Nvidia Ampere Discussion [2020-05-14]

daw213cktg.png


https://brianlovin.com/hn/24628189
 
Curious if the benefits from using a feature like SAM would be included if talking about IPC gains, or is considered a separate metric?
 
Nvidia claims Ampere already supports “Smart Access Memory” and says they’re working on implementing it in drivers.


That's great news, hopefully this becomes more widespread and we see some really gains from it. Once it becomes the standard rather than the exception game developers can start targeting it and in some ways the PC with its separate components can take advantage of something akin to a UMA.

Does anyone know if GPU's can read directly from system memory in the same way? I *think* that's what HBCC allowed but you had to manually set a potion of the system memory to be used as VRAM.
 
That doesn’t explain why AMD didn’t enable it earlier. They certainly needed the performance.
To me, it looks like they've already enabled it on driver level for DX12 games on current GPUs.
It would be weird not to do it since host addressable video memory has been available in Vulkan for years and both NVIDIA and AMD supported it for quite a while (256 MBs though, but this might be enought to store hot descriptors/constants/other stuff which changes on frame to frame basis and is hard to cache on GPU, especially with some devs' love).
I see RX 5700 XT being able to compete with RTX 2080 in some recent titles like Valhalla in 1440p and below, while still being much slower in 4K and it's not the only example.
It would be nice if somebody tested this title with the most recent driver vs year old driver so that we can verify whether there is something in recent drivers, which can boost performance a lot in titles like Valhalla.
 
That doesn’t explain why AMD didn’t enable it earlier. They certainly needed the performance.

Maybe not worth the driver work needed ? I mean, some games are not 100% stable with HBCC on, while perfectly fine withtout it (ffxv come in mind for me). So I guess they didn't want to check if this was working ok with older gpus ?
 
To me, it looks like they've already enabled it on driver level for DX12 games on current GPUs.
It would be weird not to do it since host addressable video memory has been available in Vulkan for years and both NVIDIA and AMD supported it for quite a while (256 MBs though, but this might be enought to store hot descriptors/constants/other stuff which changes on frame to frame basis and is hard to cache on GPU, especially with some devs' love).
I see RX 5700 XT being able to compete with RTX 2080 in some recent titles like Valhalla in 1440p and below, while still being much slower in 4K and it's not the only example.
It would be nice if somebody tested this title with the most recent driver vs year old driver so that we can verify whether there is something in recent drivers, which can boost performance a lot in titles like Valhalla.
That 256MB is always supported AFAIK no matter what API or GPU, the difference is that now that 256MB can be turned into however many gigabytes there is VRAM on each particular card
 
sebbbi talks resizable RBAR support and why it's good.

I wonder if Nvidia has planned this as part of their DirectStorage support and are now pulling this particular feature forward so that they can get some early gains in the benchmark wars.


Sebbi's been on a tear lately with the advice and explanations. Love it and love seeing RBAR as an open standard as compared to smart access memory. Really need one standard across all platforms, hopefully looking as much like consoles as possible as well, to reduce the amount of individual platform support that's needed.
 
You would think at least more tech savvy people on this forum knew that SAM is just how AMD calls it's RBAR implementation on Windows platform. I mean there was literally a video posted right after announcement where Frank Azor explains in detail how it works.

If Nvidia can confuse people with RTX it’s only fair that AMD can do the same with SAM.
 
Back
Top