Ray tracing on Apple Silicon?

[leman]

They do, and so far no one has seemed to take them up on the offer (the one game they touted during last years WWDC [Metro Exodus] didn't bother to even add those effects in the basic version of the engine).

Not really surprising. The performance is simply not there for RT to be practically viable. I mean, you could probably sprinkle some soft shadows here and there and it will work well enough on M1-class hardware, but why would you bother?

Because Nvidia says they are a separate block/core? ??‍♂️

Page 12 has figure 3 which shows the separate texture units (4 per sm) and the RT cores below it.

As I said, reality is often more complex than what companies say in their marketing material. Nvidia also claims that their cores are "scalar", but in reality they use large SIMD ALUs. If you are interested in the topic, I suggest you have a look at the paper I linked. It's pretty cool, and they make fairly convincing guesses to what the Nvidia "RT cores" actually do. Hint: it's probably not what you might think.

Granted what Nvidia seems to provide isn't 100% the same as the ISA papers AMD provides for RDNA, but from what I can tell AMD ISA paper doesn't really talk about how their Ray Accelerators work, from a code perspective, either.

If I understood AMD documentation correctly, they simply give you bounding-box intersection instruction, probably implemented with specialised fixed-function hardware. That is the "easy" thing to do. If this is the extent of AMD's implementation, there is no surprise that Nvidia's RT is faster. The "secret source" is work reordering and memory access batching.

 

[Bodhitree]

The crazy thing to me is no one else is using IMGTec stuff but Apple. For it to be as good as it is, with no one else using it is weird.

In smartphones and tablets that seems to be the case, not sure whether it holds true for cars and other embedded devices. The marketing on Imagination’s website is aimed at Mobile, Data Centre and Automotive applications.

It really depends what market segments put a premium on low-power, high-performance graphics that are better than ARM’s default Mali implementation and are people making their own SoCs. If you’re using a whole chip from Qualcomm (as most Android smartphone manufacturers do) then you’re almost certainly also going to be stuck with their Adreno graphics core.

I can imagine some data centres using Imagination’s technology, since they are UK-based they would make a good fit for the European supercomputer initiative for which SiPearl is designing the CPU, and given the power-efficiency benefits they might not stack up badly against mainstream graphics cards in performance-per-watt.

 

[diamond.g]

Not really surprising. The performance is simply not there for RT to be practically viable. I mean, you could probably sprinkle some soft shadows here and there and it will work well enough on M1-class hardware, but why would you bother?



As I said, reality is often more complex than what companies say in their marketing material. Nvidia also claims that their cores are "scalar", but in reality they use large SIMD ALUs. If you are interested in the topic, I suggest you have a look at the paper I linked. It's pretty cool, and they make fairly convincing guesses to what the Nvidia "RT cores" actually do. Hint: it's probably not what you might think.



If I understood AMD documentation correctly, they simply give you bounding-box intersection instruction, probably implemented with specialised fixed-function hardware. That is the "easy" thing to do. If this is the extent of AMD's implementation, there is no surprise that Nvidia's RT is faster. The "secret source" is work reordering and memory access batching.

From my understanding, AMD RT isn't nearly as bad as shown in games, per se. It is the denoising that crushes AMD hardware (from the RT thread on Beyond3d IIRC).

EDIT: now I cannot find the B3D thread, lol.

 

[crazy dave]

The crazy thing to me is no one else is using IMGTec stuff but Apple. For it to be as good as it is, with no one else using it is weird.

For the ImgTech ray tracing IP it could be just that it’s new. For TBDR, TBDR is hard to get right (why ImgTech was so sure Apple couldn’t just leave them) and most of the world’s graphics is built with IMR in mind and it doesn’t help with compute. So unless you say … control the full product stack and can push your developers towards a certain direction, it won’t give you the desired benefits without additional software optimizations that you wouldn’t otherwise be doing. That’s my understanding anyway.

 

[diamond.g]

In smartphones and tablets that seems to be the case, not sure whether it holds true for cars and other embedded devices. The marketing on Imagination’s website is aimed at Mobile, Data Centre and Automotive applications.

It really depends what market segments put a premium on low-power, high-performance graphics that are better than ARM’s default Mali implementation and are people making their own SoCs. If you’re using a whole chip from Qualcomm (as most Android smartphone manufacturers do) then you’re almost certainly also going to be stuck with their Adreno graphics core.

I can imagine some data centres using Imagination’s technology, since they are UK-based they would make a good fit for the European supercomputer initiative for which SiPearl is designing the CPU, and given the power-efficiency benefits they might not stack up badly against mainstream graphics cards in performance-per-watt.

