AI Mission: Government changes tender norms to enable low-cost GPUs – Technology News

According to insiders, the Ministry of Electronics and Information Technology (MeitY) has modified the tender norms in a bid to enable the use of low-cost graphics processing units (GPUs) in its Rs 10,000-crore artificial intelligence (AI) mission. .

The move is significant because the compute unit specifications included in the earlier tender would have limited bidders from purchasing Nvidia GPUs such as the A100, H100 and Blackwell B200, among others, which are used for advanced training of large artificial intelligence models. Industry experts said this would greatly limit compute units from third-party companies like AMD and even low-cost options from Nvidia.

Now, based on industry consultations and demand from start-ups, the government has modified the tender norms, lowering parameters such as computational efficiency, memory load and lower energy requirements. For example, in an earlier version of the tender, the AI ​​compute unit specifications required GPUs to have FP32 performance at 15 TFLOPS or higher, FP16 performance at or above 300 TFLOPS, and AI compute memory of 40 GB or more.

The 300 TFLOPS performance requirement for FP16 would limit the qualification of the AMD Instinct MI210 GPU, an industry official said in comments on the tender document to MeitY.

In the revised specifications, the government has lowered some parameters such as 150 TFLOPS for FP16 GPUs and AI compute memory to 24 GB. However, the updated tender includes a clause that at least 70% of the AI ​​computing units offered to meet this criterion should have double precision (FP6) capabilities.

This means that reducing computational performance will pave the way for small-scale tuning and inference, which will reduce energy requirements and costs. Similarly, according to an industry executive, reducing the amount of memory would help reduce costs, as memory is one of the most expensive components of high-performance GPUs.

Of the Rs 10,000-crore IndiaAI mission, MeitY has allocated Rs 5,000 crore to set up AI computing infrastructure in the country. The government has finalized a profitability gap financing model that will see it subsidize computing costs by up to 50% for startups that want to use GPUs. The tender is currently limited to the purchase of 1,000 such GPUs and will involve data center providers, cloud service providers or authorized partners of such solutions providers.

According to officials, the change in GPU specifications occurred because most startups use small-scale tuning and inference to develop and deploy AI and generative AI applications. Additionally, to avoid delays due to supply constraints of high-end Nvidia GPUs, it was necessary to pave the way for more low-cost GPUs.

Under previous specifications, only four to five of the 12 types of GPUs would meet the eligibility criteria. “The more commonly used GPUs were not included in the eligibility criteria, but now several others will also be covered,” the official said.

However, a representative of one of the companies considering submitting to the tender said that the reduced specification leaves little room for other graphics processors.

According to the tender documentation, the selected bidder will have to provide the installed capacity of at least 1,000 AI computing units within six months from the date of signing the contract.

Additionally, empaneled companies are asked to place orders for the anticipated GPUs within three months from the date of empanelment, failing which the government will honor the bank guarantee submitted by the companies along with the application.

For the anticipated 500 AI compute units, the value of the bank guarantee to be provided by the company or consortium will be Rs 50 crore, according to the updated tender document.

Similarly, in a bid to enable greater participation of smaller companies in the AI ​​computing bidding, the government has reduced the average annual turnover of a bidder in the last three financial years to Rs 50 crore from Rs 100 crore earlier.

In case of a consortium, the non-core members of the consortium should have a minimum average annual turnover of Rs 25 crore for the last three fiscal years. The same has been reduced from Rs 50 crore.

October 16 is the deadline for submitting offers in the tender for AI computing infrastructure.