India Can Build Indigenous GPU by 2029, Says C-DAC Bengaluru Chief

India can develop an indigenous GPU as a physical chip within the next two to three years, believes S D Sudarsan, Executive Director of the Centre for Development of Advanced Computing (C-DAC) Bengaluru.

In an interview with AIM, he shared that C-DAC is ramping up work on a homegrown computing stack spanning processors, GPUs, AI accelerators, and servers.

The centre is developing its own GPU and AI accelerator architecture to produce a chip capable of handling AI workloads that are relevant by the end of the decade.

“We are predicting that we’ll likely have GPU specialisation by= 2029. We will have that kind of a specification from a capability perspective; we are building on that,” Sudarsan said.

He added that the target is to develop an indigenous chip comparable to other GPUs available in the market at the time.

For the upcoming GPU, however, C-DAC is deliberately avoiding locking the design to today’s AI models. Sudarsan said the centre expects AI workloads and model architectures to change considerably over the next three years.

The organisation is therefore developing both a GPU for more demanding parallel workloads and a simpler AI accelerator for workloads that do not necessarily require a conventional GPU.

“One is for a GPU type of workload. The other one is an AI accelerator, which has a much more simpler design and won’t be as complex as a traditional GPU,” he said.

The immediate milestone is to create an indigenous GPU chip. “Very soon we can have a technology demonstrator of our own GPU as a physical chip in front of you,” he said.

The GPU effort is part of a broader attempt by C-DAC to build an indigenous high-performance computing ecosystem rather than relying entirely on foreign processors and accelerators. 

C-DAC already has the Rudra series of servers, with platforms based on Intel and AMD technologies. Rudra-I uses Intel Xeon processors, while the newer Rudra-GN platform uses AMD EPYC processors and AMD Instinct MI300X accelerators.

C-DAC is also working towards RISC-V-based server platforms. Its VEGA processor family, developed under the Digital India RISC-V (DIR-V) programme, includes 32-bit and 64-bit single-, dual- and quad-core processors based on the RISC-V instruction set architecture. 

The Ministry of Electronics and Information Technology launched the DIR-V programme in 2022 to build homegrown, open-source microprocessors.

“By 2030, we'll have a DIR-V-based HPC (high-performance computing) system, which is completely indigenous,” Sudarsan said.

Sudarsan said C-DAC expects to eventually have four server variants spanning x86 and RISC-V architectures. Two would use x86 technologies from Intel and AMD, while two would be based on RISC-V, with one variant using a fully Indian IP stack.

C-DAC is also working on chiplets based on its VEGA IP, covering CPUs, GPUs, and AI accelerators, according to Sudarsan.

India Building the Software Layer

C-DAC’s push for indigenous computing extends beyond silicon. Sudarsan said the organisation has continued developing the system software stack for its supercomputers, including compilers, integrated development environments, profilers, linkers, and loaders.

“Throughout the journey from 1988 till date, the system software stack has always been the compilers, the IDEs, as they call them, the profilers, the linkers, the loaders. They all have been always done by C-DAC, and it has always been indigenous for the supercomputers,” he said.

C-DAC has developed and taped out multiple RISC-V-based processors and SoCs under the government’s Microprocessor Development Programme. 

The organisation has also developed DHRUV64, a 1GHz 64-bit dual-core processor-based SoC that C-DAC describes as India's first fully homegrown dual-core processor-based SoC. DHRUV64 is the third chip based on the VEGA family, following THEJAS32 and THEJAS64. The chip integrates C-DAC's VEGA AS2161 dual-core processor and targets embedded and compute-intensive applications.

Sudarsan said these processors and boards are already being tested through real-world applications. 

“There are use cases of drone-based applications. There are use cases of image processing applications. We have also seen radar-type, automobile-based applications,” he said.