ThinkStation P620 (3945WX) + RTX 5070 Ti vs Ryzen 9 7900X Custom Build – Which Would You Pick for AI/ML?
Posted by Reasonable-Bear-9788@reddit | LocalLLaMA | View on Reddit | 6 comments
I’m deciding between two builds for mostly AI/ML (local LLMs, training/inference, dev work) and some general workstation use.
Option A – ThinkStation P620 (used, 1yr Premier onsite warranty) – ~1890 CHF total
- Threadripper PRO 3945WX (12c/24t)
- 128GB ECC DDR4 (8-channel)
- 1TB NVMe
- 1000W PSU
- 10GbE
- Added RTX 5070 Ti 16GB (850 CHF, bought and installed separately)
Option B – Custom build – ~2650 CHF total
- Ryzen 9 7900X (12c/24t) - used
- 64GB DDR5 5600
- Gigabyte X870E AORUS Elite WIFI7 ICE- used
- 2TB Samsung 990 EVO
- 1000W RM1000x
- RTX 5070 Ti 16GB
GPU is the same in both.
Main differences:
- 128GB RAM + workstation platform vs newer Zen 4 CPU + DDR5
- \~750 CHF price difference
- ThinkStation has 10GbE and more PCIe lanes
- Custom build has better single-core + future AM5 upgrade path
For mostly GPU-based ML workloads, is the newer 7900X worth the extra \~750 CHF? Or is the 128GB workstation platform better value?
Would appreciate thoughts from people running similar setups.
RoughOccasion9636@reddit
For pure GPU inference where the model fits in VRAM, both perform identically. The 5070 Ti is the bottleneck on either platform.
The real differentiator shows when you run 70B+ at Q4_K_M (~40GB), which overflows 16GB VRAM. The P620's 8-channel DDR4 gives ~200GB/s bandwidth for layer offloading vs the 7900X's ~90GB/s on dual-channel DDR5. Plus 128GB vs 64GB means the 7900X hits the ceiling faster on larger quantized models.
Paying 750 CHF more for Option B to get half the RAM and lower offload bandwidth doesn't make sense for LLM workloads. Unless you specifically need AM5's upgrade path or plan to run CPU-heavy preprocessing, the P620 is the clear pick.
mon_key_house@reddit
Are you sure about the LRDIMM RAM OK for the P620? I'm considering moving from my P520 and want to reuse my sticks.
RoughOccasion9636@reddit
On your GPU expansion question for the P620: it has a dual-slot riser configuration and typically supports 2x full-length GPUs. The specific slot layout depends on which riser card came with yours, but the 1000W PSU is the real bottleneck since a second 5070 Ti would push you close to 800W under load. Worth checking the exact P620 configuration guide for your revision on Lenovo support.
A couple things others have not mentioned:
The 3945WX has 128 PCIe 4.0 lanes from the Threadripper PRO platform vs the 7900X which is limited to 24 usable lanes. If you ever add NVMe drives or a second GPU later, you will not be chasing lane allocation on the P620.
ECC RAM matters more than most people admit for long-running inference jobs. A bit flip mid-computation on a 2-hour fine-tuning run is silent and produces garbage output. On consumer platforms you only find out when the results look wrong. The P620 gives you hardware error detection and correction throughout.
For your use case, Option A is the clearer choice unless you specifically need AM5 for future CPU upgrades. The Threadripper PRO platform ages better for this workload.
Reasonable-Bear-9788@reddit (OP)
Great thanks! I already ordered it now, thank you so much for the help!
I can never understand why people downvote such posts, but the answers are super helpful!
FlanOk7029@reddit
personally i'd go with option a, the extra 64gb ram is huge for running larger models locally and workstation platform gives you more headroom for expansion in future
Reasonable-Bear-9788@reddit (OP)
Thank you so much for the feedback! I am a bit worried about the lenovo case, will it have space to add more GPUs? I am not super experienced in this, and not able to get a clear answer online.