Meta reveals AI chip (3x faster than before)

Meta has launched the next-gen version of its AI chip, called Meta Training and Inference Accelerator (MTIA), previously known as Artemis, to power Meta’s ranking and recommendation ad algorithms on Facebook and Instagram, providing higher-quality recommendations to users.

What will Meta’s new AI chip be used for?

MTIA v2 will make the training of these algorithmic models more efficient and has the potential to be used, in the future, to train Meta’s generative AI Large Language Model, Llama

The first-gen AI chip, MTIA v1 (which was revealed last year) was designed in-house and was built for use in data centers.

MTIA v2 will also target data centers, with the chips already live and in use across 16 data center regions. Reports show that MTIA v2 is already performing 3x better than its predecessor, with Meta claiming that the “new version of MTIA more than doubles the compute and memory bandwidth of our previous solution”

Why has Meta developed this new AI chip?

Meta is reportedly planning to spend around $18B this year, on AI hardware, to reduce its reliance on NVIDIA, which currently controls 90% of the AI chip market. The release of MTIA v2 aligns with this strategy, although they have revealed concerns that, although “there’s a need for these chips”, they may struggle to “keep up with the advancements from the likes of NVIDIA”

But Meta isn’t the only one looking to reduce reliance on NVIDIA by developing their own supply of AI chips. Intel has recently revealed its newest chip–Gaudi 3–which is reportedly faster and more cost-effective than the NVIDIA equivalent. Google has released Axion, its first dedicated chip for running AI models. Amazon already has several custom chips under its belt, and Microsoft entered into the chip race last year with the Azure Maia AI Accelerator and the Azure Cobalt 100 CPU.

‍