Datacenter systems · B300 measured first-party · Updated July 2026
What Is the NVIDIA DGX B300?
The NVIDIA DGX B300 is NVIDIA's pre-built AI server: eight B300 Blackwell Ultra GPUs in a 10U chassis, rated at 192 petaFLOPS for inference and 70 petaFLOPS for training. Those are system numbers for all eight GPUs at once. Which is the problem. Nobody publishes what one B300 does, because NVIDIA is selling you the box. We rented one and measured it across 21 workloads, so this page answers both questions.
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 DGX B300 is easy to describe and hard to understand. Eight B300 Blackwell Ultra GPUs, dual Intel Xeon 6776P CPUs, a 10U air-cooled chassis, twelve power supplies in N+N redundancy, NVIDIA's full software stack pre-loaded. NVIDIA rates the finished system at 192 petaFLOPS for inference and 70 petaFLOPS for training. It costs roughly $400,000 to $500,000. The hard part is that every published number describes the box, not the chip. If you want to know whether a DGX B300 will serve your model fast enough, you need to know what one B300 does, and that number doesn't exist publicly. So we rented one.
What one B300 does, LLM throughput, single stream, Q4_K_M
Measured 2026-07-12 on a rented B300. Multiply by 8 for a DGX B300's ceiling if you're running eight independent jobs. But that is not how a DGX is normally used.
Every one of those fit. That sounds unremarkable until you look at the rest of our board, where a 16GB card turns down half our suite and even a 24GB card can't load FLUX.1-dev at BF16. The B300's 288GB meant we never once hit a wall, peak VRAM across the whole 21-workload run topped out around 60.5GB, on Qwen-Image-Edit. Less than a quarter of the card. So what do you do with the per-GPU number? Two things, and it matters which.
NVIDIA's own system specs disagree on how much memory a B300 has
| System | GPUs | NVIDIA's stated memory | Implied per GPU | Matches 288GB spec? |
|---|---|---|---|---|
| GB300 NVL72 | 72 | 20.7 TB | 287.5 GB | Yes |
| DGX B300 | 8 | 2.1 TB | 262.5 GB | No, 26GB short |
| HGX B300 (Supermicro) | 8 | 2.3 TB | 287.5 GB | Yes |
We can't tell you definitively why the DGX figure is lower, the likely explanation is usable versus raw capacity after ECC and reserved memory. We flag it because if you're sizing a model against 288GB, what you actually get may be closer to 262GB.
There's one more thing worth flagging, because it changes how you should think about a DGX B300 entirely. Across all 21 workloads, the B300 never drew more than 76% of its 1,400W rating. On language models it averaged between 286.2W and 391.6W, under 30%, at 12% to 40% utilisation.
One B300's power draw vs its 1,400W rating, by workload type
1,400W board rating, never reached in any of our 21 workloads
Average draw, logged at 1 Hz. Single-stream token generation is bandwidth-bound and leaves most of Blackwell Ultra's compute idle; diffusion saturates it.
Our verdict
A DGX B300 is eight B300 GPUs, 2.1TB of pooled memory, 192 PFLOPS of rated inference and a $400,000-500,000 price tag in a 10U box. The number NVIDIA doesn't give you is the per-GPU one: on our bench a single B300 does 47.97 tok/s on Llama 3.3 70B and 9.71 it/s on FLUX.1-dev, and never fails to fit anything. Multiply by eight if you're running eight separate jobs; don't, if you're running one big one. And before you spec one: our telemetry says single-stream work uses under 30% of a B300's power budget. If your workload isn't batched, you're buying capacity you'll never touch.
FAQ
- How many GPUs are in a DGX B300?
- Eight. The DGX B300 packs 8 B300 Blackwell Ultra GPUs into a 10U air-cooled chassis alongside dual Intel Xeon 6776P CPUs. NVIDIA quotes 192 petaFLOPS of inference and 70 petaFLOPS of training, both totals across all eight GPUs, not per-GPU figures.
- How fast is one B300 in a DGX B300?
- From our measurements on 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 (all Q4_K_M, single stream), 14.6 it/s on SDXL and 9.71 it/s on FLUX.1-dev at BF16. Multiply by eight for a rough ceiling if you'd run eight independent jobs, but in production the eight GPUs are linked by NVSwitch and work on one job together, which changes the maths completely.
- Why does NVIDIA say a DGX B300 has 2.1TB of GPU memory when 8 × 288GB is 2.3TB?
- This is a real discrepancy in NVIDIA's own published figures. Eight B300s at 288GB each is 2,304GB, about 2.3TB. NVIDIA's DGX B300 spec says 2.1TB, roughly 262GB per GPU. Meanwhile NVIDIA's GB300 NVL72 spec quotes 20.7TB across 72 GPUs, which is about 287.5GB each, the full amount. The likely explanation is usable versus raw capacity after ECC and reserved memory. We flag it because if you're sizing a model against 288GB, what you get may be closer to 262GB.
- What is the difference between DGX B300 and HGX B300?
- HGX B300 is the GPU baseboard platform NVIDIA sells to server makers like Supermicro and Dell, who build it into their own systems. DGX B300 is NVIDIA's own finished server built on that platform, with NVIDIA's chosen CPUs, networking, storage and software stack pre-installed. Same silicon, different amount of assembly required.
- What is the difference between DGX B300 and DGX B200?
- The GPUs. DGX B200 uses B200 (Blackwell) at 192GB per GPU; DGX B300 uses B300 (Blackwell Ultra) at 288GB. NVIDIA quotes the B300 generation as delivering 1.5× the dense FP4 performance and 2× the attention performance. The extra 96GB per GPU is the headline. It's what lets a single B300 hold a 70B model with room to spare.
- Can I rent a DGX B300 instead of buying one?
- Yes, and for most people it's the only realistic option, a full DGX B300 runs roughly $400,000-500,000. B300 GPUs are available by the hour from cloud providers, and that's exactly how we got our measurements: we rented one for a few dollars rather than buying anything.
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.