Hacker News new | past | comments | ask | show | jobs | submit login

Pretty obvious stuff, right? I mean, you don't even need HBM for that, you just need a TON of memory channels. Sure, that kind of setup would only be efficient for highly coalesced reads/writes, but that's what you need these days for inference - highly coalesced reads and writes. You could even get by with 64GB of DDR5. DDR5-4800 (rather modest) is 38.4GB/s per channel. To get 1TB/s you'd only need 26 channels. With the more expensive DDR5-6400 you'd only need 20. That doesn't at all sound insurmountable for a company of Intel's caliber. Heck, break up the dies (and the channels) across several chiplets even, if the interconnect is decent it'll still run really well.



Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: