stop GPUs this powerful from idling. You can verify our half of that yourself for the price of a coffee.
Our verdict
The GB300 NVL72 is 72 B300 GPUs and 36 Grace CPUs in one liquid-cooled 48U rack: 20.7TB of unified HBM3e, 130 TB/s of NVLink, about 120kW, roughly $3-4M. The memory figure divides cleanly to 287.5GB per GPU. The performance figures don't divide at all. They're batched, sharded and rack-scale by design. Our measured single-GPU numbers (47.97 tok/s on Llama 3.3 70B, 9.71 it/s on FLUX.1-dev) are the floor to reason from, not a number to multiply by 72. And the reason the rack exists is visible in our power log: one B300 running one stream uses under a third of its rated watts, because nothing smaller than a rack can keep it busy.
FAQ
- How many GPUs are in a GB300 NVL72?
- 72 B300 Blackwell Ultra GPUs paired with 36 Grace CPUs, arranged as 18 compute trays each holding 4 GPUs and 2 Grace CPUs, plus 9 NVLink switch trays. All 72 sit in a single NVLink domain with 130 TB/s aggregate bandwidth, which is what lets the rack present itself to software as one very large GPU rather than 72 small ones.
- How much memory does a GB300 NVL72 have?
- 20.7TB of HBM3e across the 72 GPUs, and that figure divides cleanly: 20.7TB ÷ 72 is roughly 287.5GB per GPU, matching the B300's 288GB spec. Add the Grace CPUs' LPDDR5X (roughly 472GB each, about 17TB total) and NVIDIA quotes up to 40TB of fast memory for the rack. Worth noting: the same arithmetic on NVIDIA's DGX B300 spec gives 262GB per GPU, so NVIDIA's own published system figures aren't consistent with each other.
- How fast is one B300 GPU in a GB300 NVL72?
- We measured a single B300: 47.97 tok/s on Llama 3.3 70B, 83.68 tok/s on Qwen3 32B, 287.23 tok/s on Llama 3.1 8B (single stream, Q4_K_M), 14.6 it/s on SDXL and 9.71 it/s on FLUX.1-dev at BF16. Be careful scaling that. For 72 independent jobs, our figure × 72 is a reasonable ceiling, roughly 3454 tok/s aggregate. For one job across the whole rack, it isn't: the NVLink fabric exists precisely so the rack does something a pile of separate GPUs can't, and NVIDIA's rack-level claims come from batched, sharded FP4 workloads that look nothing like our single-stream test.
- What is the difference between B300 and GB300?
- GB300 is B300 silicon plus Grace CPUs in a rack-scale configuration. A B300 is the GPU itself, 288GB, 8 TB/s, deployed in conventional 8-GPU SXM6 nodes attached to an x86 host. GB300 refers to the same GPU deployed in NVL72 racks where 72 of them pair with 36 Grace CPUs over NVLink-C2C, sharing one coherent memory space. Same chip, radically different system around it.
- What is the difference between GB300 NVL72 and GB200 NVL72?
- Mostly memory. GB300 swaps GB200's 192GB B200 GPUs for 288GB B300s, taking total rack memory from about 13.5TB to 20.7TB. FP4 compute rises from roughly 1.44 to 2.16 exaFLOPS per rack, and Blackwell Ultra adds roughly 2× attention compute. The NVLink fabric and Grace layout are otherwise the same. If your models fit in GB200's memory, GB300's main advantage doesn't apply to you.
- How much power does a GB300 NVL72 use?
- Roughly 120kW for the rack, with reported peaks around 120.8kW under mixed-precision training. Each B300 is rated at 1,400W. But put our telemetry next to that rating: across 21 measured workloads a single B300 never exceeded 1057.7W, and language-model inference averaged 286.2-391.6W, under 30% of rating. The 120kW figure is a training-shaped number. An inference-shaped rack would draw substantially less.
- How much does a GB300 NVL72 cost?
- Industry reporting puts a full rack at $3-4 million, sold through NVIDIA's partner network and hyperscalers rather than off a shelf. Individual B300 GPUs run roughly $53,000 each. For almost everyone the realistic access path is renting B300 capacity by the hour. Which is exactly how we got the measurements on this page, for a couple of dollars.
How we test
Every number on this page describing a single B300 is our own measurement. We rented a B300 and ran the GPU Battle AI Suite v2 across 21 workloads, the core 12 that every GPU on this site runs, plus 9 Blackwell Ultra extras. LLMs run on llama.cpp (llama-bench) at Q4_K_M with -p 512 -n 128; diffusion and video run on diffusers/ComfyUI at BF16 (SDXL at FP16). Each workload gets a warmup pass plus multiple timed runs (5 for small LLMs, 3 for large models and images, 2 for video). We publish the mean as the result and the minimum as the 1% low. Run-to-run variance across our fleet is under 0.5%. Telemetry is sampled at 1 Hz from nvidia-smi for the duration of every run: power, temperature, utilisation, clocks and peak VRAM. Every wattage, temperature and tokens-per-watt figure on this page is logged draw, not a board rating. The B300 was measured on 2026-07-12 on a CUDA 13 stack (harness 2.1.0-b300-cuda13, driver 580.95.05, torch 2.13.0+cu130), which differs from the CUDA 12.8 stack the rest of our fleet runs. We label it rather than hide it. The caveat that matters most: these are single-GPU, single-stream, batch-size-1 numbers. That is the honest way to measure what one chip does, and it is deliberately not how a datacenter runs a B300. Vendor and MLPerf figures use large batches across many GPUs and will be far higher. Neither is wrong, they answer different questions. Ours answers 'what does one of these actually do'. System-level specifications (GPU counts, memory totals, rack power) come from NVIDIA's published documentation and are cited as such; we have not taken a rack apart. Where NVIDIA's own published system figures disagree with each other, we show the arithmetic rather than pick a side.