First Build Ever! AI/ML/Computer Vision Focus

[Nate9]

Budget (including currency): 2000-3000 (USD)

Country: United States

Games, programs or workloads that it will be used for: Geospatial Image Processing, Computer Vision, Machine Learning, General Software Development

Other details: https://newegg.io/5ce0910 (replace MSI Ventus GeForce RTX 3080 with Founders Edition), allowing space for future upgrades

 

Hi all!

 

TLDR: computer vision with really big images.

 

This is my first PC build ever and would love any and all advice you all might have for the current build I have attached. 

Some background, I'm a former Computer Science PhD student (dropped out after 3 years due to overwhelming stress at the time), but I've come to miss the research and meaningful work I was doing. I'd like to build a PC to get back into computer vision research and start rebuilding a portfolio of projects in that area, particularly with geospatial/satellite/remote sensing imagery. I mention that because these images can be HUGE -- often upwards of 400x400 pixels and sometimes with more than 3 channels.

 

So for me, while I'd like to do other software development in addition and maybe just a bit of gaming, the focus is certainly the GPU. I'm curious to see if you all might recommend downgrading certain parts of the build in order to allow for better cooling or even a better/more than one GPU. I know I may be overspending in certain areas unnecessarily, so please let me know what you think! I'm excited to learn more about the hardware side of things.

 

[DD08]

Hi Nate

What's size files are you working with, or intend to work with. I'm asking because it relates to being focused primarily on computing power(e.g. CUDA cores) or on RAM.
Are the programs using specific code/devices(again... CUDA-cores/NVML) to the processing? Is ECC RAM relevant for you?
Will the software allow "SLI"/NVLINK/SPLIT COMPUTE ?
 

I'm running systems relying on CUDA-cores, so more or less a high-end gamer PC.

/DD08

[igormp]

11 hours ago, Nate9 said:

Other details: https://newegg.io/5ce0910 (replace MSI Ventus GeForce RTX 3080 with Founders Edition), allowing space for future upgrades

The only change I'd do is to go with 2x32gb instead of 2x16gb, and also go for a 12gb 3080 instead of the 10gb version (but that's up to your models and batch sizes).

[brob]

I would suggest a better performing cpu cooler. The Dark Rock Pro 4 comes to mind. 12th gen Intel CPU performance can be improved significantly the cooler it is able to run.

 

Consider a 2x32GB DDR4-3200 CL16 memory kit. Corsair Vengeance LPX would be a good pairing with the Dark Rock cooler.

 

Consider a less expensive case with excellent airflow. Something like the Bequiet Pure Base 500DX perhaps.

 

An 850W RMx PSU should be sufficient for the build.

 

I'd be tempted to start with a 2TB NVMe drive and no HDD.

[Nate9]

Thank you for your responses!

@DD08

• I think it's safe to assume that the absolute largest images I'd be working with would be around 15-20MB. I can imagine I'll probably downsample or split those into multiple images before processing
• Programs will definitely be using CUDA-cores when processing the images -- lots of matrix computation and linear algebra
• The software will allow SLI/NVLINK, but from what I've read it doesn't always make a huge difference for machine learning and image processing. Not sure I can explain that too well, but from what I understand, it's more about parallelizing computation -- rendering is not as much the focus

@igormp good catch on those recommendations, thank you!

 

@brob really appreciate the specific suggestions. I was especially nervous about cooling/airflow.

[DD08]

I would grab a Lian Li Air Mini case + some Arctic p14 fans, and probably the Arctic Liquid freezer 2 280 in push-pull... or a Fractal Torrent(Superb cases - Torrent for performance, Lian Li for quality/aesthetics). Phanteks P500A, P600S and Enthoo Pro 2(BIG) are also nice cases. Arctic Liquid freezer 2 420 is a beast cooler, but also BIG.
Switch to a DDR5 setup(e.g. 32GB Kingston 4800)... yes, it is more expensive(roughly 150), but i reckon it's a less "painful" investment, these days.
The MoBo is a good "no nonsense" board, which has also been pointed out by BuildZoid. Just take the DDR5 version of the same.
PSU... i would choose the Be quiet Straight Power 1000, or up it to the 1200W or Corsair HX1200(they have both served me well).
Clearly go for 2 TB Nvme, if you can/will.
The Ventus 3080 looks to be the best bang for the buck, but I have no experience with those... 
Well... that was just my thoughts...
Good luck!

Cheers, DD

[igormp]

14 minutes ago, DD08 said:

Switch to a DDR5 setup(e.g. 32GB Kingston 4800)... yes, it is more expensive(roughly 150), but i reckon it's a less "painful" investment, these days.

2x32gb ddr5 kits are hella expensive and won't bring much benefit for OP's use case.

[DD08]

7 hours ago, igormp said:

2x32gb ddr5 kits are hella expensive and won't bring much benefit for OP's use case.

