
Prospective: A brand new report reveals the huge array of Nvidia GPUs utilized by Microsoft and the improvements it has taken in arranging them to assist OpenAI practice ChatGPT. The information comes as Microsoft introduced a significant improve to its AI supercomputer to additional its homegrown generative AI initiatives.
According to Bloomberg, OpenAI educated ChatGPT on a Microsoft supercomputer constructed from tens of hundreds of Nvidia A100 GPUs. Microsoft this week introduced a brand new array utilizing Nvidia’s up to date H100 GPU.
The problem for these corporations started in 2019, when Microsoft invested $1 billion in OpenAI and agreed to construct an AI supercomputer for the startup. However, Microsoft didn’t have the interior {hardware} to fulfill OpenAI’s wants.
After buying Nvidia’s chips, Microsoft needed to rethink the right way to organize such a lot of GPUs to forestall overheating and energy loss. The firm wouldn’t disclose the precise value of the hassle, however Executive Vice President Scott Guthrie estimated the determine to be within the a whole bunch of hundreds of thousands of {dollars}.
Also learn: Is Nvidia Winning the AI Training Market?
Running all of the A100s on the identical time compelled Redmond to consider the right way to place them and their energy provides. It should additionally develop new software program to enhance effectivity, guarantee community tools can stand up to the deluge of information, design new cable trays that may be manufactured independently, and use quite a lot of cooling strategies. Depending on the altering local weather, cooling applied sciences embrace evaporative, swamp coolers and outdoors air.
Since ChatGPT’s preliminary success, Microsoft and a few of its rivals have begun growing parallel AI fashions for engines like google and different purposes. To speed up its generative AI, the corporate launched the ND H100 v5 VM, a digital machine that may use wherever from eight to hundreds of Nvidia H100 GPUs.
The H100s are related by way of NVSwitch and NVLink 4.0 with 3.6TB/s bisected bandwidth between every of the 8 native GPUs in every digital machine. Each GPU has 400 Gb/s bandwidth by way of Nvidia Quantum-2 CX7 InfiniBand and 64GB/s PCIe5 connection. Each VM manages 3.2Tb/s over non-blocking fat-tree networking. Microsoft’s new system can be geared up with 4th technology Intel Xeon processors and 16 channels of 4800 MHz DDR5 reminiscence.
Microsoft plans to make use of the ND H100 v5 VM for its new synthetic intelligence Bing search engine, Edge net browser and Microsoft Dynamics 365. The digital machine is on the market in preview now and can turn into commonplace throughout the Azure portfolio. Potential customers can request entry.