←back to thread

Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference

(cerebras.ai)
426 points benchmarkist | 4 comments | 19 Nov 24 00:15 UTC | HN request time: 0.001s | source
1. easeout ◴[19 Nov 24 04:35 UTC] No.42180121[source]▶
How does binning work when your chip is the entire wafer?
replies(1): >>42180161 #
2. shrubble ◴[19 Nov 24 04:47 UTC] No.42180161[source]▶
They expect that some of the cores on the wafer will fail, so they have redundant links all throughout the chip, so they can seal off/turn off any cores that fail and still have enough cores to do useful work.
replies(1): >>42180731 #
3. why_only_15 ◴[19 Nov 24 07:05 UTC] No.42180731[source]▶
My understanding is that they mask off or otherwise disable a whole row+column of cores when one dies
replies(1): >>42180846 #
4. wtallis ◴[19 Nov 24 07:29 UTC] No.42180846{3}[source]▶
That's way too wasteful.

Take a look at https://fuse.wikichip.org/news/3010/a-look-at-cerebras-wafer... and specifically the diagram https://fuse.wikichip.org/wp-content/uploads/2019/11/hc31-ce...

The fabric can effectively route signals diagonally to work around an individual defective core, with a displacement of one position for cores in the same row from that defect over to the nearest spare core. That's how they get away with a claimed "1–1.5%" of spare cores.