Researcher Questions Hardware Needs for 3T-Parameter Kimi 3
Zephyr suggests Huawei SuperPoDs or NVL72 systems are required.
Entities: François Fleuret, Huawei
AI researcher François Fleuret asked on X what hardware can run Kimi 3 well, saying he had read that 8 H100s might do it but questioning whether that sounds like enough for a model with roughly 3 trillion parameters. A reply on X from @zephyr_z9 pointed instead to "Huawei SuperPoDs or NVL72s." That leaves the public conversation centered on an attributed hardware question, not a settled account of Kimi 3's actual deployment needs.
Combined views
49.3K
5 posts, first seen 3h ago
Reactions from ranked influencers
5 posts
What hardware can run Kimi 3 well? I read 8 H100s but that seems not much for ~3T parameters, no?
Huawei SuperPoDs or NVL72s
What hardware can run Kimi 3 well? I read 8 H100s but that seems not much for ~3T parameters, no?
@francoisfleuret likely intended system
@francoisfleuret you can get pretty far away with 2-bit quant, SSD streaming, custom kernels and other tricks, see e.g. https://github.com/antirez/ds4
It's kinda fun that "pod" is becoming de facto industry standard for "bunch of highly connected accelerators" AFAIK "TPU Pod" was an ocean pun (a group of whales/dolphins) because TPUs have ocean animal codenames (jellyfish, dragonfish, ...) cc @JeffDean correct?
After the K3 drop, Huawei has now released a real 950 SuperPoD 1 EFLOPS fp8 & 2 EFLOPS fp4 256TB of unified memory