Poll: What GPU are you running?

[AidenShaw]

I'm curious what the "Other" vote is.

Other

​

 

[chriz_r]

Haha nice

 

[Blair Paulsen]

Unflashed nVidia Maxwell GTX980/4GB on web driver in 5,1 12c cMP

 

[ActionableMango]

Other

View attachment 672806​

I meant this poll for Mac Pros, but I'll let that slide since it is awesome.

 

[t0rr3s]

MSI GTX970. The 5770 I use for when I need bootscreen (which is hardly the case).

 

[AidenShaw]

I meant this poll for Mac Pros, but I'll let that slide since it is awesome.

And awesome that a system has power connectors for five 8+6 GPUs (they're Titan X (Maxwell)).

 

[Flint Ironstag]

Aiden, can you share the full specs of that beast? Very curious about the motherboard, chassis, and power supply(s) chosen.

 

[AidenShaw]

Aiden, can you share the full specs of that beast? Very curious about the motherboard, chassis, and power supply(s) chosen.

OK.

• Standard BTO configuration of HPE ProLiant DL580 Gen9 4U rack server
• Quad socket HPE mobo, C602J chipset, 96 DDR4 DIMM slots
  • Max of 6 TiB of RAM supported
  • This system has 1 TiB (actually 31 sticks of 32 GiB - the Maxwell GPU can't address 1 TiB)
• Four E7-8890 v3 processors
  • 18 cores / 36 threads per processor
  • 72 cores / 144 threads total
• Quad 1500 watt 240 volt power supplies (N+1 redundancy at max TDP)
• Nine PCIe 3.0 slots
  • Five PCIe 3.0 x16 (three double-spaced so a double-width card doesn't hide another slot)
  • Four PCIe 3.0 x8 (all physical x16)
• Five Titan X (Maxwell) GPUs
  • 12 GiB VRAM each
  • 3072 CUDA cores each
• Storage
  • 400 GB SAS SSD 12Gbps system drive
  • Five 1.6 TB NVMe drives in 8 TB RAID-0 stripe for data
    • In between the two Titans on the right is a PCIe 3.0 x8 to PCIe 3.0 x24 switch that fans out x4 connections for the NVMe drives
• Network
  • Two 10 GbE SFP+ converged NICs
    • 10 GbE with offload over fibre, copper or twin-ax
    • iSCSI offload for 10 GbE
    • 8 Gbps FibreChannel with offload
  • Quad 1 GbE with offload
  • Independent management NIC for console/KVM/maintenance

Since the power is unusual, a couple more pics. Here is the power distribution panel inside the chassis. Each of the white 10-pin connectors can supply over 300 watts. I have five 10-pin to 8+6 pin cables connected. (Note the fat red and black cables coming out of the power supplies (to the right, just outside the image).

​

On the back, four modular hot-swap power supplies (and four of the GPUs).

​

(Yes, there's another system under it - I have three of these configs.)

 

[chriz_r]

OK.

If I have one of these badboys, i'll probably justify keeping it on for heating purposes and maybe get some rendering work done. What kinds of tasks/workloads are you throwing in these, if you don't mind me asking?

 

[theitsage]

@AidenShaw how does the system decide the total number of PCIe lanes and distribution? Are the GPUs all getting PCIe 3 @ x16?

 

[AidenShaw]

If I have one of these badboys, i'll probably justify keeping it on for heating purposes and maybe get some rendering work done. What kinds of tasks/workloads are you throwing in these, if you don't mind me asking?

Training runs for machine learning.

@AidenShaw how does the system decide the total number of PCIe lanes and distribution? Are the GPUs all getting PCIe 3 @ x16?

Since there are four CPUs, there are 160 PCIe lanes available. Four GPUs on x16, rightmost GPU on x8. NVMe switch on x8. One x8 slot open, one x8 and one x16 slot hidden by the two right-hand GPUs. All of the embedded devices (RAID, NICs, etc) use lanes on CPU 1.

​

 

[Flint Ironstag]

Thanks for sharing, that's one helluva box! Next time you're e-wasting one of these, I'll PayPal you $5 + shipping to my place.

Cheers

 

[beaker7]

Also ticked "other."

Running quad Pascal Titans on a Supermicro X10DRG-Q. V-Ray, Maxwell, etc.

 

[pat500000]

(Tim Cook hides behind the corner after looking at those pics)

 

[dlindsey100]

980TI from MVC

 

[seveej]

A 5770 which I bought used (was originally sold as an aftermarket update) and installed in place of an identical card which had bitten the dust.

Wishing I could afford* a professionally flashed or Mac Edition card of something more up-to-date.

RGDS,

* Afford = money and justification.

 

[pat500000]

Is GTX 680 a "plug and play?"

 

[flowrider]

^^^^A Mac Edition is. A PC version is, except no boot screen. Some PC editions can be flashed by the user to show a boot screen.

Lou

 

[pat500000]

^^^^A Mac Edition is. A PC version is, except no boot screen. Some PC editions can be flashed by the user to show a boot screen.

Lou

good to know. Thanks.

 

[pat500000]

^^^^A Mac Edition is. A PC version is, except no boot screen. Some PC editions can be flashed by the user to show a boot screen.

Lou

Lou, is it required to get driver for this? I was checking out ebay and it said CUDA require driver. is that right?

 

[gpzjock]

Sapphire Vapor-X HD5870 1GB (flashed) in a Mac Pro 3,1, Palit Jetstream GTX 770 4GB (unflashed) in a Hack.

 

[kschendel]

Still running the original GT120. Partly because the machine is a software dev machine and I don't care about GPU performance much, partly because it spends most of its time running Linux and I want a card reasonably well supported by the nouveau driver, and partly because I use a Mac Pro for the near-silence (among other things) and this particular GT120 is inaudible.

If I knew of a dead-quiet, 4k boot screen capable, nouveau compatible card, I'd probably upgrade, but it would be pretty far down on the nice-to-have list.

 

[AidenShaw]

Still running the original GT120. Partly because the machine is a software dev machine and I don't care about GPU performance much, partly because it spends most of its time running Linux and I want a card reasonably well supported by the nouveau driver, and partly because I use a Mac Pro for the near-silence (among other things) and this particular GT120 is inaudible.

If I knew of a dead-quiet, 4k boot screen capable, nouveau compatible card, I'd probably upgrade, but it would be pretty far down on the nice-to-have list.

So, your main criterion is that the card is supported by the worst Nvidia driver on the planet?

"Disable nouveau" is the first thing that I do with any Linux system running on hardware with an Nvidia GPU.

 

[slughead]

I had a flashed 7970 but it got so hot the room I'm in would start heating up whenever I even just watched youtube + did web dev. I switched to Gt120 + unflashed gtx980

man, isn't it nice to have PCIe slots, how you doing nMP owners !? :X

 

[flowrider]

Lou, is it required to get driver for this? I was checking out ebay and it said CUDA require driver. is that right?

Nvidia releases 2 drivers for their GPUs. The Web Driver to drive the card in the Mac environment and the CUDA Driver for folks that utilize Cuda. The Web Drivers are specific for each OS point release. The Cuda Driver usually covers a couple of OSs. The latest CUDA Driver is 8.0.51 and is available here:

http://www.nvidia.com/object/macosx-cuda-8.0.51-driver.html

And works with 10.11.x and 10.12.x

And no, the CUDA driver is not needed if you are not using anything that uses CUDA.

Lou