Crappy HD 4000 beating HD 7660D in openCL

[mrmt]

Funny no? Who would have guessed?

http://semiaccurate.com/2013/04/29/a-look-at-intels-opencl-performance/

Since this references GPU OpenGL performance and the thread title specifically refers to GPUs, thread moved from CPUs to GPUs
-ViRGE

 

[Abwx]

Clueless , as usual , Charlie and you dont know that OCL
perf is also CPU dependant...

So much for comparing a low end APU with a high end i7.

http://clbenchmark.com/result.jsp

 

[NTMBK]

Nice job from Intel there. Should be interesting to see how Haswell OpenCL does (and I expect Haswell's OpenCL on the CPU to improve a lot as well, with AVX2, FMA3 and gather).

Of course don't forget that VLIW4 was not that great at GPGPU- hopefully AMD will improve their OpenCL performance in their APUs when they move to GCN.

 

[BLaber]

Lol Open CL benchmark doesnt mean all benchmarking will run 100% on GPU , so obviously Intel CPU prowess will show in tasks where only so much can be offloaded to gpu.

best way to test this is to run same test on Intel proc with Hd 7660 card plugged into it vs same proc with HD4000 & system set-up.

On the contrary above test shows how decent tested AMD APU is

 

[SPBHM]

the new drivers gave a nice boost for OpenCL perf on the GPU

compared to 123 from the old drivers (GPU only)


but semiaccurate used some CPU+GPU tests,

 

[mrmt]

Lol Open CL benchmark doesnt mean all benchmarking will run 100% on GPU , so obviously Intel CPU prowess will show in tasks where only so much can be offloaded to gpu.

I agree with the criticism here, and sure, from a programmer POV it's interesting to know what is being processed where, because only this way you can optimize for your hardware.

But on the other hand, thinking about the consumer out there, this further weakens the case for an AMD APU. Performance is an important component in the purchase equation, for the consumer it doesn't matter whether the code is being processed in the GPU or in the CPU, he wants performance, and performance/watt in the case of mobile and servers. If not even in a highly optimal scenario AMD can get a lead, what's the point of that massive die and higher power consumption?

It would be nice to a performance/watt analysis of those benchmarks. We already know that Intel got the performance lead, but what about efficiency, who has the more efficient approach for GPGPU?

And... thinking about Haswell. Unless you assume that Intel iGPU does almost nothing of the OpenCL heavy lift, Haswell will bring significant improvements to openCL performance, far more than Richland can bring to the table, which means that Intel will have an even greater lead in what was supposed to be one of AMD competitive arenas.

 

[BLaber]

I agree with the criticism here, and sure, from a programmer POV it's interesting to know what is being processed where, because only this way you can optimize for your hardware.

But on the other hand, thinking about the consumer out there, this further weakens the case for an AMD APU. Performance is an important component in the purchase equation, for the consumer it doesn't matter whether the code is being processed in the GPU or in the CPU, he wants performance, and performance/watt in the case of mobile and servers. If not even in a highly optimal scenario AMD can get a lead, what's the point of that massive die and higher power consumption?

It would be nice to a performance/watt analysis of those benchmarks. We already know that Intel got the performance lead, but what about efficiency, who has the more efficient approach for GPGPU?
.

We are taking about Opencl benchmarks in this thread so any interested & serious consumer/buyers reading this thread does care of which task runs on CPU or GPU when it comes to Open Cl benchmarking.

Below is Ryans assessment from quoted SA article :

AMD’s APU on the other hand tends to perform best when it’s doing simulations or putting its raw GPU power to the test in synthetic benchmarks.

Which means in Open Cl tasks AMD APU Beat competing Intel CPU at least in performance , not sure about performance/watt but we can guess who wins performance/dollar if your are running purely open cl task on integrated graphics on these 2 tested cpus.

 

[jcniest5]

Not sure if it's accurate, but my EVGA GTX 680 SC only scored a 145,443 points RayTrace. Another GTX 680 scored a 176,118 on CLIBench's site. From looking at the score comparison on that site, my GTX 680 is slower than a GTX 580 (which scored a 148,809 points). Is that normal? And this is on a Z77 board with a 3770K CPU.

 

[zlatan]

Not sure if it's accurate, but my EVGA GTX 680 SC only scored a 145,443 points RayTrace. Another GTX 680 scored a 176,118 on CLIBench's site. From looking at the score comparison on that site, my GTX 680 is slower than a GTX 580 (which scored a 148,809 points). Is that normal? And this is on a Z77 board with a 3770K CPU.

