MacPro Multi GPU onboard

[yfile]

If this was your question...

...a lot of people has answered it. You just don't like the answers.

IMGTEC claiming they have GPU reaching the TFLOPS and "advanced scalable compute cluster architecture".

I don't like answers: "I think it's impossible".

 

[Asgorath]

IMGTEC claiming they have GPU reaching the TFLOPS and "advanced scalable compute cluster architecture".

I don't like answers: "I think it's impossible".

My point is that once they actually ship such an architecture, they'll still be way behind the high-end solutions from AMD and NVIDIA. Once Imagination has a 1 TFLOPS product (and yes, the article you linked said "100s of GFLOPS, up to the TFLOPS range"), the latest NVIDIA product may very well be in the 10 TFLOPS range. What if it actually takes several chips to reach that 1 TFLOPS mark?

I don't think I ever actually said it's impossible for them to try this. My point is that I think it's highly unlikely that they'll succeed if they do. Imagination is arguably one of the dominant players in the growing mobile space. I'm not sure why they'd want to get into a niche high-end desktop market, which is already saturated with mature products that are miles ahead of where they are.

 

[goMac]

There are a few problems with this thread...

The 1 TFLOP Power6 is not suitable for mobile phones, don't compare to what is in the iPhones. It is a Power6 designed for game consoles and the like. 1 TFLOPS is the very high end for non mobile devices, the chip actually starts at 100 GFLOPS.

http://www.nordichardware.com/Graph...ses-powervr-6-more-powerful-than-hd-7750.html

PowerVR G6200 and G6400 are graphics porcessors that will appear in future system processors for smartphones and other units. For the actor that wants to develop a system processor and is not limited by a tight energy budget the new G6230 and G6430 are perfect. Imagination says that the PowerVR 6 series will offer performance from 100 GFLOPS (Giga Floating-Point Operations Per Second) up to 1 TFLOPS for the most powerful alternatives. This sets the new candidates above Radeon HD 7750 in performance, at least in theory.

So likely the 1 TFLOPs chip will be a large, expensive, power sucking chip.

Meantime the Radeon 7970 can do 3.8 TFLOPs for probably a very similar, if not cheaper price, with no custom board design. Why would Apple consider the PowerVR 6 in desktops?

 

[666sheep]

PowerVR chips are designed for SoC and due that they don't have their own memory or PCIe controller. They won't work "onboard" without those.

More likely you'll see PowerVR GPU in MBA (if it will be based on one of future ARM Cortex chips).

 

[goMac]

More likely you'll see PowerVR GPU in MBA (if it will be based on one of future ARM Cortex chips).

With news today that Intel is working on very low power x86 chips and Apple will possibly adopt them for the iPad, I think the chances of Apple moving the Macbook Air to ARM are extremely unlikely, especially in since that will just boost Apple's dependence on Samsung.

 

[yfile]

I don't think I ever actually said it's impossible for them to try this. My point is that I think it's highly unlikely that they'll succeed if they do. Imagination is arguably one of the dominant players in the growing mobile space. I'm not sure why they'd want to get into a niche high-end desktop market, which is already saturated with mature products that are miles ahead of where they are.

The point is to transform this to the mass workstation market - standard OpenCL hardware used by all MAC pro software.

----------

So likely the 1 TFLOPs chip will be a large, expensive, power sucking chip.

Soon we will see how good (or bad) are these processors.

----------

PowerVR chips are designed for SoC and due that they don't have their own memory or PCIe controller. They won't work "onboard" without those.

OK, so such cluster (SoC chip or system) need own memory and PCIE controller.

 

[goMac]

The point is to transform this to the mass workstation market - standard OpenCL hardware used by all MAC pro software.

They already have that. It's the graphics card. Software is free to adopt it as it likes.

 

[Asgorath]

They already have that. It's the graphics card. Software is free to adopt it as it likes.

Exactly this. Every Mac already has access to an OpenCL device, it's the GPU in the system. Mac Pros are lucky in that they get access to the most powerful GPUs of any Mac, and thus the most GPGPU horsepower. It's not like a PCIe card with 10 PowerVR chips on it will bring anything new to the table, the Mac Pro cards have supported OpenCL from day one.

 

[yfile]

Exactly this. Every Mac already has access to an OpenCL device, it's the GPU in the system. Mac Pros are lucky in that they get access to the most powerful GPUs of any Mac, and thus the most GPGPU horsepower. It's not like a PCIe card with 10 PowerVR chips on it will bring anything new to the table, the Mac Pro cards have supported OpenCL from day one.

I know that. FCPX is an example of such great software.
I though about creating the best possible mainboard for Mac. Faster than anything in workstations world. Later you could add more graphic cards or "clusters".

 

[Asgorath]

I though about creating the best possible mainboard for Mac. Faster than anything in workstations world. Later you could add more graphic cards or "clusters".

