Skip to content
EMARQUE AI

H200 vs B200 vs B300

Three generations of NVIDIA's data-centre GPU: Hopper (H200), Blackwell (B200) and Blackwell Ultra (B300). Every figure here is NVIDIA's own, with sparse and dense marked where NVIDIA marks them.
Published . Figures from NVIDIA's HGX and H200 pages.
Eight-GPU platforms

HGX B200 and HGX B300, side by side.

NVIDIA figureHGX B200HGX B300
GPUs8 × NVIDIA Blackwell SXM8 × NVIDIA Blackwell Ultra SXM
GPU memory, total1.4 TB2.1 TB
FP4 Tensor Core144 PFLOPS sparse, 72 dense144 PFLOPS sparse, 108 dense
FP8 / FP6 Tensor Core72 PFLOPS sparse72 PFLOPS sparse
FP16 / BF16 Tensor Core36 PFLOPS sparse36 PFLOPS sparse
TF32 Tensor Core18 PFLOPS sparse18 PFLOPS sparse
INT8 Tensor Core72 POPS sparse3 POPS sparse
FP32600 TFLOPS600 TFLOPS
FP64 / FP64 Tensor Core296 TFLOPS10 TFLOPS
NVLink5th gen, 1.8 TB/s per GPU, 14.4 TB/s total5th gen, 1.8 TB/s per GPU, 14.4 TB/s total
Networking bandwidth0.8 TB/s1.6 TB/s
Attention performance1x2x

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.

Hopper

H200, per GPU.

GPU memory141 GB HBM3e
Memory bandwidth4.8 TB/s
FP4 Tensor CoreNot listed by NVIDIA
FP8 Tensor Core3,958 TFLOPS sparse
FP16 / BF16 Tensor Core1,979 TFLOPS sparse
TF32 Tensor Core989 TFLOPS sparse
INT8 Tensor Core3,958 TOPS sparse
FP64 / FP64 Tensor Core34 TFLOPS / 67 TFLOPS
NVLink900 GB/s per GPU
PowerUp 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.

What changes from one generation to the next.

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.

More GPU memory at each step

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).

NVLink doubles per GPU

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.

B300 trades FP64 and INT8 for AI throughput

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.

Head to head.

B200 vs B300

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).

H200 vs B200

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).

H200 vs B300

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.

Which fits which work.

Large models and long context

B300: 2.1 TB of GPU memory and twice the attention performance of B200 (NVIDIA figures).

Scientific computing beside AI

B200: 296 TFLOPS of FP64 against B300's 10 (NVIDIA figures), so simulation and other double-precision work stay practical.

INT8 inference pipelines

B200: 72 POPS of INT8 against B300's 3 (NVIDIA figures, sparse).

Already running H200

Models served in FP8 or FP16 keep running. Moving to Blackwell adds FP4 and doubles NVLink bandwidth per GPU.

Common questions.

  • Does the H200 support FP4?
    NVIDIA lists FP8 as the lowest precision for H200, at 3,958 TFLOPS with sparsity per GPU, and gives no FP4 figure. FP4 arrives with Blackwell.
  • Is the B300 faster than the B200?
    For most AI work, yes on NVIDIA's figures: dense FP4 rises from 72 to 108 PFLOPS across eight GPUs and attention performance doubles. FP8 and FP16 are the same on both, and FP64 and INT8 are far lower on the B300, so for double-precision or INT8 work the B200 is the stronger of the two.
  • How much GPU memory does each have?
    141 GB per H200 GPU, so 1,128 GB across eight (our arithmetic). 1.4 TB across eight B200 GPUs and 2.1 TB across eight B300 GPUs (NVIDIA figures).
  • What is the difference between HGX and DGX?
    HGX is NVIDIA's eight-GPU platform that server makers build into their own systems. DGX is NVIDIA's own complete system. The figures above are NVIDIA's HGX figures; our DGX B200 and DGX B300 pages list each DGX system's own specification.
  • How does GB300 NVL72 compare with H200?
    They are different sizes of thing. GB300 NVL72 is a whole rack: 72 Blackwell Ultra GPUs and 36 Grace CPUs with 20 TB of HBM3e, liquid-cooled (NVIDIA figures). H200 is one GPU with 141 GB. Our GB300 NVL72 page has the rack's specification.
  • What comes after B300?
    NVIDIA HGX Rubin NVL8: eight Rubin GPUs with 2 TB of HBM4 at 130 TB/s and 400 PFLOPS of NVFP4 for inference with sparsity (NVIDIA figures). Our AI HGX Rubin is built to order on it.
  • Can you help us choose?
    Yes. Tell us the models, context lengths and precision you plan to run, and we recommend a platform in writing, in MYR.

Tell us about your workload.

Tell us the models, context lengths and precision you plan to run. We recommend a platform in writing, in MYR.