NVIDIA H20 (2024)
Product Overview
NVIDIA H20 is the export-compliant AI accelerator NVIDIA launched for the China market, based on the Hopper architecture (same architecture as the H100/H200), officially released in early 2024.
Restricted by US export controls (the October 2023 rules), H20's compute was deliberately cut to comply:
- FP16: 148 TFLOPS (~15% of the H100)
- FP8: 296 TFLOPS (~15% of the H100)
- Memory: 96GB HBM3 (4.0 TB/s bandwidth)
H20's positioning: keep NVIDIA's ecosystem advantages as much as possible while staying compliant, competing with domestic AI chips (Hygon DCU, Kunlunxin, MetaX, etc.).
Core Specifications
| Parameter | Value |
|---|---|
| Architecture | Hopper (export-compliance cut-down) |
| Process | TSMC 4N (4nm) |
| FP8 | 296 TFLOPS |
| FP16 | 148 TFLOPS |
| TF32 | 74 TFLOPS |
| FP32 | 24 TFLOPS |
| INT8 | 296 TOPS |
| Memory Capacity | 96GB HBM3 |
| Memory Bandwidth | 4.0 TB/s |
| Interconnect | NVLink 900 GB/s (downgraded) |
| TDP | 400W |
| Release | Early 2024 |
| Discontinued | July 2025 (NVIDIA notified channel partners) |
Export Compliance Background
US Export Control Timeline
| Date | Event |
|---|---|
| 2022-10 | US BIS issues export controls restricting A100/H100 exports to China |
| 2023-10 | New rules tighten further, adding a "performance density" threshold |
| 2023-11 | NVIDIA launches H20 / L20 / L2 (compliant versions) |
| 2024-02 | H20 begins shipping to Chinese customers |
| 2025-07 | NVIDIA notifies channel partners that H20 will be discontinued (export controls tighten again) |
Compliance-Limited Parameters
| Parameter | Compliance Threshold | H20 Actual |
|---|---|---|
| FP16 compute | < 300 TFLOPS | 148 TFLOPS |
| Performance density | < a certain threshold | Compliant |
| Memory bandwidth | No explicit limit | 4.0 TB/s (uncut) |
Strategy: NVIDIA chose to keep memory bandwidth (important for LLM inference) while cutting compute sharply, staying competitive within compliance.
Comparison with Domestic Chips
H20 vs Hygon DCU K100
| Metric | NVIDIA H20 | Hygon DCU K100 | Difference |
|---|---|---|---|
| FP16 | 148 TFLOPS | 192 TFLOPS | K100 +30% |
| Memory | 96GB HBM3 | 64GB HBM3 | H20 +50% |
| Bandwidth | 4.0 TB/s | 3.2 TB/s | H20 +25% |
| Ecosystem | CUDA (complete) | ROCm (compatible) | H20 advantage |
| Supply | Discontinued in 2025 | Stable supply | K100 advantage |
| Price | ~$25k (est.) | ~$15k (est.) | K100 cheaper |
H20 vs Kunlunxin P800
| Metric | NVIDIA H20 | Kunlunxin P800 | Difference |
|---|---|---|---|
| FP16 | 148 TFLOPS | 345 TFLOPS | P800 2.3x |
| Memory | 96GB HBM3 | 32GB HBM3 | H20 +200% |
| Bandwidth | 4.0 TB/s | Not disclosed | H20 advantage |
| Ecosystem | CUDA | Baidu Paddle | Each has merits |
Conclusion: H20 still leads in memory capacity and bandwidth, but its compute has been overtaken by domestic chips. With H20 discontinued in 2025, domestic substitution is accelerating.
Use Cases
- ✅ Transition period for NVIDIA ecosystem migration (CUDA code needs no changes)
- ✅ LLM inference (96GB memory + 4.0 TB/s bandwidth advantage)
- ✅ Short-term projects (still procurable before July 2025)
- ❌ Long-term training clusters (discontinued in 2025, supply chain risk)
- ❌ Cost-effectiveness first (domestic chips offer higher compute at lower prices)
- ❌ Self-reliance requirements (the US government can tighten export controls at any time)
Discontinuation Impact (July 2025)
In July 2025, NVIDIA notified channel partners that H20 would be discontinued, because:
- US export controls could tighten further
- H20 margins were squeezed (deliberately cut compute)
- NVIDIA prioritized capacity for H200/B200
Impact on the China market:
- Customers who bought H20: no short-term impact; repairs/expansion difficult long-term
- Customers who did not: accelerated shift to domestic chips (Hygon, Kunlunxin, MetaX, etc.)
- Domestic substitution timeline: brought forward 1-2 years
Vendor Information
| Parameter | Value |
|---|---|
| Company | NVIDIA Corporation |
| Architecture | Hopper (export-compliance cut-down) |
| Release | Early 2024 |
| Discontinued | July 2025 (channel partners notified) |
| China market positioning | Export-compliant version replacing the H100/A100 |
| Competitors | Hygon DCU, Kunlunxin, MetaX, Biren, etc. |
Related Cards
- NVIDIA H100 — Full Hopper architecture (banned from sale in China)
- NVIDIA H200 — Upgraded H100 (banned from sale in China)
- Hygon DCU K100 — Domestic alternative, stronger FP16
- Kunlunxin P800 — Domestic alternative, 2.3x FP16
- MetaX C600 — Fully domestic GPU