Apple Silicon and Ray Tracing Performance.

[APCX]

I’m curious if anyone is aware of any ray tracing performance benchmarks or scores for the A17 Pro. Hopefully if the M3 arrives soon with ray tracing then it might give us an insight into expected performance.

So far all I”ve seen is 3DMark’s Solar Bay benchmark. The scores are a little strange though:

We can see the new Snapdragon 8 gen 3 gets really good performance, but I’m a little disappointed in the A17 Pro. It seems to get only 50% more than the A16. I thought I remembered Apple saying up to a 4x improvement. I don’t expect 4x on every test or application but this seems low.

Is Solar Bay actually using ray tracing on the iPhone? Are any special measures needed to use the RT ability on the A17 Pro? I haven’t seen any apis for this.

Chart taken from here:

Qualcomm Snapdragon 8 Gen 3: First benchmarks and analysis

As part of its Snapdragon Summit in Hawaii, Qualcomm has showcased, amongst other things, its forthcoming top-tier SoC for 2024: the Snapdragon 8 Gen 3. Once again, the CPU cores are getting reorganized, ensuring more performance and we also have the first benchmark results which we would like...

www.notebookcheck.net

 

[Pressure]

The first phones with that SoC doesn't get near that score though 😅

The Xiaomi 14 gets 7688 in 3DMark Solar Bay (which is still mighty impressive) and faster than the A17 Pro SoC that reaches 6731.

 

[APCX]

The first phones with that SoC doesn't get near that score though 😅

The Xiaomi 14 gets 7688 in 3DMark Solar Bay (which is still mighty impressive) and faster than the A17 Pro SoC that reaches 6731.

For sure. I’m mainly wondering why the A17 doesn’t improve much on the A16. There seems to be very little discussion of the new RT ability.

 

[Pressure]

For sure. I’m mainly wondering why the A17 doesn’t improve much on the A16. There seems to be very little discussion of the new RT ability.

Well, the test doesn't run on the same APIs so I am unsure how we can compare them. On Windows and Android it uses Vulkan 1.1 🤷🏼‍♂️

 

[APCX]

I’m comparing the A16 and A17.

 

[diamond.g]

I’m comparing the A16 and A17.

Solar Bay is a very light RT test compared to Port Royal or Speedway. I think it is only RT reflections.

 

[APCX]

Solar Bay is a very light RT test compared to Port Royal or Speedway. I think it is only RT reflections.

Ahh very interesting. Many thanks.

 

[diamond.g]

Ahh very interesting. Many thanks.

UL is pretty good about explaining how the test are set up and how the scoring works.

Solar Bay Engine

3DMark Solar Bay uses a cross-platform graphics engine, based on the engine used in 3DMark Speed Way. It has been optimized for lightweight PCs and high-end Android devices. The engine was developed in-house with input from members of the UL Benc...

support.benchmarks.ul.com

 

[altaic]

I’m curious if anyone is aware of any ray tracing performance benchmarks or scores for the A17 Pro. Hopefully if the M3 arrives soon with ray tracing then it might give us an insight into expected performance.

So far all I”ve seen is 3DMark’s Solar Bay benchmark. The scores are a little strange though:
View attachment 2302524

We can see the new Snapdragon 8 gen 3 gets really good performance, but I’m a little disappointed in the A17 Pro. It seems to get only 50% more than the A16. I thought I remembered Apple saying up to a 4x improvement. I don’t expect 4x on every test or application but this seems low.

Is Solar Bay actually using ray tracing on the iPhone? Are any special measures needed to use the RT ability on the A17 Pro? I haven’t seen any apis for this.

Chart taken from here:

Qualcomm Snapdragon 8 Gen 3: First benchmarks and analysis

As part of its Snapdragon Summit in Hawaii, Qualcomm has showcased, amongst other things, its forthcoming top-tier SoC for 2024: the Snapdragon 8 Gen 3. Once again, the CPU cores are getting reorganized, ensuring more performance and we also have the first benchmark results which we would like...

www.notebookcheck.net

It's even more perplexing than that chart illustrates-- the A17 Pro has 5 GPU cores vs the A16's 4 GPU cores. If it had 6 cores, that's 6/4 = +50% right there; 5 cores indicates that the score improvement per core is only really ~25%.

