Vcache only helps when you want to access lots of tiny chunks of data that fit inside the 96-128mb cache.
During inference you have to read the entire several Gb model for each token generation, so your botleneck is still the Ram bandwidth.
Community to discuss about Llama, the family of large language models created by Meta AI.
Vcache only helps when you want to access lots of tiny chunks of data that fit inside the 96-128mb cache.
During inference you have to read the entire several Gb model for each token generation, so your botleneck is still the Ram bandwidth.
In the article they said that that is what was expected but the gains impacted the entire ramdrive and the concept has been proven now. The test used a 500mb+ block so bigger than the cache alone.
https://www.tomshardware.com/news/amd-3d-v-cache-ram-disk-182-gbs-12x-faster-pcie-5-ssd
180GB/s isn't really all that fast.
Maybe, but it's a lot faster than what we can do right now and its only the start.
So there are CPU intrinsics for prefetching data. If we can get better at anticipating the next pieces of data that need to be calculated you can speckle in those preload instructions and achieve more speed.
There are actually TSVs for 3D Cache on the AMD 7900 series, but AMD doesn't use them. Presumably because it makes the chip run hotter, so they'd have to downclock it.
But I think it would be a great candidate for an ML card. Not for directly accelerating models, but for basically fitting any kind of intermediate calculations in cache to preserve all the RAM bandwidth for model weights.