Experimental RADV Code Allows Vulkan Ray-Tracing On Older AMD GPUs

[skeevy420]

Very cool!
Is there a way to use a secondary GPU (not driving a display) to handle the RT rendering, or would the PCIe overhead of that negate the performance benefits of a 2nd GPU?
I imagine for dual-GPU cards, this could be beneficial.

I was wondering the same thing but thought that my APU calculations would be bottlenecked by ram speeds and latency when going from GPU to APU and back again.

[ms178]

Is there a chance we get to see other unexposed features of Vega with a little more love from community developers, such as NGG?

[remenic]

Nice writeup!

It has a typo or two, though: "Yesteryesterday and yesterday I".

It's a funny way to say "the day before yesterday and yesterday".

[arQon]

I vaguely remember a demo from Crytek a couple of years ago that managed a pretty impressive level of RTRT on a 5700.
* digs through history...

https://www.youtube.com/watch?v=1nqhkDm2_Tw

Obviously that's very much a custom system that probably doesn't map too well to the APIs nvidia defined for their HW-assisted aspects, but it's still a pretty impressive demo of how far you can get just by offloading calcs to the GPU, even without dedicated RT processing cores built into the card itself.

[FireBurn]

I was wondering the same thing but thought that my APU calculations would be bottlenecked by ram speeds and latency when going from GPU to APU and back again.

I thought APU's and GPU's could now access both system memory and video memory directly

[smitty3268]

I'm assuming this will end up being way too slow to actually be useful in any games, right? I guess it could be interesting for debug purposes? Or just some compat feature for broken games that refuse to start without RT being advertised even if they never use it.

Edit - ahh, I read the blog post and it makes sense now.

[skeevy420]

I thought APU's and GPU's could now access both system memory and video memory directly

I haven't looked into that.

Most I've ever tried in that department has been GPU offloading so I could play games on my dGPU (RX 580) while using the iGPU (Vega 7/4650g) to power the desktop display and I didn't have very good results because it kept wanting to always use the Vega. I assume it is because the Vega is newer and that's as smart as whatever DRI_PRIME does whenever you have two AMDGPUs in your system.

[Spacefish]

Nice work.
Down the road this might later on enable us to run games / applications that only do light usage of raytracing APIs (for sound reflection calculation for example) on non-raytracing Hardware!

Might be funny to have a "Raytracing" game running on Linux via DXVK (granted DXVK implements the DirectX Raytracing APIs) which don´t work on the same "non-raytracing" GPU on windows

[schmidtbag]

I was wondering the same thing but thought that my APU calculations would be bottlenecked by ram speeds and latency when going from GPU to APU and back again.

Unless I'm mistaken, I don't think RT is all that RAM intensive. Where APUs struggle most is anything that demands a lot of memory bandwidth. However, there are only x8 lanes, so that would substantially slow down data transfers.

[torsionbar28]

Very cool!

Reason number 317 for why AMD with full FOSS driver stack will *always* be superior to NVDA and their monolothic binary blob. Linus and his middle finger were right.