Maybe not that much, but the primary reason was to adopt to the "new platform". OP would anyway have to decide, whether it is worth the 150 extra, instead of running a "hybrid" system.

[igormp]

6 minutes ago, DD08 said:

Maybe not that much, but the primary reason was to adopt to the "new platform". OP would anyway have to decide, whether it is worth the 150 extra, instead of running a "hybrid" system.

If OP had an unlimited budget that would be fine, but those extra 150 bucks would be better spent on a GPU upgrade IMO, or trying to get 128gb of ddr4, or even getting a bigger NVMe so the GPU won't get data starved.

[Admiral Thrawn]

Here is my recommended build. It is just about the same as yours but with a focus on rendering performance and data transfer speeds.

'

[PCPartPicker Part List](https://pcpartpicker.com/list/WWjLcb)

Type|Item|Price
:----|:----|:----
**CPU** | [Intel Core i7-12700K 3.6 GHz 12-Core Processor](https://pcpartpicker.com/product/ZbddnQ/intel-core-i7-12700k-36-ghz-8-core-processor-bx8071512700k) | $376.98 @ Amazon 
**CPU Cooler** | [Corsair iCUE H100i RGB PRO XT 75 CFM Liquid CPU Cooler](https://pcpartpicker.com/product/B6pmP6/corsair-icue-h100i-rgb-pro-xt-75-cfm-liquid-cpu-cooler-cw-9060043-ww) | $112.82 @ Newegg 
**Motherboard** | [Gigabyte Z690 AERO G ATX LGA1700 Motherboard](https://pcpartpicker.com/product/P6PQzy/gigabyte-z690-aero-g-atx-lga1700-motherboard-z690-aero-g) | $302.99 @ Newegg 
**Memory** | [G.Skill Trident Z5 RGB 64 GB (2 x 32 GB) DDR5-6000 CL30 Memory](https://pcpartpicker.com/product/DtFbt6/gskill-trident-z5-rgb-64-gb-2-x-32-gb-ddr5-6000-cl30-memory-f5-6000j3040g32gx2-tz5rk) |-
**Storage** | [Samsung 870 QVO 2 TB 2.5" Solid State Drive](https://pcpartpicker.com/product/R7FKHx/samsung-870-qvo-2-tb-25-solid-state-drive-mz-77q2t0bam) | $179.99 @ Amazon 
**Storage** | [Samsung 980 Pro 2 TB M.2-2280 NVME Solid State Drive](https://pcpartpicker.com/product/f3cRsY/samsung-980-pro-2-tb-m2-2280-nvme-solid-state-drive-mz-v8p2t0bam) | $268.00 @ Amazon 
**Video Card** | [NVIDIA GeForce RTX 3080 10GB 10 GB Founders Edition Video Card](https://pcpartpicker.com/product/RnDkcf/nvidia-geforce-rtx-3080-10-gb-founders-edition-video-card-9001g1332530000) |-
**Case** | [Lian Li Lancool II Mesh ATX Mid Tower Case](https://pcpartpicker.com/product/d82bt6/lian-li-lancool-ii-mesh-atx-mid-tower-case-lancool-ii-mesh-rgb-black) | $188.57 @ Amazon 
**Power Supply** | [EVGA SuperNOVA GA 850 W 80+ Gold Certified Fully Modular ATX Power Supply](https://pcpartpicker.com/product/j6NgXL/evga-supernova-ga-850-w-80-gold-certified-fully-modular-atx-power-supply-220-ga-0850-x1) | $89.99 @ EVGA 
 | *Prices include shipping, taxes, rebates, and discounts* |
 | **Total** | **$1519.34**
 | Generated by [PCPartPicker](https://pcpartpicker.com) 2022-05-19 15:22 EDT-0400 |

 

You also have an upgrade path here and are still getting that insane 980pro speed. 

 

Hope this was helpful

- C.M

[DD08]

21 hours ago, igormp said:

If OP had an unlimited budget that would be fine, but those extra 150 bucks would be better spent on a GPU upgrade IMO, or trying to get 128gb of ddr4, or even getting a bigger NVMe so the GPU won't get data starved.

Whatever rocks the boat. The currently selected GPU would need additionally 250 bucks to upgrade to the cheapest 3080ti.
The DDR5 "route" would yield a presumably less expensive "upgrade-path", depending on resale prices and cost on new MoBo and RAM.
I don't know whether 128GB of RAM would be needed, when the GPU only has 10-12GB.

@ Admiral Thrawn: The PSU seems a bit too tiny and the 64GB RAM is 750 bucks.

[igormp]

3 hours ago, DD08 said:

I don't know whether 128GB of RAM would be needed, when the GPU only has 10-12GB.

Oh boy, trusted me, it will be needed. Most of the time spent when dealing with data is trying to curate it before you do any work with the GPU. 

And, in case OP is working with pandas for that, a rule of thumb is that you need 5~10x more ram than the size of the dataset you're working with (so a 10gb dataset, which is really small, would need 50~100gb to work with). 

