Replies: 1 comment
-
Hi @gitcnd, yes it will work - Blackwell supports OpenCL too - and can use the full 128GB unified memory capacity since FluidX3D v3.0 update. But I can't yet estimate how fast it will be. So far Nvidia only published useless marketing nonsense. FP4 bit sludge might work for some AI stuff but is useless for anything else; I find it misleading to call that "1 PetaFlop". There is no info out yet on CUDA core count, clock frequency, memory bandwidth. What I know though is that FP4:FP32 ratio on Blackwell is 32:1, so this GP10 GPU will be somewhere at ~31 TFlops/s in FP32. It's basically an RTX 5070 with 128GB LPDDR5X. That makes it just another glorified mini-PC, not an "AI Supercomputer"... Critical question for if it's worth $3k is what memory bandwidth will be. Currently you can get an Intel Meteor Lake mini-PC with 96GB unified DDR5 memory @90GB/s for ~$1400. And M1/M2 Ultra Macs with 128GB @800GB/s are ~5k and non-repairable. Kind regards, |
Beta Was this translation helpful? Give feedback.
-
https://www.nvidia.com/en-us/project-digits/
Beta Was this translation helpful? Give feedback.
All reactions