Well no auto manufacturers have admitted to using it so far.

 

[diamond.g]

For the ImgTech ray tracing IP it could be just that it’s new. For TBDR, TBDR is hard to get right (why ImgTech was so sure Apple couldn’t just leave them) and most of the world’s graphics is built with IMR in mind and it doesn’t help with compute. So unless you say … control the full product stack and can push your developers towards a certain direction, it won’t give you the desired benefits without additional software optimizations that you wouldn’t otherwise be doing. That’s my understanding anyway.

I know of at least 1 company (outside of Apple) that likes to control it's own destiny and they chose to partner with AMD instead of finding something with this in it (and if we believe them this is supposed to be better Perf/W than the AMD alternative they went with).

 

[crazy dave]

I know of at least 1 company (outside of Apple) that likes to control it's own destiny and they chose to partner with AMD instead of finding something with this in it (and if we believe them this is supposed to be better Perf/W than the AMD alternative they went with).

Are you referring to consoles?

 

[diamond.g]

Are you referring to consoles?

Nah Tesla's.

 

[diamond.g]

Nah Tesla's.

I don't think Microsoft will go off x86 for consoles before just not making consoles at all.

 

[crazy dave]

Nah Tesla's.

That’s still a console it’s just hooked up to a car battery. That’s easy they went with a proven desktop GPU that’s already used in consoles and I believe it’s a Linux OS? Again unless software is written with TBDR in mind, I don’t think you get much benefit. So just stick an IMR GPU simplifies any porting that it needs if any (I’m assuming it’s a Vulkan API?).

I don't think Microsoft will go off x86 for consoles before just not making consoles at all.

? You could pair an ImgTech GPU with an x86 CPU.

 

[diamond.g]

That’s still a console it’s just hooked up to a car battery. That’s easy they went with a proven desktop GPU that’s already used in consoles and I believe it’s a Linux OS? Again unless software is written with TBDR in mind, I don’t think you get much benefit. So just stick an IMR GPU simplifies any porting that it needs if any (I’m assuming it’s a Vulkan API?).



? You could pair an ImgTech GPU with an x86 CPU.

For Tesla's purpose TBDR could work great with keeping responsiveness.

I would be surprised for the next gen of consoles to stop using APUs.

 

[Bodhitree]

I reckon for Xbox and Playstation unless there is some very substantial reward they are going to stick with AMD’s CPU plus GPU combinations. If Imaginations raytracing is massively better than AMD’s then perhaps but I just don’t see it happening.

For Nintendo it might be interesting, for the Switch they’ve relied upon Nvidia’s Tegra chips, but they don’t seem to be developing quickly. IMG‘s low power graphics would do very well in a Switch 2 console, but who would design the SoC? Perhaps they might make a deal with Samsung?

 

[Xiao_Xi]

unless software is written with TBDR in mind, I don’t think you get much benefit.

How different are APIs for IMR and TBDR GPUs? Does Metal work on TBDR and IMR GPUs?

if we believe them this is supposed to be better Perf/W than the AMD alternative

Are TBDR GPU more efficient than IMR GPU?

who would design the SoC?

Imagination has RISC-V SoC.

 

[leman]

Again unless software is written with TBDR in mind, I don’t think you get much benefit.

TBDR is more efficient on average, so you can get away with less memory bandwidth for the same performance for example.

How different are APIs for IMR and TBDR GPUs? Does Metal work on TBDR and IMR GPUs?

Metal works on Apple, AMD and Intel GPUs. Apple has some optional functionality that exposes the inner workings of their TBDR GPUs and will thus only work on Apple GPUs.

Are TBDR GPU more efficient than IMR GPU?

In general, yes. The way how TBDR is set up improves locality of data accesses and shader utilisation. Caveats apply of course.

 

[Xiao_Xi]

How good is Imagination's ray tracing?

Demo:

Explanation:

 

[diamond.g]

How good is Imagination's ray tracing?

Demo:

Explanation:

Neat. Assuming this is using Vulkan I wonder if they plan on releasing the demo to be ran on consumer hardware "freely".

 

[diamond.g]

I reckon for Xbox and Playstation unless there is some very substantial reward they are going to stick with AMD’s CPU plus GPU combinations. If Imaginations raytracing is massively better than AMD’s then perhaps but I just don’t see it happening.

For Nintendo it might be interesting, for the Switch they’ve relied upon Nvidia’s Tegra chips, but they don’t seem to be developing quickly. IMG‘s low power graphics would do very well in a Switch 2 console, but who would design the SoC? Perhaps they might make a deal with Samsung?

