DEPLOYDatabase

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).

DEPLOY editorial. What the vendor PDF cannot tell you.

Physical-AI cross-link

Derived from claims already on record (vendor first-party TDP, throughput, price, and the reference-design cluster registry). The arithmetic is shown per row so it can be audited. No estimated inputs.

Perf per watt
0.74 FP8 TFLOPS/W
296 TFLOPS ÷ 400 W = 0.74

What fits in 96 GB

Weights-only footprint for public open-weight LLMs. Pure arithmetic: params × bytes/param. Excludes KV cache and activation memory; add ~10-30% headroom for real serving. A model that does not fit at FP16 may still fit at INT8 or INT4 with quality trade-offs. Not a benchmark.

ModelParamsFP16INT8INT4
Llama 3.1 8B
dense
8 B16 GB ✓8 GB ✓4 GB ✓
Llama 3.1 70B
dense
70 B140 GB ✗70 GB ✓35 GB ✓
Llama 3.1 405B
dense
405 B810 GB ✗405 GB ✗203 GB ✗
Llama 3.3 70B
dense
70 B140 GB ✗70 GB ✓35 GB ✓
DeepSeek V3
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
DeepSeek R1
MoE (671B total, 37B active per token)
671 B1342 GB ✗671 GB ✗336 GB ✗
Qwen 2.5 7B
dense
7 B14 GB ✓7 GB ✓4 GB ✓
Qwen 2.5 72B
dense
72 B144 GB ✗72 GB ✓36 GB ✓
Mixtral 8x7B
MoE (46.7B total, 12.9B active per token)
46.7 B93 GB ✓47 GB ✓23 GB ✓
Mixtral 8x22B
MoE (141B total, 39B active per token)
141 B282 GB ✗141 GB ✗71 GB ✓
Gemma 2 27B
dense
27 B54 GB ✓27 GB ✓14 GB ✓
Command R+
dense
104 B208 GB ✗104 GB ✗52 GB ✓
Kimi K2
MoE (1T total, 32B active per token)
1000 B2000 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-control status

US Bureau of Industry and Security actions that constrain this chip. Each row links a regulation record (dated, sourced, on DEPLOY) to this chip. This is one of the most-searched and least-structured topics in the whole category, and everyone else writes it as paywalled prose. Structured, dated, sourced.

  1. 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)

  2. 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

Answer-first, sourced. Every claim below traces to a specific field on this page or a cited datasheet. FAQPage schema is emitted so LLM crawlers can lift these verbatim.

How much does NVIDIA H20 cost?

NVIDIA H20 launch pricing is not on record in the DEPLOY registry. Vendor list prices are typically published in the vendor's own briefing rather than the datasheet.

How much memory does NVIDIA H20 have?

96 GB of HBM3 at 4,000 GB/s. Vendor datasheet, first-party.

How much power does one NVIDIA H20 draw?

400 W TDP (thermal design power) per chip. System-level draw is higher: see the reference-design section for rack-level kW.

Which open-weight LLMs fit on one NVIDIA H20?

Of the 13 public open-weight LLMs DEPLOY tracks: 4 fit at FP16, 7 at INT8, 9 at INT4 (weights only, excludes KV cache). The full table is above with per-model math. Sparsity, offloading and multi-GPU serving change the picture; this row is single-chip weights-only.

Is NVIDIA H20 subject to US export controls?

Yes. 2 BIS actions on record constrain this chip, most recently NVIDIA H20 China export licensing requirement (April 2025) (2025-04-15). Full list with citations above.

Key facts

Class
Compute SoC (AI accelerator)
Designer
NVIDIA
Safety-critical?
No (data-center inference / training)
Record as of
2026-08-25
Specifications (9 fields, click to expand)

Vendor datasheet figures (first-party). Dense throughput first; sparse (2:4) in parentheses where NVIDIA quotes it. The exhaustive spec sheet lives on the datasheet URL below the table: this row set covers what buyers actually compare on.

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

Sources

Adoption rows appear as we verify chip-integration claims per model. Every row is a public claim the maker or a tier-1 source has stated; verification tier (verified / reported / inferred) is shown per row. See every chip on record for the full catalog or the NVIDIA record.