Meta has done something that will get Nvidia and AMD very, very worried — it gave up on GPU and CPU to take a RISC-y route for AI training and inference acceleration
The RISC-V movement has received a huge boost after Meta revealed that it is ready to begin massive development of technology-centric hardware and products.
RISC, an alternative Instruction Set Architecture (ISA) to the x86 architecture used by the likes of Intel and AMD in their top CPUs, is dominated by Arm. This ISA is largely capable of replicating many of the functionalities of the best GPUs and processors out there.
RISC-V is an open source alternative that is attractive because it lowers the barrier to entry due to its open nature and the fact that organizations using it do not have to pay fees to license the use of CPUs. In recent months, companies have been out Alibaba.com Unpleasant Qualcomm have thrown their support behind the RISC-V ecosystem and announced plans to migrate in varying degrees.
Now Meta is migrating from CPUs to RISC-based components due to the ISA's energy efficiency, performance, lower latency, and flexibility to support different workloads. This is according to the senior technical director, who spoke at the RISC-V Summit, as reported by Next platform.
Meta's bet on RISC-V is in a phase of acceleration, after four years of planning, and the company is not only rolling out production hardware but also preparing to produce future custom RISC-V silicon .
For example, the company built some of its video transcoding hardware on RISC-V alone, replacing 85% of its CPUs. Known as the Meta Scalable Video Processor (MSVP), this package is already deployed and handles all video uploads across a range of social media services, including Facebook, Messenger and Instagram.
But the company is also eyeing RISC-V for use in developing a chip that could be used in AI training and inference, at a time when most of the industry is clamoring to get its hands on Nvidia's leading GPUs to get.
The A100, followed by the H100 and H200 chips have been crucial to the rise of generative AI, but Meta plans to skip this phase altogether and instead throw its weight behind RISC-V when it comes to building AI processors.
Meta's first internal RISC-V silicon for AI acceleration is a 7nm component that operates at a frequency of 800MHz, has 128MB of on-chip memory and supports up to 128GB of LPDDR5 RAM. They are currently committed to accelerating models in both training and inference.