[SOLVED] Max GPUs supported with AMD ryzen 9 3900x

Sep 6, 2020
4
0
10
Hello

I know that AMD ryzen 9 3900x has 24 PCIe v4.0 lanes. On one hand, according to this question, AMD ryzen 9 3900x could handle 2 GPUs. On the other and, in this post it's said that they are sort of equivalent to 40 PCIe v3.0 lanes.

So my doub is: Given that this cpu has the same amount of cores and threads as AMD threadripper 1920x and that it has 40 PCIe v3.0 lanes, How many GPUs could AMD ryzen 9 3900x handle?

Thank you!
 
Solution
...

I'm planning to build a deep learning PC.
...

Thanks!
That's doubtless a use case what will not follow the same rule sets as gaming for hardware optimization and utilization.

In most gaming the CPU is going to tend to be a bottleneck at low resolution (720p or 1080p), when the GPU has an easy time pumping pixels. But at high resolution (4K), therefore a lot of pixels for each frame the CPU prepares, even the most powerful GPU's bog down and leave even mid-range CPU's, like a 3600/X, loafing along. But during all of that PCIe bandwidth has very little impact, except during events like loading levels when a lot of data is pushed across the bus in a short time.

How that relates to 'deep learning' is a bit of mystery...
Hello

I know that AMD ryzen 9 3900x has 24 PCIe v4.0 lanes. On one hand, according to this question, AMD ryzen 9 3900x could handle 2 GPUs. On the other and, in this post it's said that they are sort of equivalent to 40 PCIe v3.0 lanes.

So my doub is: Given that this cpu has the same amount of cores and threads as AMD threadripper 1920x and that it has 40 PCIe v3.0 lanes, How many GPUs could AMD ryzen 9 3900x handle?

Thank you!
As many as you can fit on your motherboard.

Edit add: I don't think you have a good grasp of how PCie works, or maybe it's that you can't explain what you're thinking, but it just doesn't work like what I'm reading. The CPU will use all the lanes the motherboard chipset makes available to EACH PCIe socket at the highest bandwidth the motherboard chipset supports. The way the GPU uses it's lanes will depend on how the GPU works and it's drivers, not the CPU.

If you're asking whether two GPU's (or more, up to X570's limitations) on PCIe gen 4 x 16 sockets will be bottlenecked by a 3900X the answer is, as always: it depends. Which GPU? what resolution? what refresh rate are you pushing? And I'm pretty sure it won't be PCIe interface (gen, lanes whatever) that creates a bottleneck: just the ability of the GPU to push pixels or the CPU to prepare the frames.
 
Last edited:
Sep 6, 2020
4
0
10
As many as you can fit on your motherboard.

Edit add: I don't think you have a good grasp of how PCie works, or maybe it's that you can't explain what you're thinking, but it just doesn't work like what I'm reading. The CPU will use all the lanes the motherboard chipset makes available to EACH PCIe socket at the highest bandwidth the motherboard chipset supports. The way the GPU uses it's lanes will depend on how the GPU works and it's drivers, not the CPU.

If you're asking whether two GPU's (or more, up to X570's limitations) on PCIe gen 4 x 16 sockets will be bottlenecked by a 3900X the answer is, as always: it depends. Which GPU? what resolution? what refresh rate are you pushing? And I'm pretty sure it won't be PCIe interface (gen, lanes whatever) that creates a bottleneck: just the ability of the GPU to push pixels or the CPU to prepare the frames.

Thanks for your inputs... I intentionally didn't want to provide further information to avoid any bias... but as you are pointing out, it will depend. From what I understand from your post, CPU will be able to handle all GPUs connected to the mother board. However, the GPU will be bottlenecked if more than 2 GPUs are added, but would it be that bad? Here's the missing bits of info that could help answering my original question.

I'm planning to build a deep learning PC. Right now i have a Gtx 1080 ti and eventually I will add 2 more GPUs, an Rtx 2080 ti and 3080 when they're available (and not so expensive).... I think the ideal CPU would be an AMD Threadripper, either 1950 (discontinued but still available on ebay) or 2950 (but its price is so high compared to Ryzen 9 3900x). Hence, I was wondering if someone could provide more information on the subject and if AMD Ryzen 9 3900x is not a bad choice for handling 3 GPUs or more

Thanks!
 

Turtle Rig

Prominent
BANNED
Jun 23, 2020
772
104
590
Thanks for your inputs... I intentionally didn't want to provide further information to avoid any bias... but as you are pointing out, it will depend. From what I understand from your post, CPU will be able to handle all GPUs connected to the mother board. However, the GPU will be bottlenecked if more than 2 GPUs are added, but would it be that bad? Here's the missing bits of info that could help answering my original question.

I'm planning to build a deep learning PC. Right now i have a Gtx 1080 ti and eventually I will add 2 more GPUs, an Rtx 2080 ti and 3080 when they're available (and not so expensive).... I think the ideal CPU would be an AMD Threadripper, either 1950 (discontinued but still available on ebay) or 2950 (but its price is so high compared to Ryzen 9 3900x). Hence, I was wondering if someone could provide more information on the subject and if AMD Ryzen 9 3900x is not a bad choice for handling 3 GPUs or more

Thanks!
Like said above it depends on the motherboard is CPU is more then capable enough to drive 3 video cards especially with 24 threads. Of course you can game one video card at a time. You can't dual or three game but if you need 3 video cards in there then invest in a good motherboard that allows for it which would be pretty much the high end models of all vendors. You can put in up to 4 video cards now a days even on a consumer box noot a HEDT box. So have no worries there. If your going to use all 3 at the same time then your going to put a heavy burden on the CPU that is for sure. 🤷‍♂️😲💯👌✝
 
...

I'm planning to build a deep learning PC.
...

Thanks!
That's doubtless a use case what will not follow the same rule sets as gaming for hardware optimization and utilization.

In most gaming the CPU is going to tend to be a bottleneck at low resolution (720p or 1080p), when the GPU has an easy time pumping pixels. But at high resolution (4K), therefore a lot of pixels for each frame the CPU prepares, even the most powerful GPU's bog down and leave even mid-range CPU's, like a 3600/X, loafing along. But during all of that PCIe bandwidth has very little impact, except during events like loading levels when a lot of data is pushed across the bus in a short time.

How that relates to 'deep learning' is a bit of mystery, and problem doesn't very well. It also could depend an awful lot on the particular applications you use.
 
Solution
Sep 6, 2020
4
0
10
That's doubtless a use case what will not follow the same rule sets as gaming for hardware optimization and utilization.

In most gaming the CPU is going to tend to be a bottleneck at low resolution (720p or 1080p), when the GPU has an easy time pumping pixels. But at high resolution (4K), therefore a lot of pixels for each frame the CPU prepares, even the most powerful GPU's bog down and leave even mid-range CPU's, like a 3600/X, loafing along. But during all of that PCIe bandwidth has very little impact, except during events like loading levels when a lot of data is pushed across the bus in a short time.

How that relates to 'deep learning' is a bit of mystery, and problem doesn't very well. It also could depend an awful lot on the particular applications you use.

Thanks again for inputs. I found them very useful.

Regards!