5 Comments
User's avatar
Hellish.TX's avatar

Waste of sand.

Expand full comment
Ricardo Bánffy's avatar

The partitioning mechanism seems like a toned down version of IBM’s Telum cache architecture

Expand full comment
Chester Lam's avatar

It's really not the same. From a conversation with IBM folks it sounds like they do it based on capacity pressure on each cache slice, inferred by how often evictions are happening. Then a slice's capacity gets different core-private/shared splits depending on how much cache capacity the associated core needs.

Huawei appears to do it based on cacheline state, not capacity pressure.

Expand full comment
Peter W.'s avatar

I'd be curious what you heard about the Kunpeng 950, which is the most current one from Huawei. To me, the 920 was more of an early dress rehearsal that tried out fancy packaging. And, I'd also be curious if Huawei was able to replace CoWoS; or can they still use TSMC for packaging?

Expand full comment
Peter W.'s avatar

And somehow this important sentence was omitted: Thanks Chester for another great deep dive into a processor we here in the West don't know or see much around!

Expand full comment