Silicon surge: Meta’s strategic spending on hardware development
Friday 26 May 2023
Portfolio insights
Meta invests heavily in hardware infrastructure to support its extensive video content and AI workloads. We wrote about the scale of some of these investments in November last year in Meta – opportunity for contrarian long-term value investors. These investments are significant, and we expect Meta to spend around $36 billion in 2023 on R&D, with a meaningful portion of this going toward developing its in-house silicon. We highlight two of the recent developments below.
MSVP
The first of these developments is the Meta Scalable Video Processor (MSVP). This chip can process videos 9x faster than traditional software encoders while maintaining the same quality and using half the energy. It’s been co-designed with Broadcom and can transcode up to 4K resolution at different frame rates. The MSVP is crucial for platforms like Instagram and Facebook, where billions of videos are uploaded and watched daily. Not only will this chip benefit users, particularly those in countries with slower internet speeds, but it frees up compute resources at Meta, making its infrastructure far more efficient. You can read more about the driving force behind MSVP at Meta’s engineering blog.
MTIA
Meta is also the largest buyer of Nvidia’s H100 GPUs, as it requires GPUs for both AI training and inference. These systems are state-of-the-art. However, they are expensive and general-purpose, meaning they may not perform as well on specific tasks as a specialised chip. This is why the company has developed the Meta Training and Inference Accelerator (MTIA), a specialised chip for Deep Learning Recommendation Model inference (DLRM) – a critical aspect of its AI workload. Like the MSVP, these chips will improve the user experience and increase hardware efficiency. While the first-generation MTIA does not perform as well as Nvidia’s latest GPUs in some workloads, it’s a promising start towards building more specialised hardware for Meta’s unique requirements.
Plans for the MTIA involve enhancements for supporting larger, more complex models, with the second-generation MTIA being fabricated on TSMC’s 5nm technology and codenamed Artemis.
Meta’s focus on enhancing its hardware infrastructure to support its growing video processing and AI needs is a testament to its commitment to innovation and forward thinking. By investing heavily in these areas, Meta is ensuring its own growth and shaping the future of technology.
About the Author
