MeitY halves revenue threshold in GPU tender to Rs 50cr

The Ministry of Electronics and Information Technology (MeitY) has tweaked tender norms in a bid to allow the use of low-cost graphic processing units (GPUs) under the Rs 10,000-crore Artificial Intelligence (AI) mission, according to people in the know.

The move is significant as the specifications of compute units under the earlier tender would have restricted bidders from procuring Nvidia GPUs like A100, H100, and Blackwell B200, among others, which are used for high-end training of large AI models. It would have also largely restricted compute units of other companies such as AMD, and even low-cost options from Nvidia, industry experts said.

Now, based on industry consultations and startups’ demand, the government has tweaked the tender norms by downgrading parameters such as computational performance, memory workloads, and lower power requirements. For example — in the earlier version of the tender, the specifications for AI compute units mandated GPUs to have performance for FP (floating point) 32 with 15 TFLOPS (tera floating point operations per second) or above, performance for FP16 with 300 TFLOPS or above, and AI compute memory of 40 GB or above.

The asked performance of 300 TFLOPS for FP16 would restrict AMD Instinct MI210 GPU to qualify, an industry executive said in comments on the tender document to MeitY.

In the changed specifications, the government has reduced some parameters such as 150 TFLOPS for FP16 GPUs and AI compute memory to 24 GB. However, the updated tender has a clause that a minimum of 70% of AI compute units offered towards meeting this criterion should have double precision (FP64) capabilities.

This means a reduction in computational performance will pave the way for small-scale tuning and inference, which would lower power requirements and costs. Similarly, a reduction in memory would help cut costs since memory is one of the most expensive components in high-performance GPUs, according to an industry executive.

Of the Rs 10,000-crore IndiaAI mission, MeitY has earmarked Rs 5,000 crore to create AI compute infrastructure in the country. The government has finalized a viability gap funding model, wherein it will subsidize compute costs up to 50% for startups, which want to utilize GPUs. The tender is currently limited to procuring 1,000 such GPUs, which will see participation from data center providers, cloud services providers, or authorized partners of such solution providers.

According to officials, the change in GPU specifications has been made as a majority of startups are using small-scale tuning and inference for developing and deploying AI and generative AI applications. Also, in order to avoid delays owing to supply constraints of Nvidia high-end GPUs, it was essential to pave the way for more low-cost GPUs.

As per the earlier specifications, only four to five out of 12 types of GPUs would have met the eligibility criteria. “More commonly used GPUs were not covered in the eligibility criteria, but now a few more would also get covered,” an official said.

A representative of one of the companies considering an application for the tender, however, said the reduced specifications give only a little room for other GPUs.

As per the tender document, the bidder selected would have to make available an installed capacity of a minimum of 1,000 AI compute units within six months from the date of signing of the agreement.

Additionally, the empaneled companies are asked to submit purchase orders for the anticipated GPUs within three months of empanelment, or else the government will encash the bank guarantee submitted by companies with the proposal of application.

For 500 anticipated AI compute units, the bank guarantee value to be submitted by a company or consortium will be Rs 50 crore, as per the updated tender document.

Similarly, in a bid to allow more participation from smaller companies in the AI compute tender, the government has reduced the average annual turnover in the last three financial years of the bidder to Rs 50 crore from Rs 100 crore earlier.

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

The last date for submission of bids for the AI compute infra tender is October 16. Financial Express