From what I understand of Nintendos dev tools, they are pretty much wholly written by Nvidia at this point, so them switching to another vendor is probably going to be a major set back for them (at least in terms of third party support).


I also wonder what the minimum number of RT cores IMGTec has for a viable RT solution (taking the 6500XT as an example of not enough, lol).

 

[Bodhitree]

Imagination has RISC-V SoC.

RISC-V is largely an unknown quantity. I don’t think Nintendo will move away from ARM, if only because they have been working with the tool chain for a long time. Whereas Samsung has foundries, a cpu core in Exynos, and experience in producing tablet and mobile SoCs.

It’s speculation, but to bring real-time raytracing to a gaming tablet screen Nintendo may make the jump, it would bring them to parity with the big consoles. It would make a lot of sense.

 

[diamond.g]

RISC-V is largely an unknown quantity. I don’t think Nintendo will move away from ARM, if only because they have been working with the tool chain for a long time. Whereas Samsung has foundries, a cpu core in Exynos, and experience in producing tablet and mobile SoCs.

It’s speculation, but to bring real-time raytracing to a gaming tablet screen Nintendo may make the jump, it would bring them to parity with the big consoles. It would make a lot of sense.

Nintendo is going to target parity with the big consoles about the same time Apple makes a home console for less than $199.

 

[Bodhitree]

I also wonder what the minimum number of RT cores IMGTec has for a viable RT solution (taking the 6500XT as an example of not enough, lol).

From what the marketing materials say the technology stretches from one core supplying 1.3 GRays to a maximum of 7.2 GRays total in a multi-core data centre or desktop setup. They said the whole GPU was 9 TFlops maximum, which doesn‘t really compare with a top-end Nvidia card which i believe are about 40 TFlops. But on a 720p Nintendo Switch 2 screen it might be just the ticket…

 

[Bodhitree]

Nintendo is going to target parity with the big consoles about the same time Apple makes a home console for less than $199.

Well, why not make a big splash when you have the opportunity to do so?

 

[diamond.g]

From what the marketing materials say the technology stretches from one core supplying 1.3 GRays to a maximum of 7.2 GRays total in a multi-core data centre or desktop setup. They said the whole GPU was 9 TFlops maximum, which doesn‘t really compare with a top-end Nvidia card which i believe are about 40 TFlops. But on a 720p Nintendo Switch 2 screen it might be just the ticket…

Interesting, so per watt it seems like they would be processing more rays than an equivalent Nvidia solution.

Well, why not make a big splash when you have the opportunity to do so?

Because no one buys Nintendo hardware for the fancy graphics (which has been true since the Nintendo 64 days) and at this point Nintendo would rather not lose money even trying (they have admitted as much before).

 

[diamond.g]

Not really surprising. The performance is simply not there for RT to be practically viable. I mean, you could probably sprinkle some soft shadows here and there and it will work well enough on M1-class hardware, but why would you bother?

That would have been low hanging fruit for Blizzard to implement with their native port of World of Warcraft.

As I said, reality is often more complex than what companies say in their marketing material. Nvidia also claims that their cores are "scalar", but in reality they use large SIMD ALUs. If you are interested in the topic, I suggest you have a look at the paper I linked. It's pretty cool, and they make fairly convincing guesses to what the Nvidia "RT cores" actually do. Hint: it's probably not what you might think.

Yeah that paper is interesting, I wonder what changed for Ampere to be so much faster than Turing at RT then.

 

[Xiao_Xi]

per watt it seems like they would be processing more rays than an equivalent Nvidia solution.

Nvidia's RTX 2060 and Innosilicon's Fantasy 1 (Imagination Technologies' PowerVR architecture) has around 5 FP32 TFLOPS. But, Nvidia's GPU needs 160W, while Innosilicon claims its GPU only needs 50W.

Chinese Fenghua GPU Aims for GeForce RTX 3060 Compute Performance

The Fenghua GPU is rather powerful, though it can't match high-end GPUs.

www.tomshardware.com

 

[diamond.g]

Nvidia's RTX 2060 and Innosilicon's Fantasy 1 (Imagination Technologies' PowerVR architecture) has around 5 FP32 TFLOPS. But, Nvidia's GPU needs 160W, while Innosilicon claims its GPU only needs 50W.

Chinese Fenghua GPU Aims for GeForce RTX 3060 Compute Performance

The Fenghua GPU is rather powerful, though it can't match high-end GPUs.

www.tomshardware.com

I look forward to seeing these cards in the wild.