NVIDIA B200: power and efficiency vs. A100, H100 and H200

The B200 draws more power than any GPU before it — and does far more with each watt. Here's what changed from the A100 and Hopper generations, in plain numbers.

The B200 is the first data-center GPU of NVIDIA’s Blackwell generation, and the workhorse of most new AI deployments in 2026. The short version: it uses about 40% more power than an H100, and in return delivers more than twice the FP8 compute, more than twice the memory, and roughly 2.4× the memory bandwidth. For a business, that means more work per server, per rack and per dollar of electricity.

01 / FOUR GENERATIONS SIDE BY SIDE

A1A100 (2020): 80GB HBM2e, 2.0 TB/s, ~400W. No FP8 support — its fastest AI format is FP16/BF16 at about 312 TFLOPS.

H1H100 (2022): 80GB HBM3, 3.35 TB/s, ~700W. Introduced FP8 at about 1,979 TFLOPS.

H2H200 (2024): 141GB HBM3e, 4.8 TB/s, ~700W. The same compute as the H100, with more and faster memory.

B2B200 (2025): 180–192GB HBM3e, 8 TB/s, ~1,000W. About 4,500 TFLOPS of FP8, plus a new FP4 format at about 9,000 TFLOPS.*

02 / MORE POWER, MORE WORK PER WATT

Power draw going up is not the same as efficiency going down. Divide FP8 compute by power and the B200 delivers about 4.5 TFLOPS per watt, against about 2.8 for the H100 and H200 — roughly 1.6× the work for every watt. Memory tells the same story: about 0.19GB per watt on the B200, against 0.11 on the H100.

~1.6× FP8 per watt

4,500 TFLOPS ÷ 1,000W on the B200 vs. 1,979 TFLOPS ÷ 700W on the H100 and H200. Our arithmetic from public spec figures, not a measured benchmark.*

FP4 widens the gap further. Models served in FP4 use half the memory of FP8 and run on the B200’s fastest math path; Hopper GPUs don’t support FP4 at all. Measured results point the same way: in MLPerf Training v4.1, the B200 roughly doubled per-GPU performance on GPT-3 pre-training compared with the H200.

03 / WHAT IT MEANS FOR YOUR BUSINESS

One 8×B200 node holds about 1.5TB of GPU memory, enough to serve a 300-billion-parameter model on a single server where Hopper needed two or more. Fewer servers means fewer racks, less networking and less to operate. The catch is density: around 1,000W per GPU means a B200 node needs more power and cooling per rack than an H100 node did, which is one reason most B200 deployments land in purpose-built data centers rather than older server rooms.

For most production AI — inference, fine-tuning, training and standard-context retrieval — the B200 is the right fit. If your workloads are limited by memory rather than compute, see the NVIDIA B300 or read which GPU fits your workload.

The B200 costs more to power than the GPUs before it, and does far more with that power. For sustained AI work, efficiency per watt is the number that compounds.

* 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.