AIXX Pedia The encyclopedia of compute hardware

AIXX Pedia › Model Comparison › H100 × H20

H100 vs H20

Model Comparison · Verdict First · Specs Side by Side · AIXX Index Side by Side · AIXX Pedia

Verdict First VERDICT

HOW TO CHOOSE · One-Paragraph Verdict
Inference and large-memory workloads (single-GPU 70B-class inference, full-scale R1 on 8 GPUs, all-in-one appliance production lines), choose H20 (96GB HBM3 / 4.0TB/s, actually surpassing the H100, roughly 400W for easy deployment, and previously sold as a compliant model for mainland China); Training, faster fine-tuning, high compute density, and HPC/FP64 workloads , choose the H100 (full Hopper compute, the largest ecosystem). The H20's Tensor compute is roughly one-tenth of the H100's on a like-for-like basis (per media reports) — in one sentence: the H20 sells memory and market access; the H100 sells compute.

Specs Side by Side SPECIFICATIONS

ParameterNVIDIA H100 · Hopper GenerationNVIDIA H20 · Hopper China-market Variant
ArchitectureHopperHopper (China-market Variant derived from the GH100)
ProcessTSMC 4NTSMC 4N (per media reports)
Board Memory80GB HBM3 (640GB across 8 cards)96GB HBM3 (768GB across 8 GPUs; upgraded version reportedly ~148GB HBM3e)
Memory Bandwidth3.35TB/s per card4.0TB/s per GPU (media reports, above the H100's 3.35TB/s)
GPU-to-GPU InterconnectNVLink 4 · 900GB/sNVLink 4 · 900GB/s (HGX module spec; PCIe standalone cards have no NVLink)
Board Power (TBP/TGP)Up to 700W (configurable)~400W TDP (media reporting)
Form Factor & CoolingHGX H100 8-card baseboard (mostly air-cooled towers · OEM liquid-cooled variants available)HGX H20 8-GPU baseboard · PCIe standalone card (400W-class air-cooled tower)
Launch / DeliveryAnnounced 2022-03 · shipping since 2022-10Launched in 2024 (per media reports)
China-market TimelineSubject to China export controls since 2022-10 · not on the approved list for ChinaChina-market Variant · licensing required 2025-04 → approval to resume sales 2025-07 → supply fluctuating with policy shifts (per public reports)
Mainland China Street Price BandApprox. 5.02M per system (2026-10-07 live market floor · 1 unit · whole-lot asking basis)None currently in stock in this library (verified 2026-10-08); historical anchor: ~1.1M per 8-GPU system in 2025-02 (per Cailianshe reporting)

Specs are based on NVIDIA official public documentation and the two entry pages; price bands reflect actual on-sale listings in this library's marketplace and fluctuate with inventory—see the entry page "price band" for the latest. A dash in both columns means the entry does not yet include this parameter.

AIXX Index Side by Side WANAN INDEX

AIXX Index (0–100): estimated by this library using a weighted public-spec model—memory 40% + interconnect 30% + ecosystem 30%, weights adjusted by scenario. 90+ flagship · 75+ Excellent · 60+ usable · below 60 not applicable. The method is transparent and verifiable and does not constitute a performance promise; figures match the published figures on the two entry pages, so this page does not score separately.

Scenario IndexH100H20
LLM Inference8588
Model Training / Fine-tuning8855
HPC / FP64 Computing9240
Cost-sensitive Inference9085
Large-memory scenarios (long context / full R1)5890
Condition / Reliability / Value Retention (Experience Index)Pending hands-on testingPending hands-on testing

"—" = the machine's entry page has no rating for this scenario; this page does not fill it in. The Experience Index (condition/reliability/value retention) is pending hands-on testing—inspection batch data is being accumulated and will light up unit by unit as Inspection Reports are published.

Which to Choose for Which Scenario WHICH ONE

Choose the H20 when —
Inference-first with large memory needs: 96GB per card / 768GB across 8 cards fits full-scale FP8 R1 (~690GB) on eight GPUs, and 400W air-cooled cards drop straight into existing server rooms.
Choose the H100 when —
Choose the H100 when you need full compute: training, faster fine-tuning, HPC/FP64, and high-density workloads — the largest ecosystem, the deepest installed base, and the pricing anchor of the rental market.
Choose neither, when—
If the bottleneck is memory capacity and spot availability, pick the H200 (141GB HBM3e, approved for China in 2025-12); for FP4 and next-gen interconnect, go with the B300.

Keep Verifying LINKS

After this comparison, return to the entry pages to verify details and the latest price bands:

Footnote: This page is for reference comparison only and does not constitute a trading offer; actual price subject to the seller's quote. Data source: AIXX Compute Exchange (entry figures 2026-10; index method identical to the two entry pages, transparent and verifiable).
This Library
Editorial PolicyAbout This Library (Disclosure of Interests)Corrections Channel
Editorial Constitution Summary
Every claim sourcedDrawbacks always stated · corrections publicly loggedNo paid entries acceptedZero traffic-funneling or sales elements within entriesMarket data attached only as data modules
AIXX Family
AIXX PediaAIXX ExchangeAIXX News
Market data source: AIXX Compute Exchange · © 2026 AIXX Pedia · An AIXX Production