Mar. 4 at 9:09 PM
Rumour:
$META has been actively pursuing custom silicon development to support its AI training & inference needs, particularly for models like Llama.
However, they are having challenges in its in-house efforts & scrapped its most advanced full-scale AI chip design for training (code-name Olympus) last week due to unspecified issues
Training chips require massive scale, high memory bandwidth (eg: advanced HBM integration), extreme interconnects, & precise power/thermal mgemt - Meta struggled w/ these complexities, leading to redesign risks, potential delays, & skepticism internally about matching Nvidia's performance.
This is the 2nd training-focused design Meta has abandoned
Note:
$AVGO provides key IP, design expertise, & co-design support for MTIA v1/v2 generations which is focused on inference & lighter tasks
Meta is seeing significant cost & power savings (up to 60% in some cases) w/ its MTIA chips
