Friday, May 3rd 2024
AMD to Redesign Ray Tracing Hardware on RDNA 4
AMD's next generation RDNA 4 graphics architecture is expected to feature a completely new ray tracing engine, Kepler L2, a reliable source with GPU leaks, claims. Currently, AMD uses a component called Ray Accelerator, which performs the most compute-intensive portion of the ray intersection and testing pipeline, while AMD's approach to ray tracing on a hardware level still relies greatly on the shader engines. The company had debuted the ray accelerator with RDNA 2, its first architecture to meet DirectX 12 Ultimate specs, and improved the component with RDNA 3, by optimizing certain aspects of its ray testing, to bring about a 50% improvement in ray intersection performance over RDNA 2.
The way Kepler L2 puts it, RDNA 4 will feature a fundamentally transformed ray tracing hardware solution from the ones on RDNA 2 and RDNA 3. This could probably delegate more of the ray tracing workflow onto fixed-function hardware, unburdening the shader engines further. AMD is expected to debut RDNA 4 with its next line of discrete Radeon RX GPUs in the second half of 2024. Given the chatter about a power-packed event by AMD at Computex, with the company expected to unveil "Zen 5" CPU microarchitecture on both server and client processors; we might expect some talk on RDNA 4, too.
Sources:
HotHardware, Kepler_L2 (Twitter)
The way Kepler L2 puts it, RDNA 4 will feature a fundamentally transformed ray tracing hardware solution from the ones on RDNA 2 and RDNA 3. This could probably delegate more of the ray tracing workflow onto fixed-function hardware, unburdening the shader engines further. AMD is expected to debut RDNA 4 with its next line of discrete Radeon RX GPUs in the second half of 2024. Given the chatter about a power-packed event by AMD at Computex, with the company expected to unveil "Zen 5" CPU microarchitecture on both server and client processors; we might expect some talk on RDNA 4, too.
227 Comments on AMD to Redesign Ray Tracing Hardware on RDNA 4
There is a difference here though. In the longer term, IIRC, RT and PT will make it easier for devs to code? Once the hardware is up to scratch, it should make it better for all of us. I mean, the movie industry has used RT in its CGI for ages. Its what makes CGI look real.
Then again, to play devils advocate, I don't need my game to look real. I want to enjoy the game. Case in point, Sony make some stellar titles with amazing immersion. My fave game in a long time is Days Gone. It's got no RT. Same for the Horizon games, I think. There are indie games that don't need it either - Super Hot - I'm looking at you. 2D scrollers are another example.
In the push for realism, the graphical hardware requirement goes up and up. More fidelity needs more componentry. RT/PT is the future, but it's not the present, and that's the worst part about PC gaming right now. Nvidia are hardcore pushing the AI narrative and steaming ahead with their hardware to suit. They are undoubtedly the leaders, but they're also leading us on a ride. They need to push AI fidelity solutions because their own hardware isn't powerful enough to run what they say is the best thing ever. They're playing clever language games with us, then have the cheek to make up PR speak about paying more to save... Cringe.
But, without any contest, it's what they can do. There is always a market leader, and the consumerbase will either roll with them, villify them, or (like me) hold my hands up and say, "It's not worth it to me. I see no reason to chase this mighty green dragon." Yet, I did buy an Nvidia card this round (though for the first time in years it was 3 tiers down). My criteria was simple - 50% faster than my old 2080ti, and similar power draw with quiet fans. RT didn't matter so much, but I'll not lie, it was about 10% of my decision.
Now, as a mod - I'll say this:
This news piece is about RDNA 4 and RT. This isn't a place for rasterisation discussions, or posting benchies of non-RT raster. If you're arguing AMD's point on that here, you're definitely in the wrong thread.
On the article, I have to wonder if it's feasible to for AMD to separate dedicated RT hardware into a chiplet. This would allow them to pack more cores onto the GPU core die itself while tailoring RT performance to the market. Imagine being able to configure the number of RT chiplets to the task, allowing them to create cheaper GPUs without RT for those that want to save money and also GPUs packed with RT chiplets for those that want the best RT performance. They could even make dedicated RT accelerator GPUs for video production use, all the while using the same basic set of chiplets it uses across it's entire stack. Might also be wise to segment the AI cores from the GPU core as well.
