NVIDIA H20 96GB vs NVIDIA H200 SXM 141GB

The H200 delivers 6.7x the FP8 throughput of the H20 (1,979 vs 296 TFLOPS dense). The H200 also carries 45 GB more memory (141 GB vs 96 GB).

H20: Hopper, 2024 H200: Hopper, 2024
FP8 dense lead
6.7x
H200: 1,979 vs 296 TFLOPS
Memory
96 vs 141 GB
45 GB more for the H200
Bandwidth
4.0 TB/s vs 4.8 TB/s
H200 moves data faster
TDP
400 W vs 700 W
H20 draws 300 W less
See rental pricing from $0.79/hr Download comparison sheet

Share this page

NVIDIA H20 96GB vs NVIDIA H200 SXM 141GB

A post with this link shows the card on the Image tab as its preview.

Printable version →

NVIDIA H20 96GB

Full specs →

What is the NVIDIA H20?

The H20 is a GPU from NVIDIA on the Hopper architecture, an accelerator for AI training, inference and high-performance computing. It is an SXM module, the mezzanine format used in dense multi-GPU servers. H20 full specifications →

Can I rent the H20 in the cloud?

Flopper is not tracking any live rental listings for the H20 yet. The provider directory lists the clouds that rent comparable accelerators on demand. GPU cloud providers →

How does the H20 compare with other accelerators?

This page sets the H20 against the H200. Both parts launched in the same year. Both are NVIDIA parts, and both ship as an SXM module, so they fit the same class of system. Flopper also compares it head to head with the Zhenwu 810E, across every published precision, memory and power. H20 vs Zhenwu 810E →

Where do the H20 figures come from?

Every H20 figure on this page traces to NVIDIA's own datasheet or specification table. How Flopper verifies specifications →

NVIDIA H200 SXM 141GB

Leads on FP8 Full specs →

What is the NVIDIA H200?

The H200 is a GPU from NVIDIA on the Hopper architecture, an accelerator for AI training, inference and high-performance computing. It is an SXM module, the mezzanine format used in dense multi-GPU servers. H200 full specifications →

Can I rent the H200 in the cloud?

Yes. Cloud providers tracked by Flopper rent the H200 by the hour, and the pricing page compares their current rates and the price history side by side. H200 rental prices →

How does the H200 compare with other accelerators?

This page sets the H200 against the H20. Both parts launched in the same year. Both are NVIDIA parts, and both ship as an SXM module, so they fit the same class of system. Flopper also compares it head to head with the H100 SXM5 80GB, B300 SXM 262GB and A100 SXM4 80GB, across every published precision, memory and power. H200 vs H100 SXM5 80GB →

Where do the H200 figures come from?

Every H200 figure on this page traces to NVIDIA's own datasheet or specification table. How Flopper verifies specifications →

Specifications

H20
H200
Architecture
HopperHopper
Launch Year
20242024
Form Factor
SXMSXM
Memory
96 GB141 GB
Memory Bandwidth
4.0 TB/s4.8 TB/s
TDP
400 W700 W
Process Node
4nm4nm

Performance (TFLOPS)

H20
H200
1 TFLOPS 34 TFLOPS
44 TFLOPS 67 TFLOPS
74 TFLOPS
sparse not published
494 TFLOPS
989 TFLOPS sparse
148 TFLOPS
sparse not published
990 TFLOPS
1,979 TFLOPS sparse
148 TFLOPS
sparse not published
990 TFLOPS
1,979 TFLOPS sparse
296 TFLOPS
592 TFLOPS sparse
1,979 TFLOPS
3,958 TFLOPS sparse
FP6
No verified data available No verified data available
No verified data available No verified data available
296 TOPS
sparse not published
1,979 TOPS
3,958 TOPS sparse

Figures are dense. Where a vendor publishes a figure it sits under the dense one and never decides a winner. Hover any precision for its definition, or read the GPU glossary.

FLOPS by Precision

What actually differs

Memory bandwidth survives the H20’s export cut; compute does not. The H20 keeps 4.0 TB/s against 4.8 TB/s on the H200, so LLM decode on a model that fits stays within reach of H200 speed. That makes the H20 a serving part, not a training one.

The cut lands on prefill, long prompts and training. The H20 is rated at 148 TFLOPS BF16 against 990 TFLOPS BF16 dense on the H200, a gap no batch size closes. Compare dense to dense: the H200’s larger headline figures are with sparsity, and Flopper lists no sparse rating for the H20.

Outside export-restricted markets the H200 is the straightforward pick, rentable from $2.00 per GPU-hour on the pricing page. Inside China, weigh the H20 against the Ascend 910C.

New to the terms? The glossary explains FP64, FP32, TF32, BF16, FP16, FP8, INT8, Sparsity and TFLOPS. For how the generations relate, see NVIDIA architectures explained.

Frequently asked questions

Is the H200 faster than the H20?

On every compute row, by a wide margin: 990 TFLOPS BF16 dense against 148 TFLOPS BF16, and 1,979 TFLOPS FP8 dense against 296 TFLOPS FP8. Token generation narrows the gap, because it is bound by memory bandwidth, where the H200 leads only 4.8 TB/s to 4.0 TB/s.

Is the H20 good for LLM inference?

For decode-heavy serving of a model that fits, yes. Its bandwidth sits close to the H200's, so tokens stream at a useful rate. Long prompts, prefill and training expose the compute cut, and 96 GB against 141 GB leaves less room for KV cache, which caps batch size.

Should I choose the H20 or the H200?

Choose the H200 wherever you can buy or rent it: it does everything the H20 does, faster, with more memory. The H20 exists for buyers bound by U.S. export rules, mainly in China, where the real alternatives are domestic accelerators such as the Huawei Ascend 910C.

Can you rent an H20 or an H200?

Flopper tracks no rental listings for the H20, which sells into the Chinese market through server makers and has no NVIDIA product page. The H200 is widely available: 57 listings across 31 providers, from $2.00 to $10.60 per GPU-hour, with the lowest price on spot capacity.

Where to Rent

NVIDIA H20 96GB

No live rental listings tracked for the H20 yet.

Browse GPU cloud providers

NVIDIA H200 SXM 141GB

Rent from

$0.79 / GPU-hr

Lowest of 56 live listings across 30 providers · median $4.38

See all H200 rental prices →

Price history and every provider on the pricing page. Some links are affiliate links.

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.

© 2026 Flopper.io - Compare the GPUs Powering AI