4 ms·

I'm looking at building a PC for deep learning, and I'm just learning about the importance of PCIe lanes for GPUs. I'm confused by the Coffee Lake numbers - 16

I'm looking at building a PC for deep learning, and I'm just learning about the importance of PCIe lanes for GPUs. I'm confused by the Coffee Lake numbers - 16 on the CPU, 24 on the motherboard? Does this mean that you could have 2 GPUs are without slowdown, or does the 'x16 per GPU' advice only take into consideration the CPU lanes? reference: https://twitter.com/karpathy/status/869289384687714304 https://twitter.com/karpathy/status/869289384687714304

dirtyaura 9y ago
I'm on the same boat as you, building my first deep learning box as we speak. What I've understood is the following: if you want to build, or grow your build over time to multi-GPU deep learning machine, you are better off server-targeted Xeon family processors instead of this consumer market oriented Core line. Xeon processors typically support 32-40 PCIe lanes, while Core processors support 16. I read that with typical deep learning loads, utilizing 8 PCIe lanes per GPU is good enough, so with 32-40, you can go up to 4 GPUs. However, instead of spending tons of money and building a perfect basis for a hypothetical future DL box from the get-go, I decided to just start with cheaper parts and go with single GTX 1080 and Core processor. When I hit the limits of it, I probably know more what I need and likely the technology has advanced too.
- AlphaSite 9y ago
  Go for an AMD chip, you get 60-128 general purpose Pcie lanes.
  - dragontamer 9y ago
    Well, Threadripper AMD anyway. The key is if PCIe lanes are important, you need to buy a Motherboard that supports a high number of PCIe lanes. AMD's Threadripper is cheap (for a workstation) and supports 64+ PCIe Lanes. Intel's "extreme" line (both chips and motherboard) are the high PCIe count workhorses, as well as their Xeon chips. But they are much more expensive than AMD's solution. Even then, an Intel 6950X "only" has 40-PCIe lanes. I wonder if Intel's new line of "Extreme" will up the PCIe count to better compete vs AMD.
  - dirtyaura 9y ago
    I understood that for a hobbyist, who doesn't hand-optimizate preprocessing, good single thread performance is better with numpy. From that front Intel is better.
drudru11 9y ago
You should probably get a threadripper and max out with whatever GPU/ML processor boards you need. Will you be running Linux?

dirtyaura 9y ago

I'm on the same boat as you, building my first deep learning box as we speak. What I've understood is the following: if you want to build, or grow your build over time to multi-GPU deep learning machine, you are better off server-targeted Xeon family processors instead of this consumer market oriented Core line. Xeon processors typically support 32-40 PCIe lanes, while Core processors support 16. I read that with typical deep learning loads, utilizing 8 PCIe lanes per GPU is good enough, so with 32-40, you can go up to 4 GPUs. However, instead of spending tons of money and building a perfect basis for a hypothetical future DL box from the get-go, I decided to just start with cheaper parts and go with single GTX 1080 and Core processor. When I hit the limits of it, I probably know more what I need and likely the technology has advanced too.

AlphaSite 9y ago
Go for an AMD chip, you get 60-128 general purpose Pcie lanes.
- dragontamer 9y ago
  Well, Threadripper AMD anyway. The key is if PCIe lanes are important, you need to buy a Motherboard that supports a high number of PCIe lanes. AMD's Threadripper is cheap (for a workstation) and supports 64+ PCIe Lanes. Intel's "extreme" line (both chips and motherboard) are the high PCIe count workhorses, as well as their Xeon chips. But they are much more expensive than AMD's solution. Even then, an Intel 6950X "only" has 40-PCIe lanes. I wonder if Intel's new line of "Extreme" will up the PCIe count to better compete vs AMD.
- dirtyaura 9y ago
  I understood that for a hobbyist, who doesn't hand-optimizate preprocessing, good single thread performance is better with numpy. From that front Intel is better.

drudru11 9y ago

You should probably get a threadripper and max out with whatever GPU/ML processor boards you need. Will you be running Linux?