To me, that indicates that something isn't quite right about the Solar Bay benchmark, e.g. it's doing a lot of other stuff than ray tracing or the pipeline isn't properly optimized.

Looking into it a bit more, the Solar Bay docs state, "The Solar Bay graphics test features ray-traced reflections, volumetric lighting, particles, and post-processing effects. The scene consists of three sections, with the ray tracing workload increasing in each section. The second and third sections target an average of 2x and 3x the ray tracing workload of the first section."

When I ran Solar Bay on my 15 Pro Max (for an overall score 6532), the FPS for each test section in order were 30.1, 26.9, and 19.8. So, for 3x the ray tracing workload, I'm seeing 2/3 perf as opposed to 1/3 perf. Alternatively, one might expect 19.8 FPS at 3x to equate to 19.8*3 = 59.4 FPS at 1x rather than 30.1 FPS, i.e. I'm only seeing ~50% of the expected FPS at lower ray tracing workloads.

I think part of my disappointment stems from UL declaring Solar Bay a "ray tracing benchmark for lightweight and mobile devices" which sounds like a synthetic benchmark specifically for measuring theoretical ray tracing performance. Conversely, I would never assume 4x ray tracing perf in a real-world game test... which is what Solar Bay appears to be.

 

[leman]

To me, that indicates that something isn't quite right about the Solar Bay benchmark, e.g. it's doing a lot of other stuff than ray tracing or the pipeline isn't properly optimized.

It's not that it's badly optimised, it is simply very much dominated by traditional triangle rasterisation. RT is only used for few effects, and as you wrote yourself, it's developed for mobile devices and low-end PCs. Ideally we'd need a more RT-heavy test that for example path traces the entire scene.

 

[altaic]

It's not that it's badly optimised, it is simply very much dominated by traditional triangle rasterisation. RT is only used for few effects, and as you wrote yourself, it's developed for mobile devices and low-end PCs. Ideally we'd need a more RT-heavy test that for example path traces the entire scene.

I submitted feedback to UL that profiling stats would be wonderful. Every benchmark team already has to make that effort, so kindly expose it. Obscured benchmarks are terrible; maybe my message will encounter an engineer rather than a bean counter or PR.

 

[diamond.g]

It's not that it's badly optimised, it is simply very much dominated by traditional triangle rasterisation. RT is only used for few effects, and as you wrote yourself, it's developed for mobile devices and low-end PCs. Ideally we'd need a more RT-heavy test that for example path traces the entire scene.

Yeah UL does have tests that are heavier, but no true path tracing one so far. To be honest I think only CP2077 has a benchmark that you can run where path tracing can be enabled. Even then I am not sure if there isn't some hidden raster effects for CP2077 that are harder to point out even with PT turned on.

This feature test supposedly is designed around the ray intersection rate/hardware.

Overview of the 3DMark DirectX Raytracing feature test

3DMark feature tests are special tests designed to highlight specific techniques, functions or capabilities. The DirectX Raytracing feature test measures pure ray-tracing performance. Use this test to compare the performance of the dedicated ra...

support.benchmarks.ul.com

This test (using the "full" engine of solar bay) adds in a bunch of other features. I am not sure why they just didn't use this one for Solar Bay instead of what they went with.

Overview of 3DMark Speed Way benchmark

3DMark Speed Way 3DMark Speed Way is a DirectX 12 Ultimate benchmark for ray tracing capable gaming PCs running Windows 10 and 11. Speed Way’s engine is assembled to demonstrate the new features DirectX 12 Ultimate brings to raytraced gaming. ...

support.benchmarks.ul.com

GravityMark is another benchmark on iOS that may have the RT side enabled (I mean it does on macOS I forget if it is exposed on iOS).

https://forums.macrumors.com/threads/gravitymark-gpu-benchmark.2301185

 

[leman]

GravityMark is another benchmark on iOS that may have the RT side enabled (I mean it does on macOS I forget if it is exposed on iOS).

https://forums.macrumors.com/threads/gravitymark-gpu-benchmark.2301185

Sadly, they don't seem to support Mesh Shading or RT on iOS.

