SSD : 1TB Samsung 980 PRO SSD with Heatsink £75
PSU: CORSAIR RM850e (2025) Fully Modular Low-Noise ATX Power Supply £ 105
CPU COOLER: ARCTIC Liquid Freezer III Pro 360 - AIO CPU Cooler £72
CASE : Montech AIR 903 Base Midi-Tower, Tempered Glass - Schwarz £46
TOTAL : £1,593
What you guys think?
I am only gonna be working on medium size datasets . I know I can’t build the next ChatGpt with these specs lol
Only my 2 Cents but if you are going to work on AI or machine learning, maybe try to get a GPU with 24GB of VRAM like a RTX 3090, even used. You can maybe have a read at the following thread I link where we basically let large models like Deepseek run on modest hardware completely locally. But a GPU with 16GB VRAM is very limiting, 24GB is better.
This related to letting models run on modest hardware, I am not sure how important the VRAM is for learning how to train a model and things like that, but my guess would be that more VRAM is better too.
Thanks for the reply . I was thinking that 24 GB vram would be better . But the 3090 draws huge power around 450-550w and the 5070 only draws like 300-350 W.
I would also need to buy a 1000 W PSU instead of the 850 W.
The only reason I’m buying 5000 series is because it got the new architecture and is more efficient and 16 GB vram ain’t much but I don’t know …
I thought the 3090 also does have a TDP of 350 Watts, but to be honest I do not know how much these two cards actually pull. Other then that your build looks fine to me. But I do not have much experience with consumer hardware these days.
Maybe we can ping @MisteryAngel, she always has good input and should be up to date on the current consumer stuff.
I agree with @H-i-v-e that 24GB does give some benefits over 16GB even on an older architechture GPU.
First, I have a used 3090TI FE 24GB VRAM with a default power cap of 450W. However, in practice, for both gaming and LLM stuff I typically run it at 350 by simply running nvidia-smi -pl 350 in either windows or Linux. You can set the cap for whatever you are doing no problemo. Even with 100W less the performance is about 90% of full power cap on many applications.
Second, I did benchmark of my 3090 vs a 5080 in another thread here:
So for both LLMs and image generation and gaming personally I’d still pick the 3090TI. The only downside is that older sm86 arch does not support native fp8e4m3 dtypes, but in practice it hasn’t been an issue as triton kernels or GGUFs work plenty fine.
Personally I’d recommend getting a minimum of 2x64GB DDR5 RAM sticks. If you want to roll the dice get the 4x64GB DDR5 and hope you win the silicon lottery and get DDR5-6000MT/s. Having 128 or even 256 GB on a AM5 system (likely would want a newer mobo than you mention) would allow you to run the big MoEs like DeepSeek full size, GLM-4.5 and others with ik_llama.cpp. Given the active weights of a large MoE are farly small in comparison with overall size, having a bunch of RAM to hold all the routed experts allows you to get usable ~5ish+ tok/sec TG and likely 100+ tok/sec PP with only about 70GB/s memory bandwidth.
Final two points to consider would be maybe a Gen5 NVMe which helps reduce model load times e.g. T700 Crucial 2TB (or 4TB - the 1TB is slower). And finally consider a Zen5 CPU to get the avx_vnni real 512bit instruction set which can improve PP by almost double for some quants/configuration.
All this would cost a little more, but if you can save by finding a used 3090TI it might be about even perhaps depending on the cost of the 5070 you’re looking at. The next jump up is much more expensive anyway, so you’d be getting the most out of the least cost.
Also, they have masters in ai now? wild times! Have fun and enjoy your journey!
Depends on the model. Buying a new GPU under 16 GB doesn’t make much sense. Conversely, buying more than 16 GB is stupid expensive outside of used 3090s and 7900s. So, realistically, 16 GB is pretty much what’s happening.
Seems like the R9700 might improve the overall situation but it looks set to cost the entire build budget here and wants a 1000+ W supply, so it’s not any help.
Depends on workload, as always, but the standard basic for power supply sizing and thermal estimation’s the maximum draw the GPU’ll hit during the build’s lifetime. Which is often taken to be measured draw at stock but goes higher if the power limit’s expected to be lifted.
There’s sometimes cards which leave the power limit at manufacturer TDP but it’s commonly lifted and and it’s also common cards overrun their power limits. Default 3090 power limits I’m aware of are in the 370-420 W range with 30-80 W overruns, so 400-450 W actual draw is common. TechPowerUp measured the FTW3 hitting 507 W.
For 5070 Tis measured draws are in the 300-325 W range, much closer to the 300 W TDP. Seems like the supers might add ~50 W to that but no data yet.
3090 [Ti] is also closer to Nvidia end of lifeing it to make you to buy another GPU. Not that AMD’s much different, though 7900 is a gen newer. 7900 XT’s the most I’d put on an 850 W supply, so presumably 3090 [Ti] and 7900 XTX aren’t happening here.
9700X is Zen 5.
I don’t know of any actual data but hints exist 600 series is possibly better than 800 series for 2DPC 2R. Big MoEs presumably fall under the uni cloud credits anyways.
Other point I’d make here is the desirability of planning for mains power availability and cooling.
The III Pro’s overkill even if you’re going to lift the 9700X to 142 W PPT. Depending on distribution and shipping you might be able to free up a bit of budget with a Phantom Spirit 120, which I’d probably put towards RM850x.
Recent Samsung drives are kind of jank with small pSLCs and difficulties in sustained throughput, though most IO isn’t intensive enough to matter. I’d avoid NGFFF with stock heatsinks as they’re pretty much all minimal solutions that’ll underperform mobo armor if you take the time to check and adjust to good thermal contact. I’d also avoid Asus for all the problems, especially as ROG is where their brand tax is highest.
It always varies with local pricing and availability FWIW mostly I’ve been buying SN850X and Steel Legend lately. Teucer M.2-10 and Thermalright HR-09 Pro are a couple good default heatsinks if you get into IO armor can’t handle, though that’s unlikely in general and additionally improbable with only 1 TB in most use cases. A large model hammering the drive with coefficient reads is a potential exception, though.
Mind also 5070 Ti’s getting into elevated melt risk. This is a know your workloads thing, which seems difficult at this point, but if you’re looking mainly at deep learning with modest datasets and model complexity my guess would be 9700X + 2x32 + 5070 Ti’s probably a good tradeoff at the present budget. The machine learning I’ve been doing mostly isn’t GPU accelerated and 2x48 is minimal, so less GPU, 9900X or 9950X, and potentially 2x64 makes more sense.
There’s also a split with AMD being better supported on Linux and Nvidia on Windows. I don’t have a good sense of ROCm 6 via WSL2 on Windows and it’s too soon to tell how much ROCm 7 might simplify AMD workloads on Windows. Depends what you’re doing with local models too but I would consider 9060 XT 16 GB and 9070 of interest here. Maybe 5060 Ti 16 GB as well.