H20 (China) — Specs, Rank & Tuning
NVIDIA Datacenter Hopper · Hopper · 2024 · High-end (tier 82/100) · Export Hopper for China (memory-heavy, compute-capped)
H20 (China) — key specifications
| Brand / series | NVIDIA Datacenter Hopper |
|---|---|
| Architecture / class | Hopper |
| Launch year | 2024 |
| Memory | 96 GB HBM3 · 6144-bit · 4000 GB/s |
| Boost clock | 1980 MHz |
| FP16 compute | 74 TFLOPS |
| FP8 compute | 148 TFLOPS |
| Interconnect | NVLink / SXM5 |
| TDP | 400 W |
| Relative tier | 82/100 — High-end |
Specs: vendor datasheets / TechPowerUp database, retrieved 2026-08-29. Tier is a TweakGearPro relative performance index (estimate), not a benchmark guarantee.
Relative performance estimate
TweakGearPro anchors the ai gpu category to its flagship (highest tier) and expresses H20 (China) as an estimated percentage of that flagship, derived from the relative-performance index (tier 0–100). Planning estimate, not a benchmark.
| AI GPU (nearby tier) | Tier /100 | Est. relative perf. |
|---|---|---|
| H100 SXM 80GB | 96 | 96% |
| AMD Instinct MI325X | 96 | 96% |
| RTX PRO 6000 Blackwell | 95 | 95% |
| AMD Instinct MI300X | 95 | 95% |
| GeForce RTX 5090 | 92 | 92% |
| H100 PCIe 80GB | 88 | 88% |
| Intel Gaudi 3 | 88 | 88% |
Method: relative performance % = (part tier ÷ flagship tier) × 100, using the TweakGearPro tier index built from public benchmark / spec-class data. Source: vendor datasheets / TechPowerUp database, retrieved 2026-08-29. Real-world performance varies by workload, drivers and cooling.
Build compatibility & pairing guidance
Practical build notes derived from this part's dataset fields and TweakGearPro reference constants. Planning guidance, not purchase advice.
| Estimated system PSU | ≈ 750 W (this part's 400 W TDP + ~200 W system overhead, with 20% headroom). Confirm with the PSU Wattage Calculator. |
|---|---|
| Cooler / thermal | Plan cooling rated above this part's 400 W TDP; typical desktop safe ceiling is ~90°C (TweakGearPro reference). Check the Thermal Headroom tool. |
| Pairing balance | Match with a CPU whose tier is within ~20 of this part's tier (82/100) to avoid a bottleneck. Run the Bottleneck Estimator. |
| VRAM fit | 96 GB HBM3 — for gaming/ML sizing see the VRAM Recommendation tool or the LLM planner below. |
Reference constants (PSU headroom 20%, safe CPU/GPU ceiling 90°C) from the TweakGearPro PC reference dataset, retrieved 2026-08-29. Adapter/case clearance and exact board power vary by model — verify with vendor datasheets.
Tune & power-plan your build
Input your cooling, case airflow and goal to generate a printable tuning sheet, or size your whole-system PSU.
LLM deployment planner
VRAM math is an estimate — KV-cache sizing depends on context length and implementation.
My deployment plan — H20 (China)
Good fit for local/edge inference. Expect ~1600% of weights+KV to fit with INT4 quantization. Generated at TweakGearPro · 2026-09-13
Release & generational notes
H20 (China) launched in 2024 on the Hopper architecture as part of the Datacenter Hopper family, sitting at tier 82/100 (High-end).
| Series | Datacenter Hopper |
|---|---|
| Architecture | Hopper |
| Launch year | 2024 |
| Predecessor | H100 SXM 80GB (2022, tier 96/100) |
In the Datacenter Hopper series
H100 PCIe 80GB
2022 · Hopper · tier 88/100
Open →H100 SXM 80GB
2022 · Hopper · tier 96/100
Open →H200 SXM 141GB
2024 · Hopper · tier 97/100
Open →Generational context is summarized from official release year and architecture only; TweakGearPro does not invent patch or refresh dates. For the full generational ranking, see the AI-GPU Tier List (LLM).
Related hardware to compare
Same-tier peers plus a pairing from another category — open any profile.
Sources & methodology
- Specifications: vendor datasheets (NVIDIA, AMD, Intel, Samsung, WD, Crucial, Corsair, Seagate, Noctua, ASUS, LG, Dell and others) and the TechPowerUp database, retrieved 2026-08-29.
- Relative-performance index: TweakGearPro estimate built from public benchmark / spec-class data; flagship-anchored percentages are planning estimates, not measured scores.
- Reference constants (PSU headroom, safe thermal ceiling) from the TweakGearPro PC reference dataset, retrieved 2026-08-29.
- Tuning, PSU and VRAM figures are planning estimates — verify with your own stress tests and the component vendor's documentation.
How this page's estimates are computed
The relative-performance percentage anchors the category flagship (highest tier) at 100 and scales each part by its tier ratio. The estimated system PSU adds a base system overhead to the part's TDP and applies a 20% headroom from the reference dataset, rounded to a common PSU rating. The tuning sheet estimates load temperature from TDP, cooling class and case airflow, then suggests a voltage offset and expected clock delta — always confirm with a stress test. The LLM planner converts model parameters and quantization to an approximate VRAM footprint (weights + KV-cache) and divides by the card's usable VRAM to suggest a card count. None of these outputs are benchmark guarantees. For non-CPU/GPU parts the tier reflects the part's spec class relative to the best current model in that category.