hUMA

We know the problem of rosetta with "gpu approach" (memory workload from cpu to gpu and viceversa). Can unified memory resolve it?

---

hUMA

We know the problem of rosetta with "gpu approach" (memory workload from cpu to gpu and viceversa). Can unified memory resolve it?

Based on the little knowledge that I have, I can't comment on whether it makes it viable to use OpenCL or similar for Rosetta (or at least some modules of Rosetta), although I expect it would reduce the barriers. I think there will be beneficial side-effects of GPGPU on Rosetta which could potentially have a large impact though - Crystalwell will reportedly have a 64MB L4 cache, mainly to aid graphics but also accessible to the CPU which might be beneficial to Rosetta, as that'll be much quicker than having to go to main memory. The PS4 will apparently use GDDR5 which I believe will be quicker than any DDR3 - I presume that would help Rosetta (if it could get loaded onto a PS4 - wonder how long custom bootloaders will take if it isn't allowed natively!?!)

I presume that because HSA is being pushed from all angles (x86 and Arm) that the compilers will get better and the capabilities will increase which might bring the threshold down enough for Rosetta to run on GPUs without having to invest in re-writing and maintaining multiple code-bases...

---

I presume that because HSA is being pushed from all angles (x86 and Arm) that the compilers will get better and the capabilities will increase which might bring the threshold down enough for Rosetta to run on GPUs without having to invest in re-writing and maintaining multiple code-bases...

I think it's like "a circle":
You have new hardware - write new software
Have new software - create new hardware and so on
I hope that HSA Foundation, OpenCL 1.2 and the hUMA tecnology are the first steps of a long journey

---