When it comes to chips for AI calculations, prices skyrocket quickly. (Image: Nvidia / Pixabay)
Nvidia is known to make a lot of money from AI accelerators, also (or especially) in China. However, trade restrictions in the USA mean that the fastest models are not allowed to be sold there. This circumstance is once again worrying new
trimmed GPUs from Nvidia.
Custom AI chips: How Reuters reported, there are a total of three models. On the one hand H20 with Hopper architecture, on the other hand L20 and L2 with Ada Lovelace architecture.
According to the H20 model, it comes from Wccftech published overview (see following figure) includes, among other things, 96 GB of HMB3 memory with 4.0 TByte/s bandwidth and a computing power of 296 FP8 teraflops.
Impressive technical key data in itself. They are flanked by a price of up to $15,000, which initially also seems extreme. But in one crucial discipline, H20 is technically far behind the already available (and much more expensive) top models from Nvidia.
The data of the new Nvidia GPUs for the Chinese market according to Wccftech.com.
The conscious regression in numbers
Nvidia won’t have the H200 until the end of 2023 according to his own statements fastest GPU in the world
brought on the market.
Already the predecessor H100 was too much for the trade restrictions and sales to China. So a cropped H800 model came onto the market there. But trading in this GPU has now also been banned.
The following table shows how much slower the still extremely fast H20 GPU is in comparison. What is crucial here is the core discipline of FP8 computing power, which is important for AI calculations:
| H20 | H800 | H100 | H200 | |
|---|---|---|---|---|
| FP8 Tensor Core | 296 TFLOPS | 3.958 TFLOPS | 3.958 TFLOPS | 3.958 TFLOPS |
| FP32 | 44 TFLOPS | 67 TFLOPS | 67 TFLOPS | 67 TFLOPS |
| Storage | 96 GB HBM3 | 80 GB HBM2e | 80 GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | 4.0 TByte/s | 2.0 TByte/s | 3.35 TByte/s | 4.8 TByte/s |
| TDP | up to 400 watts | up to 350 watts | up to 700 watts | up to 700 watts |
| Interconnect | 900 GByte/s | 400 GByte/s | 900 GByte/s | 900 GByte/s |
A high price that almost seems small in comparison
Nvidia itself is holding back with official price information for the AI chips, but there are estimates. This is what the H100 chip is supposed to do sea CNBC ranging from $25,000 to $40,000. The H200 chip is logically a good deal higher.
The H800 models were sometimes even sold for $69,000 in China despite their limitations compared to the H100 (and therefore also the H200) in terms of memory bandwidth and interconnect PC Games Hardware has reported.
As a reminder, H20 is said to be priced at $12,000 to $15,000. In addition to the technical regression, there is also a regression in terms of earning potential for Nvidia.
As Reuters indicates, Nvidia also has to take into account local competition from AI chips from Huawei. More precisely, it is about the Ascend 910B model, which is similarly priced. But either way, you don’t have to worry about Nvidia earning too little.
You can find out what plans Nvidia may have in the gaming area for the RTX-4000 here:
Nvidia may not have pulled all the arrows out of its quiver yet
Background: Even if the basis of Nvidia’s gaming and AI cards is similar, very different demands are placed on them. What is particularly important for AI chips is a large and fast memory as well as the ability to quickly transfer data between individual chips. Gaming GPUs, on the other hand, primarily rely on many shader units and high clock speeds.
What do you think of this news from Nvidia? And what do you think about the fact that US trade restrictions prevent the fastest models from being sold in China? Or asked more prosaically: Do you have a graphics card from AMD or Nvidia in your computer – and which card specifically? As always, please write to us in the comments if you like.
