NVIDIA starts pre-ordering new AI chip 'H20' exclusively for China, price is about 1.8 million to 2.2 million yen
Semiconductor companies such as NVIDIA
Exclusive: Nvidia's new China-focused AI chip set to be sold at similar price to Huawei product | Reuters
https://www.reuters.com/technology/nvidias-new-china-focused-ai-chip-set-be-sold-similar-price-huawei-product-2024-02-01/
New Nvidia AI GPUs designed to get around US export bans come to China — H20, L20, and L2 to fill void left by restricted models | Tom's Hardware
https://www.tomshardware.com/pc-components/gpus/new-nvidia-ai-gpus-designed-to-get-around-us-export-bans-come-to-china-h20-l20-and- l2-to-fill-void-left-by-restricted-models
NVIDIA officials told Reuters that the company has opened pre-orders for the H20 from retailers with pricing comparable to Huawei's competitors. The H20 is the model with the highest specifications among the three graphics board models for the Chinese market developed by NVIDIA in order to avoid the US export restrictions on AI chips.
The H20 is a graphics board with lower processing power than the H100, NVIDIA's main AI chip, and the H800, which NVIDIA was selling for the Chinese market but whose export was banned in October 2023. According to anonymous sources, the H20's specs are lower than its rival, Huawei's Ascend 910B, in some key areas.
According to another source, NVIDIA seems to be setting the order price for H20 distributors in China at $12,000 to $15,000 (approximately 1.8 to 2.2 million yen) per graphic board. Another source said that some distributors have begun pre-orders for the H20 with a significant discount to approximately 110,000 yuan (approximately 2.2 million yen). In addition, the competing product Ascend 910B is sold for 120,000 yuan (about 2.5 million yen), so H20 seems to have a slight advantage in terms of price.
Reuters has asked NVIDIA if pre-orders for H20 have started, but the company declined to comment.
According to a source, an H20 distributor is selling a server with eight H20s installed for 1.4 million yuan (approximately 29 million yen). In contrast, a server equipped with eight H800 AI chips for China that NVIDIA previously sold was sold for approximately 2 million yuan (approximately 41 million yen) as of January 2023. Ta.
H20 distributors have announced to customers that they will start shipping H20 in small quantities in the first quarter of 2024 (January to March), and large quantities of H20 will be available from the second quarter (April to June). It seems that it is explained that it will be the month of the month.
Before the US government began regulating it, NVIDIA had over 90% share of China's AI chip market. However, due to the US government's export control measures, NVIDIA is facing a fierce battle for market share with Chinese manufacturers led by Huawei.
Huawei's Ascend 910B is considered to be the most competitive AI chip available in China at the time of writing, and coupled with the US government imposing export restrictions on AI chips, Ascend 910B in China 910B popularity is rapidly increasing. Baidu, the Chinese search giant, was also reported to have switched from NVIDIA to Huawei AI chips.
It is reported that Baidu is ordering AI chips from Huawei and switching from NVIDIA chips - GIGAZINE
by simone.brunozzi
H20 was originally scheduled to be released in November 2023, but that plan has been delayed . It is rumored that the delay in the release of the H20 was due to server manufacturers having problems integrating the chip.
In addition to H20, NVIDIA also plans to roll out chips such as L20 and L2 for the Chinese market. However, Reuters said that it was unable to obtain detailed information on L20 and L2, and since H20, L20, and L2 are all products for the Chinese market, NVIDIA's official website does not have a page for these products. Not.
However, Tom's Hardware reports that the specifications of H20, L20, and L2 will be as follows.
| HGX H20 | L20 PCIe | L2 PCIe | |
|---|---|---|---|
| architecture | Hopper | Ada Lovelace | Ada Lovelace |
| memory | 96GB HBM3 | 48GB GDDR6 w/ECC | 24GB GDDR6 w/ECC |
| memory bandwidth | 4.0TB/s | 864GB/s | 300GB/s |
| INT8/FP8 Tensor | 296/296TFLOPS | 239/239TFLOPS | 193/193TFLOPS |
| BF16/FP16 Tensor | 148/148TFLOPS | 119.5/119.5TFLOPS | 96.5/96.5TFLOPS |
| TF32 Tensor | 74TFLOPS | 59.8TFLOPS | 48.3TFLOPS |
| FP32 | 44TFLOPS | 59.8TFLOPS | 24.1TFLOPS |
| FP64 | 1TFLOPS | N/A | N/A |
| RT Core | N/A | Yes | Yes |
| M.I.G. | up to 7MIG | Yes | Yes |
| L2 cache | 60MB | 96MB | 36MB |
| media engine | 7NVDEC 7NVJPEG | 3NVENC(＋AV1) 3NVDEC 4NVJPEG | 2NVENC(AVI) 4NVDEC 4NVJPEG |
| TDP | 400W | 275W | TBD |
| form factor | 8-way HGX | 2-slot FHFL | 1-slot LP |
| interface | PCIe Gen5x16: 128GB/s | PCle Gen4x16: 64GB/s | PCle Gen4x16: 64GB/s |
| NVLink | 900GB/s | N/A | N/A |
Related Posts:
in Hardware, Posted by logu_ii