MoE is strictly worse for folks using discrete consumer GPUs like a 3090 AFAICT. It seems pretty good for unified memory stuff like Macs though.
MoE Performance: Discrete GPUs vs Unified Memory Architecture
By
–
By
–
MoE is strictly worse for folks using discrete consumer GPUs like a 3090 AFAICT. It seems pretty good for unified memory stuff like Macs though.