Ask HN: Does Anyone Use DGX Spark for High Performance Scientific Computing?
2 comments
Can it even be used that way with only fp4 power?
I believe it should be capable of doing 15TFLOPS (Teraflops per second) in FP64.
Edit: My source is techpowerup (not official) https://www.techpowerup.com/gpu-specs/gb10.c4342
Edit: My source is techpowerup (not official) https://www.techpowerup.com/gpu-specs/gb10.c4342
It is also pretty much accepted in the HPC community that accelerators are necessary, but the complexity of heterogeneous programming, necessity of careful design to avoid data transfer bottlenecks and the cost of accelerators are preventing commonplace adoption.
However, even though newer CUDA versions and devices like DGX Spark are addressing these concerns to significant degree, I haven't seen any benchmarks, any results or any declaration of interest from the community. If you are part of the HPC community, please share your opinions or experience.