It's normal. Kepler is generally bad at compute. Also NVIDIA don't really care about OpenCL.

 

[Stuka87]

I agree with the criticism here, and sure, from a programmer POV it's interesting to know what is being processed where, because only this way you can optimize for your hardware.

But on the other hand, thinking about the consumer out there, this further weakens the case for an AMD APU. Performance is an important component in the purchase equation, for the consumer it doesn't matter whether the code is being processed in the GPU or in the CPU, he wants performance, and performance/watt in the case of mobile and servers. If not even in a highly optimal scenario AMD can get a lead, what's the point of that massive die and higher power consumption?

It would be nice to a performance/watt analysis of those benchmarks. We already know that Intel got the performance lead, but what about efficiency, who has the more efficient approach for GPGPU?

And... thinking about Haswell. Unless you assume that Intel iGPU does almost nothing of the OpenCL heavy lift, Haswell will bring significant improvements to openCL performance, far more than Richland can bring to the table, which means that Intel will have an even greater lead in what was supposed to be one of AMD competitive arenas.

Sorry, but the consumer may care about the HUGE ($200) price difference between an AMD APU and an i7. Your comparison doesn't really make any sense since since the two CPU's are in entirely different classes. You are comparing a top end CPU with 8 threads to a lower end CPU with 4 threads. The two are NOT cross shopped.

 

[Phynaz]

Clueless , as usual , Charlie and you dont know that OCL
perf is also CPU dependant...

So much for comparing a low end APU with a high end i7.

http://clbenchmark.com/result.jsp

Are you calling mrmt clueless?

 

[Gikaseixas]

Are you calling mrmt clueless?

You're right, that was a bit harsh BUT the thread is so young, lets avoid these things.

Back on topic, i think this is a rare case where Intel uses CPU muscle to efficiently push the iGP. Some games can show us that too.

 

[Imouto]

That Luxmark score is pretty impressive. Just for comparison, my HD5850 at stock scores 610 points.

Someone running that test with HD 4000 + Discrete card would throw some interesting info.

 

[Ajay]

the new drivers gave a nice boost for OpenCL perf on the GPU

compared to 123 from the old drivers (GPU only)


but semiaccurate used some CPU+GPU tests,

Wow, lots on noise in that trace!

 

[Ajay]

It's normal. Kepler is generally bad at compute. Also NVIDIA don't really care about OpenCL.

It's really bad in OpenCL. NV refuses to update the OGL driver for Kepler and Kepler is stuck at v 1.1 - they just want to sell CUDA as long as they can (even on the desktop). It's annoying, because GK104 is fine in single precision (32b) compute (CUDA) and would be fine in SP OGL, except the driver are old an lose something like 30-40% of their likely performance.

I'd like to do some OpenCL eventually, but the performance drop from CUDA is discouraging.

 

[Imouto]

Wow, lots on noise in that trace!

Not sure if serious...

 

[Qianglong]

Wonder if one can mine LTC or bitcoin using the HD4000....

 

[Phynaz]

Also NVIDIA don't really care about OpenCL.

Color me confused as to why they are on the OpenCL BoD then.

 

[Stuka87]

Wonder if one can mine LTC or bitcoin using the HD4000....

Oh you could, it would just be horrifically slow.

 

[Jimzz]

Wonder if one can mine LTC or bitcoin using the HD4000....

I fired up my HD4000 in my 3570k and it was slower than the CPU if I remember.

 

[Yuriman]

I fired up my HD4000 in my 3570k and it was slower than the CPU if I remember.

With recent drivers it's faster than a CPU. Overall still slow - something like 6mhash, as compared to 300mhash of an HD7850.

 

[-Slacker-]

Cliff notes:

$300 CPU beats $100 CPU by an aggregate margin of 8% in tests that aren't entirely dependent on GPU performance.

I know semi accurate is desperate for hits, but can't they just stick with their usual rumor-news-thing stuff? That's why people read SA, right?

 

[Ajay]

Not sure if serious...

Yes, I'm working on ray-tracing for my Masters; that scene is objectively noisy.

 

[Imouto]

Yes, I'm working on ray-tracing for my Masters; that scene is objectively noisy.

Then you should know that there's no "finish" in a ray tracing bench like there's no "finish" in a ray tracing render.

Think about it.

 

[kevinsbane]

Yes, I'm working on ray-tracing for my Masters; that scene is objectively noisy.

That entire screenshot is crazy-noisy, including stuff that's not part of the picture. It's a 256 colour picture. You can see dithering everywhere, not just the ray-traced portion.