AMD's next-gen RDNA 4 Radeon graphics will feature 'brand-new' ray-tracing hardware

[erek]

“What this could mean for a potential Radeon RX 8800 XT is that even though rasterized or standard rendering performance is not up to the level of the flagship Radeon RX 7900 XTX - ray-tracing performance could blow it out of the water. And that would be impressive.

It would still need to compete with NVIDIA's upcoming GeForce RTX 50 Series and Intel's next-gen Arc, codenamed Battlemage, both of which are expected to boost ray-tracing performance. Either way, it's great to see AMD getting serious about RT.”

Read more: https://www.tweaktown.com/news/9794...ure-brand-new-ray-tracing-hardware/index.html

 

[LukeTbk]

Could be incredibly wrong, but does it feel a bit PS5 driven, rumoured performance target (and we can imagine reasonable power with it), RT, etc... which sound potentially really nice for them, if all that gen desktop GPU R&D spending come and amortized from the console side.

 

[NattyKathy]

If greatly improves RT perf for RDNA4 pans out, I might be in for a 8700XT / 8800XT. We'll see what I can afford come autumn... I'm definitely going to be picking up a Battlemage card and that'll take precidence. Not interested at all in RTX 5000 as a GPU Enjoyer (tm)

Could be incredibly wrong, but does it feel a bit PS5 driven, rumoured performance target (and we can imagine reasonable power with it), RT, etc... which sound potentially really nice for them, if all that gen desktop GPU R&D spending come and amortized from the console side.

I think you could be right. The PS5 Pro iGPU has been described as "RDNA3 with RDNA4 RT hardware" and Sony is definitely pushing for increased RT effects as a "premium" PS5 Pro feature.

 

[socK]

I mean, it's gotta absolutely clap cheeks to catch up in any meaningful way.

Nvidia has been practically doubling their RT performance each generation. I'm assuming the 5 series will yet again be a giant gain.

 

[NattyKathy]

Nvidia has been practically doubling their RT performance each generation. I'm assuming the 5 series will yet again be a giant gain.

A nitpick, but most of that perf gain has been because the GPUs are just faster in general and that's skewed by disproportionate gainz at the top-end. Ofc a 4090 will be much faster in RT than a 3090, but then it's also much faster in raster and every other metric too.

Looking at the "performance ratio" (how much performance is lost by enabling RT with otherwise-identical settings), normalized gen-on-gen RT perf gainz on NV have been more like 10% or less which isn't nothing but it's not double. They keep making the RT cores better, but they're stingy with the RT core count. So 4070Ti with similar raster perf to 3090Ti doesn't pull ahead in RT despite having more advanced hardware RT acceleration bc the 4070Ti has much fewer RT cores than the 3090Ti. Same for 3070Ti vs 2080Ti etc.

Spoiler: TPU RT benchies- look at the %, smaller # is better

I mean, it's gotta absolutely clap cheeks to catch up in any meaningful way.

Nitpick aside, you're absolutely right. RDNA4 would need like a 30% increase in RT performance to be in a good spot and more like 40-50% increase to truly impress- and that's just counting normalized RT vs raster perf, not overall performance gainz. AMD made a big deal about RDNA3 being faster in RT than RDNA2 but it turns out most of that is because Navi31 is just faster in general than Navi21, and its RT performance ratio is nearly the same as RDNA2 (L)

Sidenote, it's really impressive how well Intel did with RT perf on Alchemist. I'll spare the thread more TPU graphs, but looking at perf ratio and tier-for-tier perf, Alchemist is very close to NV. AMD has to lock in here and at least get in the same ballpark as NV and Intel or they're gonna be in a world of hurt as RT continues to gain game adoption and gamer acceptance.

 

[Lakados]

A nitpick, but most of that perf gain has been because the GPUs are just faster in general and that's skewed by disproportionate gainz at the top-end. Ofc a 4090 will be much faster in RT than a 3090, but then it's also much faster in raster and every other metric too.

Looking at the "performance ratio" (how much performance is lost by enabling RT with otherwise-identical settings), normalized gen-on-gen RT perf gainz on NV have been more like 10% or less which isn't nothing but it's not double. They keep making the RT cores better, but they're stingy with the RT core count. So 4070Ti with similar raster perf to 3090Ti doesn't pull ahead in RT despite having more advanced hardware RT acceleration bc the 4070Ti has much fewer RT cores than the 3090Ti. Same for 3070Ti vs 2080Ti etc.

