GPU 'upgrade' for nMP

[netkas]

Take a look http://barefeats.com/tube21.html

TitanX connected to nMP benchmarks.

Looks good for games. TB2 bus makes it slower than TitanX in cMP but still faster than D700s in OSX (no crossfire in osx)

Some details about setup: "Oh, and just one (UNO, SINGLE, ONE) system file has to be modded with our latest eEFI tech. And we have a GTX670 here that can boot and run like a standard Mac GPU, with NO SYSTEM MODS WHATSOEVER."

 

[lowendlinux]

Yea I think "upgrade" is correct. It works and that's great but..

 

[DearthnVader]

Take a look http://barefeats.com/tube21.html

TitanX connected to nMP benchmarks.

Looks good for games. TB2 bus makes it slower than TitanX in cMP but still faster than D700s in OSX (no crossfire in osx)

Some details about setup: "Oh, and just one (UNO, SINGLE, ONE) system file has to be modded with our latest eEFI tech. And we have a GTX670 here that can boot and run like a standard Mac GPU, with NO SYSTEM MODS WHATSOEVER."

Nice benchmarks, it's aways good to see someone pushing the envelope in Mac Graphics Cards. However timedemos tend to load into the gram of the graphics card in an, operation that's not timed.

In the case of the TitanX with it's 12GB of Vram, I'd say all the textures from the games and other benchmarks tested fit in the Vram, so there was no texture swapping between the Vram and main system memory.

If you used a card with less Vram in and eGPU setup, and ran the same benchmarks, would we see texture swapping between Vram and main system memory over a 20 Gbit/s bus of Thunderbolt 2?

Seems to me that cards with large amounts of Vram are ideal for eGPU, I'd be interested to see some benchmarks of cards with 2GB of Vram to see how they fair.

 

[netkas]

Interesting point. 6, 8 and 12 gb is enough these days. 4 gb are almost enough, unless you do some 4k gaming and maximize settings. these 2gb cards are taxed by vram even on pcs. Thing in the middle is 3gb cards, but they are going away slowly.

 

[Zorn]

Sigh...always great to get a reminder of the genius of the nMP design. A GPU upgrade that requires an expensive and large external box to house a full size GPU to get slower performance than if it were just plugged in internally. Joy.

 

[koyoot]

The problem is this: If the crossfire would be available for games in OS X, Dual D700 should be on par at 1440p in OS X Gaming as Titan X in cMP, and probably faster at 4K resolution, and is much faster in compute in OpenCL.

Unfortunately in OS X crossfire does not work for graphical purposes, and that is too bad.

 

[koyoot]

One more thing. If Metal is based on Mantle, this pretty much shows you the future. Because Mantle-ish APIs are allowing simultaneous work of GPUs from different vendors, you may be allowed to get performance from all 3 GPUs at the same time.

 

[thefredelement]

Does that mean two can be used with Thunderbolt3 eventually? This is pretty cool, probably more for the future than for right now but if you have 2013 nMP and want CUDA then this is awesome.

This is pretty awesome for the future, people will be able to take their rendering power with them or easily switch it between machines.

 

[DearthnVader]

Does that mean two can be used with Thunderbolt3 eventually? This is pretty cool, probably more for the future than for right now but if you have 2013 nMP and want CUDA then this is awesome.

This is pretty awesome for the future, people will be able to take their rendering power with them or easily switch it between machines.

I suppose if Apple doesn't include TB3 on the MP7,1 we have netkas

and MVC to blame.

Is PCI-E even a supported interface of TB3?

 

[ixxx69]

Is PCI-E even a supported interface of TB3?

Yes, of course. That's what TB is... an external PCIe interface (along with DP).

 

[Roykor]

One more thing. If Metal is based on Mantle, this pretty much shows you the future. Because Mantle-ish APIs are allowing simultaneous work of GPUs from different vendors, you may be allowed to get performance from all 3 GPUs at the same time.

I am not sure, but is Metal the future for Apple only? If the rest of the world with computers / software continue to stay on Cuda+OpenCL? Than you have Metal, but just some few programs support it. Right?