Of course I'm not a GPU engineer so I can't say if the latency and bandwidth would be an issue here. We know AMD couldn't do two GPU dies due to bandwidth issues, not yet at least using their current organic substrate. That said bandwidth requirements of just a fraction of that die might be a different story. I'd say it absolutely will not make it easier to code. It'll make it easier to get higher quality lighting for sure but at the end of the day you'll still want to be placing light by hands and creating custom shaders if you want to do a specific kind of visual effect.
I am told Intel is this really competitive new third player in the GPU market and they are for sure gonna destroy AMD but the fact is performance is in the dump on Intel GPUs in almost all conceivable metrics.
Also from what i'm hearing don't discount a high end card.
While ASICs optimized for RT would offer better performance under these circumstances, the gameplay remains unplayable on both low-end and dedicated RT hardware. Regardless, for those interested, RDNA4's implementation with larger blocks dedicated to RT promises significantly improved perf (and remain unplayable like the others)...
Even if RDNA4 brings RT performance/watt parity with Lovelace or even Blackwell, the goalposts will probably shift to whatever Nvidia has in a scale that AMD doesn't. With Fermi it was compute throughput, with Kepler it was geometry performance, with Turing / Ampere / Ada it was RT performance, with Blackwell it'll probably be AI.
There's no remotely possible catching up if the >80% marketshare player uses its dominant position to constantly weasel in new goalposts every new generation. Videogame developers happily adopted the 1st-gen tessellation on the X360 and TruAudio on the PS4 and XB1, yet their PC counterparts in AMD dGPUs were promptly ignored.
Besides, GPUs are traditionally tested at higher resolutions, and CPUs at lower resolutions, to avoid bottlenecks from other parts of the system.
It's true, 1080p/1440p is the range these cards can comfortably operate with settings maxed. However the A770 in particular is decent as a productivity card, and doesn't mind higher resolutions for gaming either.
The original comment thread was how does Arc compare at RT compared to the competition.
Someone tried to say the RX 7600 is a faster card, so I posted some TPU testing that shows that's not the case. RT or no.
If I wanted to play at 1080p60 with RT off I'd buy an APU.
Keep name calling though.
Thinks someone's got the needle about the Intel thread. I didn't get proven wrong there either. Gaming, Zen 4. Datacentre and servers, Zen 4. Niche cases that should use a GPU instead of the CPU, Intel or AMD.
Ah, shucks, I forgot about the "Nvidia good, AMD bad" infallible logic once again, my bad.
Even the 12 GB 4070 Ti is faster at every resolution except 4K, while using 75 W less. If my budget was £300 and I was choosing between a 7600XT and a Arc A770 I'd absolutely pick the A770.
Cheapest 7600XT is £20 more though at £320.
That's just a reality check. GPUs need to be able to handle RT in small implementations. It's not an all or nothing scenario as some might believe.
I know you are totally obsessed with RT and if it was after you literally nothing else would matter but that's not how it works in the real world. AMD can ask for a premium for better raster performance in the same way Nvidia can for RT, it must be real hard to wrap your head around the concept but it's actually a real phenomena.
The thing is Intel sucks in both raster and RT so you shouldn't be so bewildered that AMD cards are much more expensive than Intel's.
RT talk people. Raster isn't the point of this OP.
Reply bans for those unwilling to discuss RT.
We know that Intel fumbled when it comes to general rendering performance, in theory the A770 should have been faster than it is...but it failed to deliver. However, TPU review showed that the hit that Arc is taking when RT is on is lower when compared to RDNA 2 and not that far behind ampere. Now that doesn't mean that Battlemage will manage to battle it out with Blackwell and RDNA 4, but if they can figure out the general rendering performance, they might realistically target an overall parity with Ada upper midrange (and nvidia even made the job easier for them)
i've never heard anyone say that for a midrange gpu and even more so with RT on