Understand the difference between the versions of NVIDIA A100, A800, H100, and H800 in one article?

A100 A800 40GB video memory 80GB video memory PCIE version and SXM version

H100 H800 80GB memory PCIE version, SXM version NVL version

What the hell is DGX HGX?

Are these confusing? Today, through a comparison of parameters, take a big article to understand the various versions of NVIDIA A100, A800, H100, and H800

A800 NVLink 8 card module A100 Super Micro NV server A100 PCIE single card

H800 Ultra Micro NV Server H100 Ultra Micro NV Server H100 PCIE Single Card

GH200 Super Server

What is the difference between NVIDIA DGX and NVIDIA HGX - Zhihu.com

On October 7 last year, the United States introduced new regulations on semiconductor export restrictions to China , including restrictions on the export of high-performance computing chips to mainland China. And take the performance index of NVIDIA's A100 chip as the limiting standard. That is, a high-performance computing chip that meets the following two conditions at the same time is a regulated high-performance computing chip:

(1) The I/O bandwidth transmission rate of the chip is greater than or equal to 600 Gbyte/s;

(2) The sum of computing power calculated by multiplying the bit length of each operation of the "digital processing unit original computing unit" by TOPS is greater than or equal to 4800TOPS. This also makes it impossible for NVIDIA A100/H100 series and AMD MI200/300 series AI chips to be exported to China.

Subsequently, in order to meet the needs of Chinese customers while complying with the US restriction rules, NVIDIA announced on November 8 that it will launch the A800, a replacement product for the A100 that complies with the new US regulations, and will be put into production in the third quarter of this year. Judging from the officially announced parameters, A800 mainly reduces the transmission rate of NVLink from 600GB/s of A100 to 400GB/s, and other parameters are basically the same as A100.

In March of this year, Nvidia released a new generation of H100 GPU based on 4nm process, with 80 billion transistors and 18432 cores. Similarly, NVIDIA has also launched a special version of the H800 for the Chinese market. "Our 800-series products comply with export control regulations," NVIDIA said in a statement to Reuters. Reuters reported that the H800's chip-to-chip data transfer speed is half that of the H100 . The 800 series is said to have been adopted by the cloud divisions of Alibaba, Baidu and Tencent.

Since the end of last year, with the continued popularity of generative AI represented by ChatGPT, the demand for AI chips based on high-performance GPUs in the generative AI market has skyrocketed . Among them, the powerful NVIDIA AI chip is highly sought after by the market and occupies a monopoly position in the market. In contrast, AMD's AI chip market share is relatively small.

According to statistics, NVIDIA currently sells at least 9 models of AI accelerator cards, including 4 high-performance models, namely V100, A800, A100 and H100. In terms of price, the V100 accelerator card costs at least US$10,000, which is about 69,000 yuan at the current exchange rate; the price of the A800 is US$12,000, which is about 87,000 yuan, and the market once hyped it up to 100,000 yuan; the price of the A100 is 1.5 10,000 US dollars, or about 108,000 RMB; the H100 accelerator card is currently the most powerful NVIDIA, priced at 36,500 US dollars, or about 264,000 RMB.

According to news, due to the soaring market demand, the market price of Nvidia’s replacement versions A800 and H800 for the Chinese market is 40% higher than the original manufacturer’s suggested retail price, and the delivery date of new orders may be extended to December.

Previous revelations also showed that ByteDance, a major Internet company, has ordered NVIDIA GPUs worth about 1 billion US dollars , about 100,000 pieces of A100 and H800, of which A100 should be ordered before the US government ordered in August 2022.

However, according to the latest report in the "Wall Street Journal", the new restrictions being considered by the US Department of Commerce may prohibit the sale of NVIDIA A800 chips without a special US export license to China. So this means that the H800 series will also be limited.

At present, although there are many domestic GPU acceleration chip and AI chip manufacturers in China, such as Biren Technology, Muxi Integrated Circuit, Moore Thread, Haiguang Information, Cambrian, etc., their overall performance is still relatively low compared with NVIDIA and AMD. There is a large gap and cannot replace this part of the market demand.

Although for these domestic AI chip manufacturers, the new US AI chip restriction policy is good news, but for domestic manufacturers that rely on high-performance AI chips to provide AI hardware, as well as Internet manufacturers and some AI technology manufacturers that provide AI services It is a bad news, after all, without the support of powerful AI chips, the development of its AI technology and the AI ​​services it can provide will also be negatively affected.

Nvidia GPU is currently in the supply-side market, the imbalance between supply and demand is serious, and the alternative products of the same kind in the market are not strong, which also makes Nvidia called "the center of the AI ​​​​world".

The A800 graphics card has risen from more than 90,000 yuan a week ago to 130,000 yuan, an increase of about 30%, and the spot price of servers has risen from 1.2 million to about 1.4 million yuan.

According to reports, Baidu, ByteDance, Tencent and Alibaba have ordered a total of US$5 billion worth of AI chips from Nvidia. Among them, Nvidia will ship a total of about 100,000 A800 chips this year, worth $1 billion, and another chip worth $4 billion will be delivered next year.

About purchasing GPU

The GPU market in mainland China is usually divided into different types of Bank of China and OEM: the price difference between each bank of China will not be too large; the price of each type of OEM will have a certain gap; the market usually uses OEM products to compete with Bank of China, Both products can be chosen, and it is recommended to choose Bank of China.

In addition, taking the A100 40GB as an example, it is not only divided into PCIE version and SXM version, but due to the interruption of supply, there are also cases of removing old cards and reorganizing cards on the market, and the price difference is also very large.

The price of the demolished products is mainly determined based on the quality, use time, production time, output, etc. The price gap is relatively large, and it is also much lower than the price of the Bank of China.

The author encountered the situation that the user did the research in the early stage and removed the old card to compare with the author's Bank of China card. The user would feel that the author's price was very high and there was no sincerity. In fact, the author is engaged in the subdivision of scientific research servers , owns the brand of scientific research servers, and uses brand-new components of the first brand in production, so the comparison is very unfair.

If the user is just looking for the price, it can also be provided for the user if the specified requirement is demolition and reorganization.

A100\H100 is basically less and less in mainland China, and A800 is currently making way for H800. If you really need A100\A800\H100\H800GPU, it is recommended not to be picky. For most users, the difference between HGX and PCIE version is not It's very big, and you can buy it as soon as it's in stock.

In any case, choose regular brand manufacturers to cooperate . Under the current market situation where supply and demand are out of balance, most merchants in the market cannot supply, and even provide untrue information. If it is a scientific research server, Fenghu Yunlong scientific research server is the first choice. Mining, quality and after-sales service are guaranteed.

Welcome to communicate with  Manager Chen【173-1639-1579】

Has been focusing on scientific computing servers for many years, shortlisted for political mining platforms, H100, A100, H800, A800, RTX6000 Ada , a single dual-socket 192 core server is available for sale .

Guess you like

Origin blog.csdn.net/Ai17316391579/article/details/132627201