Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

+6 GBs of VRAM is a pretty solid spec.

There's some degree of worry that 10GB on the 3080 is not enough for the high-end games of the future. NVidia made a mistake leaving a huge gaping hole between the 10GB 3080 and the 24GB 3090.

A 6800 XT fills that hole with a nice little 16GB.



I doubt it was a mistake. more likely an intentional 3080ti-sized hole in the lineup.


The VRAM is a plus but isn't AMDs compute performance and infrastructure pretty bad?

I want to like their cards but they just don't seem to be able to balance the seesaw across all fronts


Compute performance is fine. 20+ TFlops (or T-Integer ops) on 128MB L3 cache on 16GBs GDDR6. I don't think anyone can complain about the raw compute performance of these cards.

Infrastructure: yeah. Without ROCm support, its hard to recommend. You still get Windows DirectCompute, OpenCL, and Vulkan compute. But ROCm is much easier to use than those other frameworks.

---------

For compute: With 40 WGP x 4 SIMDs per WGP x 32 Threads per WGP x Occupany 4, you're probably aiming at 20,480 SIMD-threads if you're going to be using compute on this card.

Occupancy 20 is supported: 20-threads per VALU, kinda like hyperthreading although you split your registers between the threads. 20 is probably too much (too few VALUs per thread), but 1 is too few (you'll be waiting on VRAM latency). So 4 to 8 is an estimate for how much occupancy a typical program probably needs to hide memory latency.

Those 16GBs / 20kiloThreads == 780kB per thread. Sooooo... yeah. RAM goes quick... really quick... on compute applications. Compute programmers really want more VRAM.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: