Frames per Joule.

We have fallen for the folly of maximum GPU performance. It is the age of 450 Watt GPUs. Why is that?

There is absolutely no need to have 450 Watt GPUs. The sole reason they use that much, is for a few additional FPS in the benchmarks. For real-world use, that 450 Watt is useless, as I have come to conclude after running an experiment.

Traditionally, I have always used NVIDIA GPUs in my Linux boxes, for stability and performance. But today, I am running an (elderly) AMD RX580 in my main rig. NVIDIA on Linux does not give you much control over clock speeds and voltages, but CoreCtrl let's me adjust the clocks for the RX580.

I already knew that for the last few percent of performance, you need a heap more wattage. This is because Wattage increases linearly with clock speed, but quadratically to the Voltage. Increasing Voltage comes at a huge cost. Let's quantify that cost.

I benchmarked my Photon Mapping engine on two settings: 900Mhz GPU, 1000 MHz VRAM versus: 1360MHz GPU, 1750 MHz VRAM.

I then checked the FPS and the (reported!) Wattage. I guess for a more complete tests, measuring power consumption at the wall socket would have been better. Maybe another time.

So, with the higher clocks, we expect better performance of course. Did we get better performance?

     Clocks    FPS     Volt   Watt
   900/1000     38    0.850     47
  1360/1750     46    1.150    115

The result was even more extreme than I expected: For 145% more power consumption, I got to enjoy 21% more frames/s.

Not only that. The high-clock setting came with a deafening fan noise, whereas I did not hear a thing at low Voltage.

Which now makes me ask the question: Shouldn't those benchmarking youtube channels do their tests differently? Why are we not measuring --deep breath-- frames per second per (joule per second) instead? (As Watt is defined as Joule/Second.)

We can of course simplify those units as frames per joule, as the time units cancel each other out.

Hot take: LTT, JayzTwoCents, GamersNexus should all be benchmarking GPUs at at lower Voltage/Clock setting. I do not care if RTX inches out a Radeon by 4% more FPS at 450Watt. I want to know how they perform with a silent fan, and lowered voltage. Consumer Report tests a car at highway speeds, not on a race track with Nitrogen Oxide and a bolted on Supercharger. We should test GPUs more reasonably.

Can we please have test reports with Frames per Joule? Where is my FPJ at?

UPDATE: I can't believe I missed the stage-4 galaxy brain. We can invert that unit to Joules/Frame instead!

The lacking state of rPi4 GPU support.

Because I am a sucker for Linux SBCs, I got myself a Sipeed Lichee Pi 4A RISCV SBC. And remarkably, it can do OpenCL too! Which means I can try my Photon-Mapping CL kernel on it.

Even though it can run OpenCL kernels, I did find that the OpenCL implementation is not the most stable one: As soon as I try to fill a buffer with clEnqueueFillBuffer() it will crash.

The performance of my kernel on this SBC was incredibly bad though: three orders of a magnitude slower than a desktop GPU. Which made me wonder what the performance would be on my Raspberry Pi 400.

Eben Upton designed the VideoCore IV GPU in the original rPi. So surely, the software support should be excellent. Sadly, there is no OpenCL for the rPi4.

CLVK to the rescue! This sofware will translate OpenCL kernels to Vulkan compute kernels. Even though RaspberryPi OS would not work with it, Ubuntu 23.04 for rPi would, as it had a newer Mesa.

Which brings us to the next disappointment: even though there is a Vulkan driver for the rPi4, this driver lacks important Vulkan extensions. It cannot handle 16-bit floats, nor can it handle 8-bit ints.

  unknown NIR ALU inst: vec1 16 div ssa_217 = b2f16! ssa_182

This leaves us with a rather lacking implementation. Since I wrote my Photon Mapper for FP16, this experiment has to be shelved.