Linux OpenCL Performance With The Newest AMD & NVIDIA Drivers

[smitty3268]

Yep, given the way that nVidia INTENTIONALLY gimps gpgpu capabilities of their consumer cards I was incredibly surprised to see 780 TI perf so close to the R9 290X and exceeding even the more modest ATI cards.

Note that Michael is only testing the 290. He doesn't have a 290X card.

[Calinou]

Kepler and Maxwell (600/700/800) series target gaming, not GPU computing, which is the inverse of what Fermi (400/500) did.

[johnc]

Kepler and Maxwell (600/700/800) series target gaming, not GPU computing, which is the inverse of what Fermi (400/500) did.

This seems to be quite the myth. In terms of GPU compute, the Kepler architecture is far more performant, especially performance per watt, than Fermi.

It is true that they removed FP64 units from Fermi to Kepler but that's because FP64 workloads are practically non-existent in the consumer space and it was a waste of die space.

Unfortunately most of the people who did benchmarks at the time of Kepler's release were idiots and don't understand GPU compute and now everyone thinks Kepler is a cripple because of it.

Nobody who does GPU compute anything is going to prefer Fermi to Kepler, especially when power requirements are taken into consideration.

[liam]

Is there a danger that GPU-based computing (OpenCL, CUDA, HSA, etc.) is going to be replaced by FPGAs? Probably certainly not in the consumer space (where these technologies are rare anyway), but in HPC, which could lead to these technologies, in time, withering on the vine.

I'm just thinking about how quickly GPU mining collapsed based on a market need to go further than what GPUs can do. Would the same pressures apply to typical HPC markets today?

Some half-interesting viewpoints from an incorrigible crank: http://semiaccurate.com/2014/06/20/i...e-desperation/

Both Intel and Microsoft are making use of fpgas. Intel is planning on including them in a Xeon socket compatible product sometime soon. Msft focused their efforts on the bing ranking engine, specifically "most of the" feature computations, but all of the machine learning model and free form expressions are what they accelerated on the FPGA.
I'm assuming Intel is going to start making these. If so I'd expect prices to come down greatly.
I've said before that including a small FPGA on embedded systems, especially industrial ones with very long life cycles, makes a lot of sense. Even on standard mobile systems I can see a use for them in accelerating media processing, aggregating/creating projections of sensor data, and, most interestingly, completely unforseen uses.

[drSeehas]

Why HD 7850 AND R9 270X?

What is the rationale behind testing HD 7850 AND R9 270X?
It is the same chip (Pitcairn)!
Instead why not test a card with a Cape Verde chip to complement the 740?

[Sdar]

Yep, given the way that nVidia INTENTIONALLY gimps gpgpu capabilities of their consumer cards I was incredibly surprised to see 780 TI perf so close to the R9 290X and exceeding even the more modest ATI cards.

Cuda performs better than Opencl on Nvidia cards, and when you try more complex kernels surpass amd for a good margin, at least until AMD decides to fix the compiler... but after three years broken I'm not hopping for a real fix anymore.

[asdfblah]

I can propose couple of another benchmarks:
1) bfgminer --scrypt --benchmark (https://github.com/luke-jr/bfgminer.git) - massively parallel computations with incredibly heavy GPU memory demands. GPU should both be good at parallel computations and provide fast memory. Can be tricky a bit in sense that best results are obtained after tuning parameters for particular GPU.
2) clpeak utility (https://github.com/krrishnarraj/clpeak.git). GPU VRAM speed benchmark. While it sounds simple, it depends on both GPU and drivers so it can be quite interesting thing to compare. This one also known good way to crash MESA+LLVM OpenCL stack, at least on AMD cards .

I also suggest "John The Ripper" with jumbo patches: https://github.com/magnumripper/JohnTheRipper

[Michael]

What is the rationale behind testing HD 7850 AND R9 270X?
It is the same chip (Pitcairn)!
Instead why not test a card with a Cape Verde chip to complement the 740?

The cards were tested for what I had in my possession...

[Marc Driftmeyer]

These are craptastic Linux OpenCL performance results and still AMD wins out.

The Windows optimized OpenCL tests shows AMD blowing the doors off of Nvidia.

[drSeehas]

The cards were tested for what I had in my possession...

I know, but why don't you simply buy a card with a Cape Verde chip (e.g. HD 7750, some R7 250)? These are not very expensive.

... craptastic ...

What's "craptastic"?