Planning an AI/ML build for school work

[gtech02]

Budget (including currency):  $1500 pre-GPU

Country:  USA

Games, programs or workloads that it will be used for:  Machine learning, mostly in Python

Other details (existing parts lists, whether any peripherals are needed, what you're upgrading from, when you're going to buy, what resolution and refresh rate you want to play at, etc):  I have one WD SN850x 2 TB SSD, one Cooler Master HAF XB EVO, one IEI Mustang-V100-MX8-R10 AI accelerator/ VPU card, and all needed peripherals (mouse, keyboard, monitor, USB sound card).

 

I'm preparing for some AI/ML heavy programing courses.  I plan to build a dual boot (Ubuntu/Windows 11) workstation that might see some medium/light gaming.  My focus is on stability as I've had models take multiple days to train in previous classes, it must churn away without crashing, for hours to days.

 

I plan to order:

Motherboard: Pro WS W680-ACE

Processor: Intel Core i5-13600K 

Windows SSD: Corsair MP600 PRO LPX M.2 2280 2TB (optional)

CPU cooler: Noctua NH-D15

RAM: either Kingston Technology Fury Beast Black 64GB 4800MT/s DDR5 CL38 XMP 3.0 Ready Computer Memory (Kit of 2) KF548C38BBK2-64 or

Kingston Technology Fury Beast Black 64GB 5600MT/s DDR5 CL40 XMP 3.0 Ready Computer Memory (Kit of 2) KF556C40BBK2-64

GPU: TBD 

WIFI: Intel AX210 IEEE 802.11ax Bluetooth 5.2 Tri Band Wi-Fi/Bluetooth Combo, the motherboard doesn't have wifi, but has a M2 Key E

Power Supply: CORSAIR RMx Shift Series RM1200x Shift 

Extra fans: Cooler Master MegaFlow 200, 2x Cooler Master SickleFlow 80 V2  

I'd love to find an IEI Mustang-T100-T5 for my second PCIe 3x4 slot

 

I have two questions, first about RAM, I want 64 GB (2x32GB) both of the options I've listed are on the motherboard's QVL the motherboard lists its memory standard as DDR5-4400, the processor says it supportsDDR5-5600,  However, the ASUS QVL lists some DDR5-6000 kits as compatible, what gives me the most stability for pushing a large dataset through a model without crashing my system.

 

Next the GPU, I plan to buy it separately, as it will probably cost as much as the rest of the system combined.  I'm not sure if I want a pair of workstation grade cards (Nvidia RTX 4000 SFF Ada), I can start with one and add another down the line.  Or I could get one RTX 4090, the Ada Lovelace architecture has some nice features for AI/ML, but do I want one massive 3-4 slot wide card or two 2 slot width cards?

 

 

 

[DrMacintosh]

16 minutes ago, gtech02 said:

I'm preparing for some AI/ML heavy programing courses.

Is this an entirely remote course? You're building a desktop which entirely locks you into one place of work, which is fine if you don't have lectures to attend in person or a professor to have one on one time with. Just a thought.

[gtech02]

1 hour ago, DrMacintosh said:

Is this an entirely remote course? You're building a desktop which entirely locks you into one place of work, which is fine if you don't have lectures to attend in person or a professor to have one on one time with. Just a thought.

100% on-line

[SorryBella]

6 hours ago, gtech02 said:

Motherboard: Pro WS W680-ACE

It wont make you get ECC like on AM4 and AM5. Intel ECC is exclusive to Xeon lineup. And besides, either RAM you picked are just basic non-ECC RAM. Besides that 2 complain, your picks are pretty solid... Except:

6 hours ago, gtech02 said:

However, the ASUS QVL lists some DDR5-6000 kits as compatible, what gives me the most stability for pushing a large dataset through a model without crashing my system.

6000C30 and 6400C34 is stable on Raptor Lake and if you can get those for little extra, its worth it for CPU side performance.

6 hours ago, gtech02 said:

Next the GPU

If its purely just raw value, RTX 3090 and RTX 3090Ti for 700-800$ a pop is a huge boon. 24GB, a lot of tensor cores, and still a very strong architecture. You can get 2 3090 for the price of 1 4090 and it bangs the solo 4090 out the door if your PSU can take it.

[igormp]

15 hours ago, gtech02 said:

Motherboard: Pro WS W680-ACE

Isn't this a bit overkill, specially given that it doesn't give you much extras over other consumer mobos?

15 hours ago, gtech02 said:

RAM: either Kingston Technology Fury Beast Black 64GB 4800MT/s DDR5 CL38 XMP 3.0 Ready Computer Memory (Kit of 2) KF548C38BBK2-64 or

Kingston Technology Fury Beast Black 64GB 5600MT/s DDR5 CL40 XMP 3.0 Ready Computer Memory (Kit of 2) KF556C40BBK2-64

What about 2x48GB instead? Reports say that those run out of the box at 5200~5600MHz without issues, and also gives you more ram to work with.

15 hours ago, gtech02 said:

Next the GPU, I plan to buy it separately, as it will probably cost as much as the rest of the system combined.  I'm not sure if I want a pair of workstation grade cards (Nvidia RTX 4000 SFF Ada), I can start with one and add another down the line.  Or I could get one RTX 4090, the Ada Lovelace architecture has some nice features for AI/ML, but do I want one massive 3-4 slot wide card or two 2 slot width cards?

Basically what SorryClaire said above. Going for workstation grade ones are not worth it for home usage, they'll be slower and cost thrice the price of a regular geforce GPU.

I believe you won't be able to fit a 4090 with your case, so you may want to look into a bigger one, but after doing so and getting a mobo with enough spacing between its PCIe x16 slots there shouldn't be any problem running 2x 3090/4090.

[gtech02]

Well, both the primary SSD and my copy of Windows 11 are delayed in shipping.  However, I snagged a 3090 TI FE on Facebook Market Place, after watching a LTT video about GPU hunting.  I doubt I could get of them in this motherboard, but that's a lot of GPU power.  Now if I could just find a video on cable management.