You're missing my main point. By the time the PowerVR cluster is released, the existing high-end desktop solutions will be even more powerful than they are now, and still bordering on an order of magnitude more than the PowerVR solution. It's not like AMD and NVIDIA are suddenly going to say "okay guys, this one is powerful enough, let's stop innovating and driving forward" so that PowerVR can catch up.

Edit: And yes, I can believe that their target market is the next-gen consoles, not the high-end desktop.

 

[goMac]

I know that. FCPX is an example of such great software.
I though about creating the best possible mainboard for Mac. Faster than anything in workstations world. Later you could add more graphic cards or "clusters".

This system already exists. It's called PCI slots and GPUs.

You know a GPU is really just hundreds of processors working together, right? A GPU is the same thing you're talking about. A board of smaller processors.

 

[yfile]

You're missing my main point. By the time the PowerVR cluster is released, the existing high-end desktop solutions will be even more powerful than they are now, and still bordering on an order of magnitude more than the PowerVR solution. It's not like AMD and NVIDIA are suddenly going to say "okay guys, this one is powerful enough, let's stop innovating and driving forward" so that PowerVR can catch up.

Edit: And yes, I can believe that their target market is the next-gen consoles, not the high-end desktop.

I don't missing your point. But it is likely that performance of PowerVR will grow much faster than AMD and NVIDIA.

----------

This system already exists. It's called PCI slots and GPUs.

You know a GPU is really just hundreds of processors working together, right? A GPU is the same thing you're talking about. A board of smaller processors.

OK, this is an example what I mean:
http://rog.asus.com/116882012/news/asus-zeus-concept-motherboard-with-dual-gpu-onboard/
But I'm looking for something more power efficient and make a room for PCIE slots.

 

[ActionableMango]

IMGTEC claiming they have GPU reaching the TFLOPS and "advanced scalable compute cluster architecture".

I don't like answers: "I think it's impossible".

Oh, I see. The only valid answers are ones that agree with you. Then why ask?

 

[Asgorath]

I don't missing your point. But it is likely that performance of PowerVR will grow much faster than AMD and NVIDIA.

Based on what, exactly? Imagination's long track record of building high-end GPGPU processors, as opposed to AMD/NVIDIA's 15+ years of building GPUs? Don't forget that PowerVR used to be a player in the desktop space, but gave up when they couldn't compete and switched focus to low-power mobile chips.

Why are you obsessed with a PowerVR solution when there are already compelling products from AMD and NVIDIA? Not sure I understand why you really want to see a PCIe card with 10+ PowerVR chips when an NVIDIA GK110 is way more interesting and powerful.

 

[yfile]

Oh, I see. The only valid answers are ones that agree with you. Then why ask?

No. I'm looking for reasonable answers and I received a few. Simple "It's impossible" is not enough.

----------

Based on what, exactly? Imagination's long track record of building high-end GPGPU processors, as opposed to AMD/NVIDIA's 15+ years of building GPUs? Don't forget that PowerVR used to be a player in the desktop space, but gave up when they couldn't compete and switched focus to low-power mobile chips.

Why are you obsessed with a PowerVR solution when there are already compelling products from AMD and NVIDIA? Not sure I understand why you really want to see a PCIe card with 10+ PowerVR chips when an NVIDIA GK110 is way more interesting and powerful.

I'm not obsessed but impressed with this little chips.
Maybe a better solution would be to use 2xGTX680MX on motherboard instead of 1 of 2 CPUs.

 

[goMac]

OK, this is an example what I mean:
http://rog.asus.com/116882012/news/asus-zeus-concept-motherboard-with-dual-gpu-onboard/
But I'm looking for something more power efficient and make a room for PCIE slots.

I don't think the Power7s you're talking about are necessarily more power efficient.

Plus you need a real GPU anyway, so you aren't really saving anything.

Apple may end up integrating a GPU, but it would be because of requirements for Thunderbolt (which is why ASUS did it), not for OpenCL.

 

[yfile]

I don't think the Power7s you're talking about are necessarily more power efficient.

Plus you need a real GPU anyway, so you aren't really saving anything.

Apple may end up integrating a GPU, but it would be because of requirements for Thunderbolt (which is why ASUS did it), not for OpenCL.

You saving OpenGL performance for example.

Integrating a GPU(s) on motherboard has nothing to do with Thonderbolt. Asus included Thunderbolt because lost PCIEs.

 

[goMac]

Integrating a GPU(s) on motherboard has nothing to do with Thonderbolt. Asus included Thunderbolt because lost PCIEs.

No they didn't. GPUs on a motherboard still use PCIe lanes. You don't gain any PCIe lanes back.

Notice the ASUS motherboard has Thunderbolt, yet two less PCIe slots. They didn't gain any PCIe slots, they lost two, one for each GPU.

Edit: Actually, scratch that, that motherboard has zero PCIe slots.