NVIDIA B300: power and efficiency vs. B200 and Hopper
Blackwell Ultra adds 50% more memory and roughly 1.5× the FP4 compute of the B200. Here's what that costs in power, and what it buys you compared with earlier generations.
The B300, also called Blackwell Ultra, is the second wave of NVIDIA’s Blackwell generation. It isn’t a new architecture: it’s a B200 rebuilt for the workloads that strain memory most — long reasoning chains, very long contexts, and large mixture-of-experts models. It is the most power-hungry GPU NVIDIA has shipped, and on the work it’s designed for, among the most efficient.
01 / HOW IT COMPARES
- H1H100 (2022): 80GB HBM3, 3.35 TB/s, ~700W, no FP4.
- H2H200 (2024): 141GB HBM3e, 4.8 TB/s, ~700W, no FP4.
- B2B200 (2025): 180–192GB HBM3e, 8 TB/s, ~1,000W, about 9 PFLOPS of FP4.
- B3B300 (2025–26): 288GB HBM3e, about 8 TB/s, 1,100–1,400W depending on the system, about 13–15 PFLOPS of FP4.*
One trade-off to know: the B300 cuts double-precision (FP64) performance sharply compared with the B200. That matters for traditional scientific simulation, not for AI training or inference.
02 / POWER AND EFFICIENCY
The B300 asks for 10–40% more power than a B200. In return, using the most commonly cited figures (15 PFLOPS at 1,400W), FP4 throughput per watt is about 10.7 TFLOPS, against 9.0 on the B200 — roughly 1.2× better. Across the full range of published figures it lands anywhere from about 9.3 to 13.6. Memory per watt improves more: about 0.21–0.26GB per watt, against 0.19 on the B200 and 0.11 on the H100.
Against Hopper, the gap is generational. Hopper has no FP4 path, so the fairest comparison is a model served in FP4 on a B300 versus the same model in FP8 on an H100: about 3.8× the compute per watt using the most common figures, and 3.6× the memory per GPU. Treat that as a ceiling from spec sheets; how close you get depends on whether your model holds its quality at FP4.
03 / WHEN THE EXTRA POWER PAYS OFF
The B300 earns its premium when memory is the bottleneck: reasoning models that think for thousands of tokens, retrieval over very long documents, large mixture-of-experts models, and high-concurrency inference where each user’s context takes memory. Fitting a model on fewer GPUs means less splitting across servers, and less waiting on the network between them. If your workload isn’t memory-bound, a B200 usually does the same job for less power and a lower rate — see which GPU fits your workload.
Plan for the power. At up to 1,400W per GPU, B300 racks typically need liquid cooling and far more power per rack than Hopper-era deployments; that is part of what a qualifying deployment specifies before you sign.
The B300 is the most power NVIDIA has put into one GPU. For memory-hungry AI, it turns that power into more work per watt than anything before it.
* Specs are per GPU, dense (no sparsity), compiled from public trackers and vendor write-ups (SemiAnalysis InferenceX, DataCrunch, Introl, Vast.ai, Glenn Klockwood), checked October 17, 2026, not from NVIDIA's own datasheets. Sources disagree on some figures — B200 memory is listed as 180GB or 192GB depending on the system, and B300 power from 1,100W to 1,400W. Performance-per-watt figures are our arithmetic from those spec numbers, not measured results; real efficiency depends on model, precision, batch size and cooling. Confirm exact numbers against NVIDIA's datasheets before specifying a deployment.