3 ms·
Inventec 96 DIMM CXL Expansion Box at OCP Summit 2024 for TBs of Memory
- magicalhippo 2y agoCurious about the latency of this. This[1] page, which is from 2022, seems to indicate that the CLX memory controllers adds about 200ns, and you got some retimers which add a few tens of ns. Then there's the cabling, assuming 2/3 the speed of light that's 5ns per meter. With the memory latency itself it seems this adds up about 350ns, while a local memory hit is about 100ns[2][3]. For random access workloads this is certainly better than not having stuff in memory, but I guess for streaming workloads a sled full of NVMe drives might be just as good. [1]: https://www.nextplatform.com/2022/12/05/just-how-bad-is-cxl-memory-latency/ https://www.nextplatform.com/2022/12/05/just-how-bad-is-cxl-... [2]: https://wccftech.com/amd-epyc-9654-96-core-genoa-cpu-tears-apart-milan-x-sapphire-rapids-cache-latency-benchmarks-up-to-30-tbps-over-2x-performance-increase/ https://wccftech.com/amd-epyc-9654-96-core-genoa-cpu-tears-a... [3]: https://arxiv.org/abs/2204.03290 https://arxiv.org/abs/2204.03290