GPU Ray Tracing Performance Comparisons [2021] *spawn*

Just to add to the dogpile, my 3080Ti finds a pretty whopping efficiency gain when poking at the undervolting options. I write a decently-sized post a few weeks ago regarding my undervoltage / overclocking findings with my card; I'll post this, go find my old post, and then edit this one to insert it here: I was happier with my laptop's GTX 1050Ti 4GB than I'll ever be with my desktop GTX 1060 3GB | Beyond3D Forum

The cliff's notes for my particular situation:

 

Hopefully AMD will also follow suit in the ML game like Intel/NV in their next approach.

 

The reason I asked about watts per RT-FPS is to find out whether "theoretical TFLOPS" is part of the difference we're seeing.

With the wide range in performance delta between RDNA 2 and Ampere (and perhaps with clues based upon Turing) I'm wondering if we can determine whether there are situations where FP throughput is a significant portion of the delta between Ampere and RDNA 2.

 

We will know more when were leaving cross gen behind perhaps.

 

RT shading performance is probably dominated by shader and memory coherence (or lack thereof) and not so much by raw flops. I don’t know how you would isolate FP throughout from all the other dependencies in an RT pass.

I ran a few traces of the Bright Memory raytracing benchmark and SM instruction throughput was pretty low during the Dispatch Rays call. What's interesting is that right after each call to Dispatch Rays there was a period of high SM utilization which makes me think RT in this benchmark is writing to a V/G buffer and the hit shading is done after the raytracing pass and not during.

It's inconclusive but the only thing I could gather is that there are periods of ~80% SM instruction throughput. So if you set aside FP32 flops for a sec and just look at overall instruction throughput Ampere may have an edge as I believe its peak instruction throughput is higher than on RDNA2. It's murky because a lot of those instructions are ALU/SFU transcendentals, type-conversions and bit-manipulation so it's not a straightforward flops comparison.

 

Those results show the same power gap just with the variances of silicon quality.

 

For those two things, RDNA and Ampere should be equivalent, ish. I suppose.

That's the standard DXR 1.0 approach, as I understand it. Favours NVidia (and soon to be, Intel). Though I assume there's a dedicated buffer set up by the driver to handle the results of despatch rays (spilling results to off-chip memory). That buffer is then consumed by the hit shaders.

Is that the correct interpretation?

Type conversion and bit manipulation should run at native SIMD rate, for at least one SIMD out of the pair in Ampere. Perhaps those only run on the "integer" SIMD?

Does 80% of SM instruction throughput imply 1.6 instructions per clock? You're implying that's the peak utilisation as far as I can tell, so the overall utilisation is lower. Also, is some of that shading unrelated to hit shaders?

AMD's ray tracing performance diagnostic tools have only just been announced, so diving deep isn't possible yet.

In the end, comparing say 6800XT and 3080, where 3080 has around 40% more FLOPS, how much of the 60%+ RT advantage (at 4K) seen in 3080 is due to FLOPS?

 

DispatchRays doesn't return until after the hit shader runs so it seems the bound hit shader isn't doing a lot of math. My guess is that it's just writing parameters out to a buffer and a separate compute shader is spun up after to actually do the material shading.

Yep, they run on the INT pipeline.

According to the profiler each SM has a peak instruction issue rate of 4.0 and peak measured is 3.2 (80%). It makes sense as an SM has 4 independent partitions each with its own instruction scheduler and execution units. What's bizarre is that the profiler reports peak issue rate for each of the two FP pipelines as 0.5 per clock. I would expect it to be 2.0 as each of the 4 partitions can issue to each FP pipeline every other clock. ALU peak is 2.0 as expected as it takes 2 clocks to process an ALU warp. SFU is 0.5 as expected as it takes 8 clocks to process an SFU warp. I don't know how to interpret the FP peaks.

 

That sounds like a very good theory.

This appears to explain the terms being used:

Advanced Learning :: Nsight Graphics Documentation (nvidia.com)

What I don't understand is why the picture in that document is not explained by the bullet points that follow. Light Pipe and Heavy Pipe?

This describes the pipelines:

Kernel Profiling Guide :: Nsight Compute Documentation (nvidia.com)

where the heavy FMA pipeline is different from the light one because it has integer dot product functionality. Other "integer" operations occur on other pipelines, the primary one being "alu".

Then we have the "fma" pipeline: "Fused Multiply Add/Accumulate. The FMA pipeline processes most FP32 arithmetic (FADD, FMUL, FMAD). It also performs integer multiplication operations (IMUL, IMAD), as well as integer dot products. On GA10x, FMA is a logical pipeline that indicates peak FP32 and FP16x2 performance. It is composed of the FMAHeavy and FMALite physical pipelines."

Notice that this isn't a real pipeline in Ampere, merely a convenience term to describe the grouping of the capability of the heavy and light pipelines.

The "alu" pipeline is for bitwise and boolean operations and some integer math.

As for the discrepancy in the peak rates, I have no idea!

Back to the subject of the benefits of FP32 throughput in NVidia ray tracing:

nvidia-ampere-ga-102-gpu-architecture-whitepaper-v2.pdf

"Ray tracing denoising shaders are a good example of a workload that can benefit greatly from doubling FP32 throughput." I have no idea what proportion of the frame time is spent denoising ray tracing results.

 

ASVGF takes about 1/5 of frame time in Q2RTX.
FYI, the best sample is a RTXDI SDK that includes latest versions of all RTX~ish technologies including NRD and you have full control to turn on and off every effect in the sample application, so you can count it by yourself.
https://github.com/NVIDIAGameWorks/RTXDI

 

In the three games there, merely comparing 99% percentiles for RT on 3070 and 2080Ti, we have:

• Battlefield - 37 v 44 - 2080Ti is 119% of 3070
• Control - 23 v 23 - 2080Ti is 100% of 3070
• Metro Exodus - 33 v 28 - 2080Ti is 85% of 3070

Without RT:

• Battlefield - 63 v 68 - 2080Ti is 108% of 3070
• Control - 38 v 29 - 2080Ti is 78% of 3070
• Metro Exodus - 52 v 52 - 2080Ti is 100% of 3070

Control seems to be the only "tough" test there, as Metro Exodus looks like it's the old version with just GI, not the Enhanced Edition.

So merely comparing NVidia to NVidia across the Turing/Ampere generation makes it hard to conclude much.

Why does Control punish the 3070 so much more than 2080Ti? 3070 has higher triangle intersection throughput, more FLOPS and "full fat" async compute. Maybe the game's implementation is old enough that it doesn't exploit async compute? With a question over the integer instruction mix for ray tracing, maybe instruction issue rate (concurrently to FMA heavy and FMA light pipes) is more important than FLOPS? Which would mean that 3070 has no advantage in instruction issue rate.

I worry whether all the results presented on that page are from the same driver.

 

This should be 120% faster. You don’t need to add another 100% when something is over 100% faster. It’s the same math as all the others.

 

Hardware-based techniques applicable for ray tracing for efficiently representing and processing an arbitrary bounding volume - NVIDIA Corporation (freepatentsonline.com)

Nvidia seems to be considering both hardware SBVH and OBB(sort of) for better traversal performance.

 

64% faster at 1440p. Good to see the 2070 super still hanging in there. It’s impressive how the 1060 and 580 are still neck and neck after all this time. Those were really well matched competitors.

 

Dont get this wrong... Everybody already knows, that RDNA2 has significantly lower Ray Tracing Performance. No need to spam the forums with every single game performance test you come across

AMD needs at least a 3x increase (which is rumored) to not be obliterated again this fall.