I'm planning my first PC build around a PNY RTX PRO 6000 Blackwell Workstation Edition with 96 GB of VRAM. The system will run local LLM inference, possibly some fine-tuning, and imaging pipelines using CPU-heavy OpenCV/NumPy preprocessing alongside PyTorch workloads under Ubuntu 24.04 LTS.
My current parts list includes an AMD Threadripper PRO 9975WX, ASUS Pro WS WRX90E-SAGE SE motherboard, 128 GB of DDR5-5600 ECC RDIMM, two Samsung 990 PRO drives, a 2,000 W ATX 3.1 power supply, a Fractal Design North XL case, and a 360 mm sTR5 AIO cooler. The build costs roughly €10,000 before the GPU.
I'd like to expand to at least 256 GB of RAM, add more NVMe storage and a large HDD, and possibly install a second RTX PRO 6000 later. My current memory plan is to use four 32 GB RDIMMs now and add four more later to populate all eight memory channels.
Would it make more sense to save about €2,300 by switching to an ASUS Pro WS TRX50-SAGE WiFi with a Threadripper 7970X? The TRX50 platform has only four DIMM slots, but I'm not sure how much the extra memory channels and upgrade capacity would matter for my workloads. I'd also appreciate advice on whether the RAM expansion plan, cooling setup, case airflow, power supply, and other parts are sensible or excessive.
3 Answers
The power and cooling choices are defensible for a potential two-GPU build, but they are overkill for the current single-GPU system. A 2,000 W supply gives useful headroom for a second 600 W card, though you should verify its connector layout, transient handling, and the actual power requirements of both cards.
A 360 mm top exhaust radiator should work if the case supports the socket hardware and radiator clearance. Keep strong front intake airflow for the GPU and add at least one rear exhaust fan. A GPU support bracket is worthwhile for a card this heavy. I’d also budget for a UPS, check Linux support for the board’s networking and sensors, and confirm that the case can physically accommodate the motherboard, radiator, GPU length, power cables, and the unusually large PSU.
With four DIMMs on WRX90, you’re using four of its eight memory channels, so the initial memory bandwidth is broadly comparable to a four-slot TRX50 system despite the much higher cost. The extra channels become more valuable once all eight slots are populated, especially for CPU-heavy preprocessing or workloads that spill model data into system RAM.
Before relying on a later upgrade to eight sticks at DDR5-5600, confirm with the memory vendor or motherboard support list that mixing kits purchased at different times is supported at that speed. If your models and fine-tuning data mostly fit inside the GPU’s 96 GB of VRAM, TRX50 may provide similar performance for much less money. WRX90 is easier to justify if you know you’ll need 256 GB or more and want maximum expansion flexibility.
I’d seriously consider installing 256 GB from the beginning if the budget allows, rather than assuming a second kit will train perfectly later. Four sticks is a reasonable starting point, but it leaves half of WRX90’s bandwidth unused. Also remember that filling all eight slots can reduce the maximum stable memory speed depending on the DIMMs and firmware.
For a single GPU, a less expensive platform may be enough. The 9975WX and WRX90 combination makes more sense when you expect sustained CPU preprocessing, large datasets, several storage devices, or a second GPU—not simply because the parts are faster on paper.
Related Questions
Lenovo Thinkpad Stuck In Update Loop Install FilterDriverU2_Reload