• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

OpenCL machine..

Caporegime
Joined
13 Jan 2003
Posts
25,086
Ok, my background is simple.

My first MBP 15" was struck down by the nvidia-gate just after the replacement program ended.

My second MBP 15" has the doom Radeon 6750M .. it's GPU is failing with hangs etc.

I want to build an OpenCL crunching machine under linux/OSX. OpenGL performance isn't so much of a problem as this isn't about gaming. The GPU should have a decent processing performance in each - minimum of 2GB ram and 512 pipelines.
The work it's doing is realtime image processing.

I have had direct AMD OpenCL compiler issues - I found a bug in the runtime OpenCL driver that doesn't have a workaround that AMD have verified as an issue but are not committing to a fix for not can they say when Apple would have a fix for. So as this seems to be a reoccurring issue with AMD that others have had.. screw them.

I'll be sticking with OpenCL as it's semi portable at least..

Now that leaves me with two options:
* Apple with Intel Iris Pro (current top MBP has a Radeon - so scratch that model)
* Linux box with nVidia/Intel GPUs


Now portability is a plus point.. but Intel Iris Pros are unknown to me.

Any thoughts?
 
Wait you're upset with AMD because they can't tell you when a fix is coming from Apple on opencl issue.......that makes sense. Remember opencl is pretty much controlled by Apple; so you should also be submitting this issue to them *but getting them to fix issues is always fun*

Have you checked to see if this issue effects Nvidia? Its also known AMD's cards are faster in opencl than Nvidia.....

With that said; AMD runs opencl better but if you're effected by a bug that you don't know when it will be fixed....well guess decide what's in your budget and go with that.

The bug, believe it or not, was discovered whilst using the clFFT library (an AMD developed library). On investigating, AMD have stated that it's an OpenCL runtime issue within the driver (AMD again) that Apple use. The additional point is that if AMD fix it, then apple has adopt the fixed version.
What happens is that for random sized textures, the GPU just simply returns from the runtime without error (so can't detect) but all you receive is a blank texture.


Yep - my mistake - nVidia GPU on the topped MBP:
Intel Iris Pro Graphics & NVIDIA GeForce GT 750M with 2GB of GDDR5 memory and automatic graphics switching.


In terms of cross platform, I think I'd probably stick with OpenCL as it allows (although lower performance) some of the ARM embedded devices to run it. It means I can produce a cut down processing pipeline for specific tasks that the embedded system could be used for.

For the larger tasks - it really looks like the nVidia GPU is the only way togo but at the same time it effectlvely removes the main mac users that have Intel or AMD (again using specific texture sizes and cropping to fit as a limp home measure).

I'm thinking that it may be better to make a processing box for the heavy/development use. Pity the MBP has been made an iPad..

quoting apple Up to 90% faster than previous-generation MacBook Pro with Intel HD Graphics 4000
Hmm that's not a good starting point lol .. and after the first two MBP failing for the same reason - I'm rather reluctant to fork out £2000 for one that would replace the one I have now in terms of capabilities..
 
Last edited:
The rMBP is very powerful bit of kit. I got one as a replacement for a problem 2011 expecting to be underwhelmed and it's a cracking bit of kit if you ignore the missing ports sacrificed for "thin-ness". The Intel Iris Pro is surprisingly capable and runs the nVidia 750 very close.

Hmm interesting.. how did you get a replacement? My system log and I have images of the GPU hangs available to back my statement up.

The annoying thing is that the MBP is a nice small piece of kit which makes it both useful for astronomy and for keeping the Mrs happy. I think the mrs would prefer that than having a linux box lying around but not the price tag (until I get back into employment after being laid off late Nov).
 
Last edited:
I'd check to see if that bug hasn't been fixed already....also which card are you using? normal gaming card or a firepro? there have been a lot of fixes; which humbug pointed out one of the reasons Apple did switch from Nvidia to AMD in their graphics as opencl is now gaining a lot of traction.

Nick that clears a bit up on it. its one of the cart before the horse things; but I'd check I've seen a lot of bugs that have been squashed lately.

The current GPU is a 6750M. The problem was passed from the AMD front end folkts to the AMD OpenCL team who came up with the diagnosis (after about 3 months!).

My thinking here is that it's good to design the system for Intel Iris *if* that is the current mainstay of people's mac experience for the use of the application. The OpenCL application is used for realtime astronomy observations - this means laptop usage next to the telescope (or inside the tent for warmth!). I also have a process in mind that would be non-realtime but the level of processing would benefit from something heavier duty - which could benefit from the power but could equally be processed on a slower GPU for a longer period.

Only thing is that my current MBP is already doomed :/ :)
 
Last edited:
Reading this.. http://www.anandtech.com/show/8568/the-geforce-gtx-970-review-feat-evga/14 it appears that the AMD R9 wipes the floor with the nVidia GPU for OpenCL - probably the GPU core design have a better performance but the nVidia cores can be clocked higher so you get more cycles resulting in a higher single precision performance.

However I'm accutely aware of the benchmark games played by both sides and that these are probably the Windows non-opensource drivers too.

I'd be happy to run with a slower GPU in the laptop but that R9 double precision is probably worth verifying the results for for a box based configuration.
 
Last edited:
Perhaps a cheap NV card and a cobbled-together Linux system in the meantime? Or one of those external laptop GPUs?

http://www.pcworld.com/article/2839...you-add-a-desktop-video-card-to-a-laptop.html

needing up to 375 watts of electrical power

Hmm the 8GB graphics cards are indicating a minimum of 650W. Nice idea though.

Not sure when Apple's next hardware refresh is - the annoying bit is an 8GB GPU linux box, even with duff drivers, is likely to return performance better than the MBP.. but without being quite as portable.

On the astronomy side of things - I've been using a mac mini and screensharing but that's not ideal due to the slow frame rate. The portable element may result in me going for another MBP with an average GPU again. It means I get to support OS X and Linux from the same machine.
 
The 6750M has peak 576 GFLOPS. Is that enough power? Because even the Intel 5000 (GT3) has 704 GFLOPS, bumped up to 832 for the Iris Pro 51/5200 (GT3/e). I can't imagine you'd have much trouble tuning for it over AMD/Nvidia GPUs.

All this assumed SP rather than DP.

If 700-830 GFLOPS is enough, and you value portability, I'd say go with the macbook.

Edit: or if saving money is a priority, you could go as low as an R7 250 (£66) with 806 GFLOPS SP, similar to the Intel offerings (with 1 GB GDDR5, if you really need 2 GB you've got to bump up to a 260/X, but they're substantially faster at 1536/1971 GFLOPS respectively).

Enough powah?!?! :D hehe - one thing to note here is that OpenGL GLFOPS often differs from OpenCL performance. nVidia's non-compute GPUs typically have disabled reduced processing for OpenCL vs OpenGL.

No - seriously, I can add more into the pipeline and make use of more GPU power and memory. The issue is that the more that is added the more memory is usually required. The MBP has 1GB so it's a good starting point as the mainstream machines have about that. Note with OpenCL you will typically get about 40% of memory max by default however it's possible to push that up but the OS will take a portion.

However I get your point - looking at the relative performance means any is really going to be a match for the 6750M...

The main issue normally is memory bandwidth - there's a massive requirement here as processing is usually doing, basically, 4K textures.
 
I had an idea - why don't you first build the rest of the machine OP and get two equivalent AMD and Nvidia cards and try both with the software you are using. You can send the card you don't want back and pay any restocking costs. That way you will know definitely what is the better solution??

That's a good idea!
 
Back
Top Bottom