Need OpenCL boost today. No way.

[asiga]

I feel sad how hard is to buy a powerful Mac nowadays. Because of a new project that I'm doing these weeks, I need high OpenCL performance right now. The only options I've is to buy a Mac Pro (not updated since 2013) or an iMac (mobile GPU, and supposedly to get an upgrade soon).

It's really disappointing. And this is happening in Apple, who almost invented OpenCL. This is really disappointing.

I was going to buy a new Mac today, but, honestly, I can't. And not because of my wallet, but because of Apple ineptitude to provide current OpenCL performance.

 

[keysofanxiety]

I feel sad how hard is to buy a powerful Mac nowadays. Because of a new project that I'm doing these weeks, I need high OpenCL performance right now. The only options I've is to buy a Mac Pro (not updated since 2013) or an iMac (mobile GPU, and supposedly to get an upgrade soon).

It's really disappointing. And this is happening in Apple, who almost invented OpenCL. This is really disappointing.

I was going to buy a new Mac today, but, honestly, I can't. And not because of my wallet, but because of Apple ineptitude to provide current OpenCL performance.

What project/application do you need? If we have an idea of your usage there's bound to be an economically viable Mac that should run this. Even with OpenCL the CPU will do a lot of the work, so don't feel too concerned about graphics performance.

 

[asiga]

It's a research project. I've written the OpenCL code myself. But my GPUs are old. I need current GPUs for the research to be on focus.

Btw, is there any easy way of getting OpenCL LuxMark results (the GPU-only result) for the complete current Apple product line? I've searched the net for hours, and no way for getting the results for the current MacPro, the current iMacs, the current Mac Mini, the current MBP...

Note that I'm interested in the GPU-only benchmark. I don't care about the CPU.

 

[SoyCapitanSoyCapitan]

It's a research project. I've written the OpenCL code myself. But my GPUs are old. I need current GPUs for the research to be on focus.

Btw, is there any easy way of getting OpenCL LuxMark results (the GPU-only result) for the complete current Apple product line? I've searched the net for hours, and no way for getting the results for the current MacPro, the current iMacs, the current Mac Mini, the current MBP...

Note that I'm interested in the GPU-only benchmark. I don't care about the CPU.

Barefeats does all the tests

 

[asiga]

Barefeats does all the tests

Yes, you're right. Thanks a lot. From what I see, the highest iMac 5k 27inch can come close to a current Mac Pro in terms of OpenCL performance, and it can be considered "current" (although its price is about the same as a Mac Pro). On the other hand, I see the new 15inch MBP has unexpectedly low OpenCL performance... in the lines of the Mac Mini 2014 Iris Graphics, even if the MBP has a new AMD GPU.

I'll continue thinking. Thanks. Anyway, buying a new Mac Pro today seems unwise to me because they're likely to be updated this year. Buying a 5k iMac seems wiser because it's newer... but price is about the same, and performance not as good as the Mac Pro

 

[mathpunk]

If you're writing new OpenCL code now, you might consider looking at OpenCL 2.0. And if you want performance right now, I think the only option is to assemble a workstation (Windows 10 or Linux) around a FirePro W9100 card.

 

[asiga]

If you're writing new OpenCL code now, you might consider looking at OpenCL 2.0. And if you want performance right now, I think the only option is to assemble a workstation (Windows 10 or Linux) around a FirePro W9100 card.

Sadly, this is the impression I'm arriving to. I'd like to avoid it because I do all my work on OSX. But, on the other hand, I see it unwise to buy any Apple product until at least the September event, or even later. How disappointing is this.

 

[koyoot]

Barefeats uses old, outdated benchmark to count performance of AMD GPUs, and favor Nvidia GPUs. For raw horsepower in compute performance he should use Luxmark 3, not 2, or 2.1.

But all in all. Yes, Fire Pro D300 and M290X will have the same performance, and D700 will have the same performance as M295X. They have similar core count, similar core clock. There is easy way to count the performance of AMD GPU:
Core conut x2/core clock. 2048 GCN cores x2/850 MHz gives 3481.6 GFLOPS of compute OpenCL power.

 

[asiga]

As an alternative solution, do you know of any remote testing lab which would allow me to upload and run my OpenCL app in several current Mac models so that I get an idea on what model would be my best investment? Doing a search I found MacinCloud, but it seems just a server without graphics, so I don't think I could test OpenCL apps there.

 

[deconstruct60]

....Doing a search I found MacinCloud, but it seems just a server without graphics, so I don't think I could test OpenCL apps there.

If it is a dedicated Mac system (i.e., no Virtual Machine or ParaVirtiualization) then it it has graphics. There are no Macs sold without graphics. The "user quota" they have on storage space may be restrictive. A little odd if no VM to be doling out dedicated drives where don't use much ( pricing to get to full disk higher , but slimmed down disk space leaves alot of space not making money and can't make money unless backups across machines ... which seems like a huge potential security hole. )


http://macstadium.com/ has dedicated server offerings ( mini and Mac Pro). Not particularly "short term" ( daly/hourly ) options though.

 

[deconstruct60]

If you're writing new OpenCL code now, you might consider looking at OpenCL 2.0. And if you want performance right now, I think the only option is to assemble a workstation (Windows 10 or Linux) around a FirePro W9100 card.

The W9100 isn't the "only option". There are other cards that are viable OpenCL 2.0 solutions.
Apparently data integrity isn't an major issue for highest end , ECC capable card isn't the only option.

You can rent HPC time/space too.

https://aws.amazon.com/hpc/

Not "maximum" GPGPU cards but work on multiple variations at the same time with multiple instances.

 

[asiga]

You can rent HPC time/space too.

https://aws.amazon.com/hpc/

Not "maximum" GPGPU cards but work on multiple variations at the same time with multiple instances.

Forgive my ignorance, it's the first time I hear about AWS, and trying to understand the docs I see a lot of web dev stuff, which isn't my area. Talking about the stuff I know and understand, can AWS be used just like Virtualbox, run a Windows image, and upload a Windows executable compiled on my real machine? Can it run a standard windows executable with OpenCL, or do I need to use a custom API or is it required to compile the source on the cloud?

I'm asking this because I use wxWidgets+OpenCL+OpenGL as my only APIs, and moreover my build system cross compiles from OSX to Windows, so I get both OSX and Windows executables from my Mac. If I can just throw one of these Windows executables into the Amazon AWS, this would save my day.