[Nate9]

Not to feed the flame of the DDR4 vs. DDR5 debate currently going on, but I've updated and made two part lists because I'm indecisive. 

DDR4: https://pcpartpicker.com/user/napool/saved/#view=VnNYdC

DDR5: https://pcpartpicker.com/user/napool/saved/#view=Kcn9CJ

 

Really appreciate everyone's suggestions and the discussions thus far. I'm not sure I'm 100% convinced on the DDR5 adoption, as I'm inclined to put that money toward a future GPU addition or RAM as @igormp suggested...or just save it lol. It is tempting to adopt the new config though. My current questions are: does the cooling + airflow + power supply config seem ok for the CPU + GPU combo? I know that GPU can get quite hot. Am I over/under-spending anywhere in that area? I had someone say the build might be a little overkill, though I think he's mostly referring to the GPU -- which unfortunately seems like basically a non-negotiable for AI/ML with big images.

 

@DD08, ironically enough I kinda like the Fractal Torrent's aesthetic the most anyway lol.

 

Side note: @igormp, I hadn't heard that memory rule of thumb before! That's really interesting, where did you hear that?

 

EDIT: oh also, I can't really seem to convince myself to go for water cooling... idk it just seems like more hassle/risk than it's worth especially for a first-time PC builder. Let me know if you all have any thoughts on that.

[brob]

2 hours ago, Nate9 said:

Not to feed the flame of the DDR4 vs. DDR5 debate currently going on, but I've updated and made two part lists because I'm indecisive. 

DDR4: https://pcpartpicker.com/user/napool/saved/#view=VnNYdC

DDR5: https://pcpartpicker.com/user/napool/saved/#view=Kcn9CJ

 

 

Is there a reason why you chose not to use the WiFi versions of the Pro Z690-A?

 

[igormp]

2 hours ago, Nate9 said:

Side note: @igormp, I hadn't heard that memory rule of thumb before! That's really interesting, where did you hear that?

 

There you go: https://wesmckinney.com/blog/apache-arrow-pandas-internals/

Quote

Nowadays, my rule of thumb for pandas is that you should have 5 to 10 times as much RAM as the size of your dataset. So if you have a 10 GB dataset, you should really have about 64, preferably 128 GB of RAM if you want to avoid memory management problems. This comes as a shock to users who expect to be able to analyze datasets that are within a factor of 2 or 3 the size of their computer's RAM.

 

2 hours ago, Nate9 said:

EDIT: oh also, I can't really seem to convince myself to go for water cooling... idk it just seems like more hassle/risk than it's worth especially for a first-time PC builder. Let me know if you all have any thoughts on that.

I wouldn't, air is simpler, and unless you're doing OCing a good air cooler should handle that i7.

 

 

[Nate9]

20 hours ago, brob said:

 

Is there a reason why you chose not to use the WiFi versions of the Pro Z690-A?

 

Honestly, no particular reason. Do you suggest the WiFi version?

[brob]

5 hours ago, Nate9 said:

Honestly, no particular reason. Do you suggest the WiFi version?

 

Yes, when they are less expensive than an add-on card, which tends to be the case these days.

[DD08]

On 5/21/2022 at 12:08 AM, Nate9 said:

EDIT: oh also, I can't really seem to convince myself to go for water cooling... idk it just seems like more hassle/risk than it's worth especially for a first-time PC builder. Let me know if you all have any thoughts on that.

I've been running water cooling for many years now, and haven't had any problems, yet (knock on wood). I don't find it to be a hassle... to be honest, the big air coolers (e.g. Noctua og BQ) are more fiddly to integrate, than the AIO's(for me at least).

The BQ cooler you've selected should go well with the Torrent case. I fairly recently completed 3 systems in Torrent cases, but with an Arctic 420 (Liquid Freezer 2) AIO in Push/Pull configuration. Had to modify the bottom fan mount to enable installation of the 420 radiator and fans. To most, the use of an AIO in the Torrent is rather ridiculous, but the system now has a higher airflow and static pressure, than the default 2 x 180mm fans(+ 3 x 140mm). Temps are low, even though the demand for cooling a 12900K and 2 x 3090 Strix , could be "interesting".

But, again, the BQ or Noctua aircoolers will fit well into a Torrent build.

 

If you can fork out the extra cash, the 980 Pro NVMe will provide potentially double the speed.

Regarding PSU, I would still recommend a good +1000watt unit, if the price is not astronomical. I just upgraded my sons PSU, as his 850watt PSU did not have enough PCI-e cable outlets(2) for his new GPU(3). He now runs a BQ Straight Power 1200w unit, so should be g2g for the next few years, if the "next-gen-nvidia-recommended-wattage" inflation takes off into the multiverse.
 

I usually go by the "RAM-scale-factor" of 3-4 times the GPU, but if the tasks at hand dictate 5-10, then... that is the way.

... I have spoken. : )