On February 1, 2024, Reuters broke the news that Nvidia has started accepting pre-orders for the new China-specific version of the AI chip from dealers, who are offering pricing comparable to Huawei's competing products.

Called the H20, this graphics card is the most powerful of the three graphics cards developed by Nvidia for the Chinese market.

The H20 naturally has less computing power than Nvidia's flagship H100 AI chip and H800, a graphics card aimed at the Chinese market, which was also banned in October last year.

According to sources, the H20's specs also seem to indicate that it is not as capable as the Huawei Ascend 910B in some key aspects.

The specifications of H20 are still relatively strong, INT8/FP8 Tensor computing power is 296T, BF16/FP16 Tensor computing power is 148T, TF32 Tensor computing power is 74T, FP32 computing power is 44T, FP64 computing power is 1T, and power consumption is 400W.

It has 60MB L2 cache, 96GB HBM3 memory, 4TB/s bandwidth, supports 7-channel NVDEC, 7-channel NVJPEG video engine, and PCIe 5.0 x16 system interface, 900GB/s NVLink interconnection bandwidth, and supports up to 7 MIGs.

Compared with H200, it has 45GB less HBM3, and the computing power is even more castrated, INT8/FP8, BF16/FP16, TF32 are as much as 93% less, FP64 is 97% less, and only FP32 is only 35% less.

But even so, the performance of the H20 FP32 is still not as good as that of the Huawei Ascend 910B, which is also the biggest advantage of the latter.

According to two sources, Nvidia's pricing for China's H20 in recent weeks is said to be only $1.2-$15,000, which is much lower than the $17,000 of Huawei's Ascend 910B. Some of these dealers have started selling chips and have increased their prices significantly, selling them for about 110,000 yuan.

The H20 can also be used in parallel to form a set of HGX servers for about $200,000, compared to about $280,000 for eight H800 servers before.

The source added that dealers have told customers that they will be able to start delivering H20 products in small batches in the first quarter of 2024 and in large quantities from the second quarter.

In terms of specs, the H20 lags behind the 910B in terms of FP32 performance, a key measure of how fast a chip can handle ordinary tasks, with less than half the standard performance of its competitors.

However, the H20 seems to have an advantage over the 910B in terms of interconnect speed, which measures the speed at which data is transferred between chips. This means that the H20 is still somewhat competitive compared to the 910B in applications where a large number of chips need to be connected together to work as a complete system.

Nvidia plans to start mass production of the H20 in the second quarter of this year.

Gartner ranks the world's top 10 semiconductor manufacturers by revenue in 2023 (in billions of dollars), and Nvidia jumps to fifth.

Previously, the Wall Street Journal reported that Nvidia now faces a more troublesome problem: China's big customers are not actively buying performance downgraded chips.

Alibaba Group, Tencent, Baidu, and ByteDance have been testing Nvidia's special chip samples since November last year, according to people familiar with the matter.

They have indicated to Nvidia that the number of chips ordered from Nvidia this year will be far less than the banned Nvidia high-performance chip (H/A800) that they had planned to buy.

· END·