00:00This is Meta's chip lab in Fremont, California.
00:03Inside, the company is developing the next generations of MTIA,
00:07short for Meta Training and Inference Accelerator, its in-house AI chip program.
00:11It's a long-term effort to build the most efficient architecture for Meta's internal
00:15workloads, with four new generations of chips planned over the next two years,
00:19from ranking and recommendations to large-scale Gen AI inference.
00:23When chips come in from the fab, this is where they're validated,
00:26tested at the chip, rack, and workload level before deployment into Meta's data centers.
00:31MTIA 300 is already in production, supporting ranking and recommendations training,
00:36helping decide what shows up in your social feed.
00:39Those go into liquid-cooled servers like these.
00:42MTIA 400 is moving towards deployment, expanding into broader AI workloads, including Gen AI.
00:48Future versions 450 and 500 push further into Gen AI inference, with deployments planned in 2027.
00:55The effort hasn't always moved as quickly as Mark Zuckerberg and Meta had hoped.
01:00Meta has made some acquisitions, it has tried to make some others,
01:04in an effort to strengthen its in-house chip talent and accelerate progress.
01:08AI models are evolving faster than traditional chip cycles, so Meta is speeding up the design process,
01:14aiming to improve performance, cost, and power efficiency at scale.
01:18At the same time, the company is striking major supply deals with leading chip makers,
01:22securing gigawatts of AI computing capacity.
01:25The strategy is buy compute at scale from NVIDIA and AMD,
01:29but also use custom silicon where Meta's workloads are uniquely its own.
01:33Because in the AI race, it isn't just about the models, it's about the compute behind them.
Comments