NVIDIA H20

What it is

The NVIDIA H20 is a data-center GPU designed specifically for the Chinese market so that NVIDIA could keep selling there while complying with US export controls ^[2]. It was the most powerful of three China-specific chips that NVIDIA based on its Hopper and Ada Lovelace architectures ^[2]. Launched at the end of 2023, it was the most advanced NVIDIA chip Chinese buyers could legally obtain for training and running models until the licence requirement introduced in April 2025 ^[3].

How it differs from the H100

The clearest way to see what export controls cost is to put the H20 next to the H100. The H20 has 78 streaming multiprocessors (SMs), the repeated blocks of cores that carry out the arithmetic, against 144 on the H100 — a 41% cut in GPU core count. NVIDIA's own comparison put its performance 28% lower ^[1].

Memory is where the two come closer. The H20 carries 96 GB of HBM3 memory with 4.0 TB/s of memory bandwidth, against 8 TB/s on the H100 ^[1]. Bandwidth often matters more than raw arithmetic when a finished model is being run rather than trained, because producing each token means streaming the model's weights through the chip, and holding plenty of memory lets a large model fit on fewer cards. A part with fewer cores but generous memory can therefore stay useful for inference even while trailing badly on compute.

The published numbers

NVIDIA's published figures for the H20 are 296 TFLOPS at INT8, 148 TFLOPS at FP16, 74 TFLOPS at TF32, 44 TFLOPS at FP32 and 1 TFLOPS at FP64 ^[1]. The last is the figure that matters in scientific computing needing double precision, and it is very low; neural networks mostly work in lower precision, so the INT8 and FP16 numbers are the relevant ones.

Other specifications include 60 MB of L2 cache, support for eight-way HGX configurations, 900 GB/s of NVLink bandwidth between GPUs and a 400 W thermal design power ^[1]. Compared with top-end training accelerators, 400 W is a modest power envelope.

Price, orders and reception

The chip was priced at about US$12,000 per unit ^[1]. NVIDIA began taking orders in February 2024, started shipping in April and reached volume shipments in May 2024 ^[2].

Orders and enthusiasm were not the same thing. Chinese distributors and server vendors described the H20 as barely worth the trouble, and one buyer said a card still cost more than one million yuan ^[2]. A cut-down part that is still expensive is a difficult sell: buyers compare it both with what they could have bought before the controls tightened and with alternatives available at home.

Where it sits

The H20 is a case study in how export controls reshape chip design. The same architecture was trimmed — fewer cores, a wider gap to the flagship — while keeping relatively large memory, so that the product lands under a performance line drawn by rules rather than by the market ^[1][2]. Its short life shows the weight of policy as much as engineering: a chip that began as the most advanced NVIDIA product legally available to Chinese buyers stopped being freely sellable once the licence requirement arrived in April 2025 ^[3].