Rasheed Sons AI Workstation — Threadripper PRO 9975WX, Dual RTX PRO 4000
The bigger brother of our 9965WX workstation. Thirty-two Zen 5 cores, 128GB of quad-channel ECC memory and two NVIDIA RTX PRO 4000 Blackwell cards for 48GB of ECC GDDR7 in total. It is assembled, cable-managed and tested in Johannesburg. Run and fine-tune AI models on your own hardware, with no cloud subscription, no per-token billing, and nothing leaving your premises.
What the second card changes
Video memory decides which models run at full speed. One RTX PRO 4000 gives you 24GB. Two give you 48GB, and inference tools like llama.cpp, vLLM and Ollama split a model across both cards. Models that had to spill into system memory on a single card now sit entirely on the GPUs.
-
7B to 14B models (Llama, Mistral, Qwen, Gemma class) run on a single card at full speed, which leaves the second card free for other work.
-
32B models run with plenty of headroom for long context windows, which is what you need for RAG over large document sets and long coding sessions.
-
70B models, quantised, fit across both cards in video memory. That is the step from "usable" to genuinely quick, and the main reason to choose this machine over the single-card version.
-
Very large mixture-of-experts models can still use the 128GB of quad-channel system memory, at roughly 205 GB/s, alongside the 48GB on the cards.
Two cards also means two jobs at once. Serve a chat model to your team on one GPU while the other generates images, transcribes audio or runs a fine-tune, and neither job slows the other.
Beyond text: Stable Diffusion and Flux image generation, Whisper transcription, local embeddings and vector search, and LoRA fine-tuning without renting an H100 by the hour.
Thirty-two cores behind it
The Threadripper PRO 9975WX adds eight cores over the 9965WX. That matters for everything around the model: preparing and cleaning datasets, tokenising, building vector indexes, running CPU-offloaded layers, and compiling. It also makes this a serious machine for rendering, simulation and heavy multitasking when it isn't running AI work.
Storage that keeps up
4TB of PCIe Gen 5 NVMe. Model weights are large and you will collect them: a single quantised 70B is around 40GB before you download a second one. Gen 5 sequential speed means a model loads in seconds rather than minutes, which matters when you switch between models during a working day.
Built for multiple cards
The RTX PRO 4000 is a single-slot blower card, and that's why this works. Each card exhausts its heat out the back of the case, so two stacked cards stay stable under sustained load where open-fan consumer cards choke on each other's hot air. The 1600W power supply has ample capacity for both, and the board's remaining PCIe Gen 5 slots leave room to grow.
ECC on both system and video memory means a bit-flip halfway through a long run gets corrected instead of silently corrupting the result.
Specification
-
Processor: AMD Ryzen Threadripper PRO 9975WX, 32 cores / 64 threads, up to 5.4 GHz
-
Graphics: 2 × Leadtek NVIDIA RTX PRO 4000 Blackwell, 24GB GDDR7 ECC each (48GB total), 8,960 CUDA cores and 672 GB/s per card, PCIe Gen 5, single slot
-
Motherboard: ASRock TRX50 WS, workstation series, PCIe Gen 5, Wi-Fi
-
Memory: 128GB G.SKILL T5 NEO DDR5-6400 ECC Registered (4 × 32GB), quad-channel
-
Storage: 4TB XPG MARS 980 PRO, M.2 2280, PCIe Gen 5 x4, with heatsink
-
Cooling: SilverStone XE360-TR5 360mm liquid cooler, full sTR5 coverage
-
Power: ASUS Pro Workstation 1600W Platinum, fully modular, ATX 3.1
-
Chassis: In Win 915 full tower, tempered glass, aluminium
