Raja Koduri, a GPU expert known for his work on graphics processors at companies like AMD, Apple, ATI, S3 Graphics, and Intel, believes that custom-designed silicon for artificial intelligence (AI) and high-performance computing (HPC) will not replace GPUs entirely. He is confident that there is room for new designs tailored specifically for these tasks even beyond GPUs. Companies like AWS, Graphcore, Google, Microsoft, and Tesla are actively exploring custom silicon solutions, which align with Koduri’s perspective on the evolving landscape of architectures.
In response to an article discussing the future of AI, Koduri emphasized the continued dominance of GPUs, stating, “We have been hearing about this since 2016, yet GPUs remain at the forefront.” He pointed out the limitations of purpose-built silicon, highlighting the inherent instability of AI despite some portraying it as a simple operation of matrix multiplications.
While GPUs initially focused on accelerating graphics workloads, they have significantly evolved to enhance AI and HPC tasks, particularly due to Nvidia CUDA’s widespread adoption. The adaptability of Nvidia GPUs in accommodating new data types has made it challenging for competitors like AMD and Intel to keep pace in the AI and HPC domains.
Koduri underscored the importance of robust system software support in silicon solutions, noting the reliance on software developers to address structural deficiencies in purpose-built designs. He expressed concerns about the shortage of new talent in system software development, emphasizing the need for a fresh influx of expertise to drive innovation in this space.
Despite his involvement with Tenstorrent, a company specializing in custom AI accelerators and HPC CPUs, Koduri remains optimistic about the potential for future architectures that are more attuned to specific workloads. He envisions leveraging the insights gained from current challenges to shape next-generation structures that could surpass the limitations of existing silicon and GPUs.
In conclusion, Raja Koduri’s forward-looking perspective hints at the emergence of novel architectures tailored to the evolving demands of AI and HPC, suggesting that the industry is on the cusp of transformative advancements in silicon design.