@frustum, now that A17 Pro has technologies dedicated to these features, do you think they could be enabled on iOS?

 

[diamond.g]

So I guess we have to wait for the hardware next week to get some RT performance results. Wonder if the games shown will be updated with RT support by the time the hardware comes.

 

[APCX]

If it’s only 2x the M2, then it’s not getting close to Nvidia in Blender. The A17 was quoted as up to 4x improvement with RT. Weird.

 

[leman]

If it’s only 2x the M2, then it’s not getting close to Nvidia in Blender. The A17 was quoted as up to 4x improvement with RT. Weird.

2x of M2 means OPTIX 4070 laptop levels for M3 Max. Which is a good improvement, especially with those RAM sizes. Looking forward to benchmarks.

 

[APCX]

2x of M2 means OPTIX 4070 laptop levels for M3 Max. Which is a good improvement, especially with those RAM sizes. Looking forward to benchmarks.

Sure, but the 4070 is available in much cheaper machines. We’ll see I suppose.

 

[jujoje]

Got to love the apple graphs on the product page:

M3 Max is 49x faster in Redshift (than a 13-inch MacBook Pro with Intel Core i7)

Pretty curious to see how things pan out in Blender / Redshift though. Wonder if the new architecture (other than raytracing) will help much.

Landing around a 4070 is pretty decent improvement, particularly going from CUDA levels of performance to OPTIX ones, and it has the added advantage of a lot more memory and not being throttled when away from a power source. For when you have to render frames from Starbucks or sth

 

[treehuggerpro]

Got to love the apple graphs on the product page:

M3 Max is 49x faster in Redshift (than a 13-inch MacBook Pro with Intel Core i7)

Yeah, head scratcher. Then I discovered more rational comparisons behind the 16" model tag.

 

[frustum]

Sadly, they don't seem to support Mesh Shading or RT on iOS.

@frustum, now that A17 Pro has technologies dedicated to these features, do you think they could be enabled on iOS?

The new GravityMark 1.85 for iOS is in review. It includes Ray Tracing and Mesh Shading support.

update: App Store and Google Play versions with Ray Tracing support are live.

 

[Chuckeee]

General ray tracing question:

Is RT a specific known algorithm (like a video compression codex)?

 

[leman]

General ray tracing question:

Is RT a specific known algorithm (like a video compression codex)?

It’s more of a tool I suppose? RT allows to quickly locate objects in a 3D scene that are intersected by a given ray. You can use this tool to accomplish a large number of tasks, from emulating light behavior to visibility detection and motion planning.

 

[frustum]

M3 Pro RT test:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

And, unfortunately, it's behind AMD Radeon 6500 XT:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

iPhone 15 Pro Max cannot run RT benchmark in 2K/200K mode (not enough memory):

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

 

[Pressure]

M3 Pro RT test:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

And, unfortunately, it's behind AMD Radeon 6500 XT:

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

iPhone 15 Pro Max cannot run RT benchmark in 2K/200K mode (not enough memory):

GravityMark Report

GravityMark GPU Benchmark

gravitymark.tellusim.com

I'm not sure there is much point in comparing different APIs like that.

The RX 6600 only gets 3,358 under Metal RT on macOS.

The M3 Pro is also 16% faster than the M2 Max.

 

[diamond.g]

I'm not sure there is much point in comparing different APIs like that.

The RX 6600 only gets 3,358 under Metal RT on macOS.

The M3 Pro is also 16% faster than the M2 Max.

Do we have positive confirmation that MetalRT actually uses Navi2 RT hardware?