FP4 arrives with Blackwell
NVIDIA lists FP8 as the lowest precision for H200 and no FP4 figure. Eight B200 GPUs reach 144 PFLOPS of FP4 with sparsity, 72 dense; eight B300 GPUs 144 sparse, 108 dense.
| NVIDIA figure | HGX B200 | HGX B300 |
|---|---|---|
| GPUs | 8 × NVIDIA Blackwell SXM | 8 × NVIDIA Blackwell Ultra SXM |
| GPU memory, total | 1.4 TB | 2.1 TB |
| FP4 Tensor Core | 144 PFLOPS sparse, 72 dense | 144 PFLOPS sparse, 108 dense |
| FP8 / FP6 Tensor Core | 72 PFLOPS sparse | 72 PFLOPS sparse |
| FP16 / BF16 Tensor Core | 36 PFLOPS sparse | 36 PFLOPS sparse |
| TF32 Tensor Core | 18 PFLOPS sparse | 18 PFLOPS sparse |
| INT8 Tensor Core | 72 POPS sparse | 3 POPS sparse |
| FP32 | 600 TFLOPS | 600 TFLOPS |
| FP64 / FP64 Tensor Core | 296 TFLOPS | 10 TFLOPS |
| NVLink | 5th gen, 1.8 TB/s per GPU, 14.4 TB/s total | 5th gen, 1.8 TB/s per GPU, 14.4 TB/s total |
| Networking bandwidth | 0.8 TB/s | 1.6 TB/s |
| Attention performance | 1x | 2x |
Source: NVIDIA's HGX page. FP4 is shown sparse and dense; the other tensor rows are sparse, and NVIDIA states dense is half the sparse figure. Attention performance is relative to Blackwell.
| GPU memory | 141 GB HBM3e |
|---|---|
| Memory bandwidth | 4.8 TB/s |
| FP4 Tensor Core | Not listed by NVIDIA |
| FP8 Tensor Core | 3,958 TFLOPS sparse |
| FP16 / BF16 Tensor Core | 1,979 TFLOPS sparse |
| TF32 Tensor Core | 989 TFLOPS sparse |
| INT8 Tensor Core | 3,958 TOPS sparse |
| FP64 / FP64 Tensor Core | 34 TFLOPS / 67 TFLOPS |
| NVLink | 900 GB/s per GPU |
| Power | Up to 700 W per GPU, configurable |
Source: NVIDIA's H200 page, H200 SXM. NVIDIA publishes H200 per GPU rather than as an eight-GPU system.
NVIDIA lists FP8 as the lowest precision for H200 and no FP4 figure. Eight B200 GPUs reach 144 PFLOPS of FP4 with sparsity, 72 dense; eight B300 GPUs 144 sparse, 108 dense.
141 GB per H200 GPU, 1.4 TB across eight B200 GPUs and 2.1 TB across eight B300 GPUs. Eight H200 GPUs hold 8 × 141 GB = 1,128 GB (our arithmetic on NVIDIA's per-GPU figure).
900 GB/s per H200 GPU, 1.8 TB/s per Blackwell and Blackwell Ultra GPU, 14.4 TB/s across an eight-GPU board.
On NVIDIA's HGX table, B300 raises dense FP4 from 72 to 108 PFLOPS, doubles attention performance and doubles networking. FP64 falls from 296 to 10 TFLOPS and INT8 from 72 to 3 POPS.
On NVIDIA's HGX figures, B300 keeps B200's FP8, FP16 and NVLink, raises dense FP4 from 72 to 108 PFLOPS across eight GPUs, and doubles attention performance and networking. It holds 2.1 TB of GPU memory against 1.4 TB. B200 keeps far more FP64 (296 TFLOPS against 10) and INT8 (72 POPS against 3).
B200 adds FP4, which NVIDIA does not list for H200, and doubles NVLink to 1.8 TB/s per GPU. Eight B200 GPUs hold 1.4 TB of memory; eight H200 GPUs hold 1,128 GB (our arithmetic on NVIDIA's 141 GB per GPU).
B300 adds FP4, at 144 PFLOPS with sparsity and 108 dense across eight GPUs, doubles NVLink per GPU and nearly doubles the memory: 2.1 TB against 1,128 GB for eight H200 GPUs (our arithmetic). H200 lists 34 TFLOPS of FP64 per GPU; NVIDIA's HGX B300 figure is 10 TFLOPS for all eight GPUs.
B300: 2.1 TB of GPU memory and twice the attention performance of B200 (NVIDIA figures).
B200: 296 TFLOPS of FP64 against B300's 10 (NVIDIA figures), so simulation and other double-precision work stay practical.
B200: 72 POPS of INT8 against B300's 3 (NVIDIA figures, sparse).
Models served in FP8 or FP16 keep running. Moving to Blackwell adds FP4 and doubles NVLink bandwidth per GPU.
Tell us the models, context lengths and precision you plan to run. We recommend a platform in writing, in MYR.