Comment by digiown
15 hours ago
I assume the use case is that you are an inference provider, and you put a bunch of models you might want to serve in the HBF to be able to quickly swap them in and out on demand.
15 hours ago
I assume the use case is that you are an inference provider, and you put a bunch of models you might want to serve in the HBF to be able to quickly swap them in and out on demand.
I think the hope is to run directly off of HBF directly, to eventually replace RAM with it entirely. 1.5TB/s is a pretty solid number! It's not going to be easy, it doesn't just drop in and replace (vastly bigger latency) but HBF replacing HBM for gobs of bandwidth is the intent, I believe.
Kioxia & Nvidia are already talking about 100M IOps SSD's directly attached to GPUs. This is less about running hte model & more about offboarding context for future use, but Nvidia is pushing KV cache to ssd. And using BlueField-4 which has PCIe on it to attach SSDs, process there. https://blocksandfiles.com/2025/09/15/kioxia-100-million-iop... https://blocksandfiles.com/2026/01/06/nvidia-standardizes-gp... https://developer.nvidia.com/blog/introducing-nvidia-bluefie...
We've already deepseek running straight off NVMe, weights runnig there. Slowly, but this maybe could scale. https://www.reddit.com/r/LocalLLaMA/comments/1idseqb/deepsee...
Kioxia for example has AiSAQ, which works in a couple places such as Milvus; not 100% clear but me exactly what's going on there, but it's trying to push work to the NVMe. And with NVMe 2.1 having computational storage, I expect we see more pushing work to the SSD.
These aren't directly the same thing as HBF. A lot is caching, but also, I tend to think there is an aspiration of trying to move some work out of ram, not merely to be able to load into ram faster.
Flash has limited write cycles. The faster you write, the faster it wears out. How do you overcome that?
they will probably use a simpler more direct protocol than NVMe