AMD has had unified memory for ages in HPC and for a while now in the Strix Halo systems. I haven't had the chance to play with one yet, but I have high hopes for some of our complex simulation workloads.
Re:Prefix sum: 20 GB/s (2.6x baseline)
by ashtonsix on Tuesday October 14, 2025 @03:04PM
(#45583550)
Oh neat. I have some related unpublished SOTA results I want to release soon: PEF/BIC-like compression ratios, with faster boolean algebra than Roaring Bitsets.
The shared memory architecture doesn't eliminate copying the data across to the device. Edit: or back.
Re:Prefix sum: 20 GB/s (2.6x baseline)
by yogishbaliga on Tuesday October 14, 2025 @06:23PM
(#45585756)
If it is unified memory, CPU can access the result of GPU processing without copying it to CPU memory (theoretically)
If the CPU touches an address mapped to the GPU doesn't it fault a page into the CPU address space? I mean the program doesn't do anything special, but a page gets faulted in I believe.
Nvidia's unified memory architecture may make it better as same memory can be shared between CPU and GPU.