 

[netkas]

afaik, metal is something between mantle and opengl. not very low-level as mantle, not high level as opengl.

 

[666sheep]

The problem is this: If the crossfire would be available for games in OS X, Dual D700 should be on par at 1440p in OS X Gaming as Titan X in cMP, and probably faster at 4K resolution, and is much faster in compute in OpenCL.

Unfortunately in OS X crossfire does not work for graphical purposes, and that is too bad.

If, if, if...

 

[ActionableMango]

Nice work Netkas. You continue to amaze me.

 

[ManuelGomes]

Yeah, nice job there netkas.
Just curious, who runs barefeats and does all the testing?

 

[goMac]

One more thing. If Metal is based on Mantle

Vulkan is based on Mantle. Metal is not. Similar sounding names, but that's the only relationship. Metal for Mac is based on Metal for iOS, and Metal for iOS has nothing to do with Mantle.

 

[DearthnVader]

Does the eGPU report a Link Speed?

Does the nMP's x16 AMD GPU Link at PCIE 2.0( 5.0 GT/s ) or 3.0( 8.0 GT/s )?

Can someone with a new MacPro look under System Report>PCI>AMD Dxxx> Link Speed for me please.

 

[MacVidCards]

The internal GPUs are PCIE 3.

TB2 is run through PCIE 2.0 x4 lanes but in reality TB overhead/latency/something means it is less than x4.

The TB2 enclosure in those tests was Sonnet with a x8 elec slot.

 

[DearthnVader]

Thanks MVC, I must be missing something, I read somewhere that the maximum throughput of PCIE 2.0 4x was 2GB/s. Yet Apple has put a 20 Gbps Thunderbolt 2 bus on this link.

http://www.trentonsystems.com/applications/pci-express-interface/

Isn't 20 Gbps 2.5GB/s, greater than the bandwidth of 4x PCIE 2.0?

I'm a little confused.

Anyway, it seems the nMP has 48 PCIE lanes, 40 PCIE 3.0 lanes on the Xeon and 8 PCIE 2.0 from the Intel C600 PCH.

On the Intel C600 PCH( PCIE 2.0 ):
1x lane goes to GigE Controller
1x lane goes to the other GigE controller
1x lane goes to the 802.11ac controller
1x lane goes to the USB3 controller
4x lanes go to the PCIE SDD controller

On the Xeon( PCIE 3.0 ):
16x lanes go to the AMD Dxxx
16x lanes go to the other AMD Dxxx

Leaving 8x PCIE 3.0 lanes for the six Thunderbolt 2 ports, but Thunderbolt 2 seems to have been built on the PCIE 2.0 spec and Apple needed a Hub to route 6 ports over three busses.

http://www.anandtech.com/show/7603/mac-pro-review-late-2013/8

I think what all this means when it comes to the eGPU is, the e GPU should be equal to 8x AGP 3.0 in terms of throughput, or close.

Apples TB 2 = 2GB/s
8x AGP 3.0 = 2.1GB/s

But with Apples TB 2 there are three busses, so we could connect 3 eGPU's to the right TB ports and not be sharing bandwidth. In the case of an nVidia card we'd have 3 Cuda devices each on their own 2GB/s link to the CPU.

 

[MacVidCards]

The nMP has been an enigma.

It is difficult to get eGPU to work with it. Somehow the bridge chips have a difficult time dealing with more than 1 high bandwidth device AT ALL.

I doubt that 3 eGPUs could be made to work, I have been unable to get 2 working.

 

[DearthnVader]

The nMP has been an enigma.

It is difficult to get eGPU to work with it. Somehow the bridge chips have a difficult time dealing with more than 1 high bandwidth device AT ALL.

I doubt that 3 eGPUs could be made to work, I have been unable to get 2 working.

Hmmm..........

That must be why no Co. ever brought a eGPU to market.

So you're saying that to connect an eGPU, one high bandwidth device, we may lose something like the ThunderBay 4, another high bandwidth device?