Gearbest Tech
[Gearbest News]In the past two years, with the popularity of AI and large model training, related companies also need stronger computing power to meet the growing computing needs. In this context, NVIDIA's professional-grade A100 graphics card is highly sought after by the market, but due to special reasons, this graphics card cannot be sold to domestic areas. Therefore, last year NVIDIA provided the domestic market with a brand new A800 graphics card. In order to meet export regulations, this product has been castrated in terms of bandwidth, so it is also called the “special version”.
Due to the reduction in bandwidth, it also has a certain gap in performance compared to A100. It is understood that the A800 graphics card will have three different versions, namely the 40GB HBM2 thread version using the PCIe interface, the 80GB HBM2e video memory version and the 80GB HBM2e video memory SXM version, to meet the needs of different customers and environments.
In terms of performance, the data transfer rate of A800 is 400GB per second, which is lower than the 600GB per second of A100, and supports a memory bandwidth of up to 2TB/s. In addition, the computing performance of 9.7TFlops FP64 double precision, 19.5TFlops FP32 single precision, 1.6-2.0TB/s memory bandwidth, up to 7 instances per card, support for single card to eight cards in parallel, PCIe 4.0 system bus, 250-400W thermal design power consumption (PCIe is dual-slot air cooling or single-slot liquid cooling) is no different from A100.
However, there was news recently that this “special edition” graphics card will also face the threat of being banned from sale. As soon as this news came out, the price of A800 skyrocketed, and now it has reached 110,000 yuan per piece. In just one week, compared with the previous price of 90,000 yuan, the price has increased by 20%-30%.
The reason for such an increase may be related to its insufficient production capacity. According to He Ning, vice president of H3C Group and general manager of the operator business unit, in an interview during the Shanghai MWC exhibition, the supply of high-end servers equipped with NVIDIA chips is still very tight, with a delivery cycle of about 30 weeks.
Since the beginning of this year, various large models such as ChatGPT have been continuously launched, and domestic related large model software has also emerged. It seems that there is a trend of catching up with foreign related software. However, with NVIDIA's insufficient GPU production capacity and the news of the ban on sales, it is inevitable that the speed of progress of some companies will be hindered, which may affect the development prospects of domestic large model software under certain circumstances.
magicCubeFunc.write_ad(“dingcai_top_0”);
Login with Google