AI chip · NVIDIA
NVIDIA H20
Second-generation China-specific variant of Hopper. Introduced late 2023 after Oct 17 BIS rule blocked H800. Kept the memory (96 GB HBM3) and interconnect (900 GB/s NVLink) but sharply reduced dense compute (~148 FP16 TFLOPS vs H100's 989) to fit under BIS's performance-density threshold.
Market position
The compute-throttled H100 variant that shipped into China through 2024. Chinese hyperscalers (ByteDance, Alibaba, Tencent) bought hundreds of thousands. Trump-admin April 2025 rule then required per-shipment licenses, which were granted in July 2025 under a 15% revenue-share arrangement (first of its kind for a chip export license).
Efficiency and power
How much work you get per watt and per dollar, and how much power a full rack draws.
- Perf per watt
- 0.74 FP8 TFLOPS/W296 TFLOPS ÷ 400 W = 0.74
What fits in 96 GB
Which open-source LLMs run on one NVIDIA H20, by precision. Weights only: add roughly 15% headroom for real serving. If a model does not fit at FP16, try INT8 or INT4 (smaller quality trade-off than most people expect).
| Model | Params | FP16 | INT8 | INT4 |
|---|---|---|---|---|
| Llama 3.1 8B dense | 8 B | 16 GB ✓ | 8 GB ✓ | 4 GB ✓ |
| Llama 3.1 70B dense | 70 B | 140 GB ✗ | 70 GB ✓ | 35 GB ✓ |
| Llama 3.1 405B dense | 405 B | 810 GB ✗ | 405 GB ✗ | 203 GB ✗ |
| Llama 3.3 70B dense | 70 B | 140 GB ✗ | 70 GB ✓ | 35 GB ✓ |
| DeepSeek V3 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✗ |
| DeepSeek R1 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✗ |
| Qwen 2.5 7B dense | 7 B | 14 GB ✓ | 7 GB ✓ | 4 GB ✓ |
| Qwen 2.5 72B dense | 72 B | 144 GB ✗ | 72 GB ✓ | 36 GB ✓ |
| Mixtral 8x7B MoE (46.7B total, 12.9B active per token) | 46.7 B | 93 GB ✓ | 47 GB ✓ | 23 GB ✓ |
| Mixtral 8x22B MoE (141B total, 39B active per token) | 141 B | 282 GB ✗ | 141 GB ✗ | 71 GB ✓ |
| Gemma 2 27B dense | 27 B | 54 GB ✓ | 27 GB ✓ | 14 GB ✓ |
| Command R+ dense | 104 B | 208 GB ✗ | 104 GB ✗ | 52 GB ✓ |
| Kimi K2 MoE (1T total, 32B active per token) | 1000 B | 2000 GB ✗ | 1000 GB ✗ | 500 GB ✗ |
Math: FP16 = params × 2 bytes; INT8 = params × 1 byte; INT4 = params × 0.5 bytes. MoE models sum every expert (full weights on disk), not the per-token active subset.
Export controls
US Commerce Department rules that restrict where NVIDIA H20 can be sold. Effective dates and full text below, with the official Federal Register filings and news coverage for each.
The Oct 17 2023 update closed the workaround NVIDIA used with A800 and H800: BIS replaced the interconnect-bandwidth threshold with a performance-density metric (total processing performance x performance density) and added a Notified Advanced Computing regime for chips that fell just below the ceiling. Direct effect: A800 and H800 could no longer be exported to China. Triggered NVIDIA's H20 as the next-generation China-specific variant.
Sources: BIS press release, Oct 17 2023 · Federal Register, Oct 25 2023 (rule text)
In April 2025 the Trump administration required NVIDIA to obtain an export license to ship the H20 (the last China-permitted NVIDIA data-center GPU) into China. NVIDIA disclosed a $5.5 B Q1 charge tied to unshipped H20 inventory. Licenses were subsequently granted in July 2025 under a revenue-share arrangement (15% of China-tied H20 revenue paid to the US government), the first such structure applied to a chip export license.
Sources: NVIDIA Q1 FY26 results (H20 charge disclosure) · Reuters, Apr 15 2025 (H20 license requirement)
Common questions
How much does NVIDIA H20 cost?
NVIDIA H20 doesn't have a public launch list price. Vendors like this one usually price through direct sales rather than a public sheet. See list price per TFLOP for chips that do publish.
How much memory does NVIDIA H20 have?
96 GB of HBM3, running at 4,000 GB/s. Straight from the vendor datasheet. See chips with the most memory for context.
How much power does one NVIDIA H20 draw?
400 W at the chip. A full server draws more once you add CPU, memory, networking and cooling: see the rack power number above. Compare to other chips on perf-per-watt.
Which open-source LLMs fit on one NVIDIA H20?
Of 13 open-source LLMs we track, 4 fit at FP16, 7 at INT8, and 9 at INT4 (for example Llama 3.1 8B, Llama 3.1 70B and Llama 3.3 70B). Full table above with each model's memory need.
Is NVIDIA H20 subject to US export controls?
Yes. 2 US Commerce Department rules restrict where this chip can be sold, most recently NVIDIA H20 China export licensing requirement (April 2025) (2025-04-15). Details and effective dates in the export-control section above.
Key facts
- Class
- Compute SoC (AI accelerator)
- Designer
- NVIDIA
- Safety-critical?
- No (data-center inference / training)
- Record as of
- 2026-09-27
Specifications (9 fields, click to expand)
Straight from the vendor datasheet. Dense throughput shown first; sparse (2:4) numbers in parentheses where the vendor publishes them. Full datasheet linked below.
- Process node
- TSMC 4N
- TDP
- 400 W
- Memory
- 96 GB HBM3
- Memory bandwidth
- 4,000 GB/s
- FP16 (dense)
- 148 TFLOPS
- FP8 (dense)
- 296 TFLOPS
- Form factor
- SXM5
- Announced
- 2023-11-01
- Released
- 2024-01-01
Source: vendor datasheet
Generation
Foundry & process
- Process node
- TSMC 4N
Sources
Adoption rows appear as we confirm each chip-in-robot pairing from a public source. See every chip we track for the full catalog or the NVIDIA page.