Spoiler: TPU RT benchies- look at the %, smaller # is better

View attachment 651695
View attachment 651696

Nitpick aside, you're absolutely right. RDNA4 would need like a 30% increase in RT performance to be in a good spot and more like 40-50% increase to truly impress- and that's just counting normalized RT vs raster perf, not overall performance gainz. AMD made a big deal about RDNA3 being faster in RT than RDNA2 but it turns out most of that is because Navi31 is just faster in general than Navi21, and its RT performance ratio is nearly the same as RDNA2 (L)

Sidenote, it's really impressive how well Intel did with RT perf on Alchemist. I'll spare the thread more TPU graphs, but looking at perf ratio and tier-for-tier perf, Alchemist is very close to NV. AMD has to lock in here and at least get in the same ballpark as NV and Intel or they're gonna be in a world of hurt as RT continues to gain game adoption and gamer acceptance.

Intel and Nvidia use “AI” acceleration in their ray tracing and upscaling, Nvidia uses CUDA, Intel uses OneAPI, but either way they have dedicated cores to accelerating it. AMD doesn’t, really hoping their talk of improvements in AI on their upcoming cards address this.

 

[NattyKathy]

Intel and Nvidia use “AI” acceleration in their ray tracing and upscaling, Nvidia uses CUDA, Intel uses OneAPI, but either way they have dedicated cores to accelerating it. AMD doesn’t, really hoping their talk of improvements in AI on their upcoming cards address this.

Yeah the ability to accelerate low-precision math with dedicated units is huge for stuff like denoising and upscaling. The performance impact of denoising is often underappreciated and NV / Intel are way ahead there.

There's also the matter of the capabilities of the dedicated raytracing accelerators separate from the "AI" stuff and AMD is also way behind in that aspect.

A few years back Imagination Technologies (who have been doing stuff with hardware accelerated RT since before NV even) put out a whitepaper suggesting a tierlist for RT acceleration levels based on how much of the BVH & intersection tasks can be offloaded onto dedicated hardware units- Link to blag post

If I'm remembering / understanding correctly the hardware capabilities, Turing/Ampere fall into "Level 3" acceleration, Ada/Alchemist are "Level 4", and RDNA2/RDNA3 are down in "Level 2"- doing the absolute minimum to be considered "hardware accelerated" RT. The "Ray Accelerators" in RDNA2/3 just aren't very capable in terms of how much of the RT pipeline they can take on, which means more is being pushed onto the shader cores (and AMD is also somehow using the ROPs for parts of the intersection calcs? Seems desperate...)

Add in the lack of low-precision high-throughput "AI" acceleration for denoising and upscaling and it's not a pretty picture for AMD. Technically RDNA3 does have "Matrix Accelerators" which on paper do much the same thing as NVs Tensor Cores or Intel's XMX cores... But I haven't really found any info suggesting they're actually being used for anything.

More robust low-precision matrix accelerators (that actually get utilized!!!) and significantly more advanced BVH & intersection hardware is a must for RDNA4. AMD kinda phoned it in for hardware RT acceleration on RDNA2/3 and that cannot continue.

 

[LukeTbk]

Looking at the "performance ratio" (how much performance is lost by enabling RT with otherwise-identical settings), normalized gen-on-gen RT perf gainz on NV have been more like 10% or less which isn't nothing but it's not double

the faster you go, the bigger the performance ratio you loose for the same amount of millisecond used by RT, we rarely see some good RT values (and I am not sure if would matter much when it goes to buy a gpu for gaming)

A 4090/4080 are probably significantly more powerful in RT than a 3090/3080 than it is in raster, the 4090 lead over a 3090 tend to get bigger heavier/purer RT the workload get (like you say a game that has RT in it benchmark would be greatly influenced by how fast the gpu is at everything else, not just the RT and its denoising)

But in game where most of the ms on a frame is not due to the RT being on, doubling your RT performance will not move the framerate that much.

Rt core
2080ti: 68
3090ti: 84
4090..: 128

RT Tflops:
2080TI: 42.9
3090TI: 78.1
4090..: 191.0

Tflops per core:
2080TI: 0.63 Tflops
3090TI: 0.93 Tflops + 47%
4090..: 1.49 Tflops + 60%

For AMD to gain 30% more frame in a RTX games, specially if it was not already doing terrible, does it need to get 60% faster at doing raytracing and its denoising ? That type of unintuitive and hard to do math is going on.

