Question Which cpu to buy? (deep learning and streaming)

[rakgb]

I'm planning to build a deep learning machine that does a lot of post processing on images on the cpu too. Currently Core i9 9900K turns out to be the fastest cpu available on market but with only 16 PCIE Lanes. My concern is I will be needing 2 GPU (RTX 2080Tis) and a 4K video capture card (Black Magic Decklink SDI 4K - link). The capture card will use upto 4 threads max, 2 for capturing raw camera feed and 2 for sending back processed feed and both my GPUs will be running parallel doing different tasks. According to this Guide the number of lanes do not really matter for deep learning performance, his study shows 32 images from ram to gpu at a time which takes 2ms if GPU is on 16x lane, 5ms if its on 8x lane and 9ms if its on 4x lane. In my case since it will be used only for inferencing it will just be 1 images at a time from ram to the two GPUs. My only concern is I will be running my entire pipeline in realtime from capturing, processing and writing back to the capture card and all of this should happen under 40-50ms per frame. The faster the cpu the better my cpu part of code runs. Coming to the confusing part, Core i9 9900K is a brilliant choice but due to the capture card and two GPUs I'm guessing there's a bottleneck somewhere! I really do not want to get the X series Core i9 9900X which has a whopping 44 lanes since its clock speed maxes out at 4.5GHz. Do I have no choice but to get the X chip to avoid the bottlenecks?

 

[InvalidError]

Does the extra 500MHz boost (if you aren't OCing) really matter that much? Can't you parallelize your CPU-side code some more to use the 9900X's extra cores and threads? Further along this train of thoughts, how about a 16C32T ThreadRipper 2 (2950X) which has up to 60 available PCIe 3.0 lanes?

 

[captaincharisma]

why are you looking at consumer grade CPU's for something like this? you need to be looking at AMD's threadripper or Intel's xenon CPU's for something like this

 

[rakgb]

Does the extra 500MHz boost (if you aren't OCing) really matter that much? Can't you parallelize your CPU-side code some more to use the 9900X's extra cores and threads? Further along this train of thoughts, how about a 16C32T ThreadRipper 2 (2950X) which has up to 60 available PCIe 3.0 lanes?

I don’t really need 60 lanes do I? That 500MHz is making my code run 5ms faster (and nope I haven’t overclocked). I do not think I can parallelize it more. And I’ve never tested my setups on an amd machine I don’t think I have the budget to gamble with amd for now.

 

[rakgb]

why are you looking at consumer grade CPU's for something like this? you need to be looking at AMD's threadripper or Intel's xenon CPU's for something like this

I’ve tested my setup on xeons and my conclusion was xeons are overpriced slower chips. The code sadly is little sequential and cannot be parallelized more so it is highly dependant on clock speed. AMD threadripper too offers about 4.4GHz which seems to be slightly slower than 9900X.