If a 4090 is twice as fast with RT on than a 3090, even if it lost 40% of its frame rate to put it on like a 3090, it needed to be twice at RT as fast to do so, because its start non-rt frame rate had less MS.

 

[Sycraft]

I mean, it's gotta absolutely clap cheeks to catch up in any meaningful way.

Nvidia has been practically doubling their RT performance each generation. I'm assuming the 5 series will yet again be a giant gain.

Doesn't have to be as powerful as nVidia to still be a nice step forward. The thing is, the better the RT AMD has, the more games that'll implement RT. Right now RT is pretty much a PC-only, nVidia-only feature. Ya you CAN use it on other stuff, but performance is often not good enough. Makes less devs interested in it, or it is implemented in a very weak way so that there is technically RT but it has a minimal hit.

If AMD desktop chips, and Intel desktop chips, and the PS 5.5 or whatever the fuck its called has better support for it, then we could see it getting more widely used. While it might not look as good as what nVidia does, that's fine. Could easily have options for resolution, number of bounces, and so on that you can turn up on GeForces and down on other stuff, but still get nice RT.

 

[socK]

A nitpick, but most of that perf gain has been because the GPUs are just faster in general and that's skewed by disproportionate gainz at the top-end. Ofc a 4090 will be much faster in RT than a 3090, but then it's also much faster in raster and every other metric too.

Looking at the "performance ratio" (how much performance is lost by enabling RT with otherwise-identical settings), normalized gen-on-gen RT perf gainz on NV have been more like 10% or less which isn't nothing but it's not double. They keep making the RT cores better, but they're stingy with the RT core count. So 4070Ti with similar raster perf to 3090Ti doesn't pull ahead in RT despite having more advanced hardware RT acceleration bc the 4070Ti has much fewer RT cores than the 3090Ti. Same for 3070Ti vs 2080Ti etc.

Spoiler: TPU RT benchies- look at the %, smaller # is better

View attachment 651695
View attachment 651696

Nitpick aside, you're absolutely right. RDNA4 would need like a 30% increase in RT performance to be in a good spot and more like 40-50% increase to truly impress- and that's just counting normalized RT vs raster perf, not overall performance gainz. AMD made a big deal about RDNA3 being faster in RT than RDNA2 but it turns out most of that is because Navi31 is just faster in general than Navi21, and its RT performance ratio is nearly the same as RDNA2 (L)

Sidenote, it's really impressive how well Intel did with RT perf on Alchemist. I'll spare the thread more TPU graphs, but looking at perf ratio and tier-for-tier perf, Alchemist is very close to NV. AMD has to lock in here and at least get in the same ballpark as NV and Intel or they're gonna be in a world of hurt as RT continues to gain game adoption and gamer acceptance.

Look at the Path Tracing numbers and it's even more egregious, where the ray tracing workload dominates and is about as heavy as you can possibly get.

https://www.techpowerup.com/review/...ty-benchmark-test-performance-analysis/6.html

You can really see the raw RT gen over gen uplift NVIDIA was churning out when it's extremely heavy. Relatively, it's gigantic.

 

[Axman]

Could be incredibly wrong, but does it feel a bit PS5 driven, rumoured performance target (and we can imagine reasonable power with it), RT, etc... which sound potentially really nice for them, if all that gen desktop GPU R&D spending come and amortized from the console side.

I've heard from a couple of sources that say Sony really threw their ass into gear on this, to the point of possibly doing most of the development.

And it's not like Sony's a second-tier developer when it comes to hardware, they got guys.

 

[Flogger23m]

Looking at the "performance ratio" (how much performance is lost by enabling RT with otherwise-identical settings), normalized gen-on-gen RT perf gainz on NV have been more like 10% or less which isn't nothing but it's not double. They keep making the RT cores better, but they're stingy with the RT core count. So 4070Ti with similar raster perf to 3090Ti doesn't pull ahead in RT despite having more advanced hardware RT acceleration bc the 4070Ti has much fewer RT cores than the 3090Ti. Same for 3070Ti vs 2080Ti etc.

And that is why ray tracing will still be a long ways off until it is very common in games. Even a 20% performance increase in ray tracing isn't that noteworthy. If you go from 50 frame rates to 60 frame rates that is a decent improvement, but still what most would consider to be too low for an average frame rate. You're still going to need DLSS, or to turn down other settings.