Rasheed Sons AI Workstation — Threadripper PRO 9965WX
A workstation built for running and fine-tuning AI models on your own hardware, under your own roof. Twenty-four Zen 5 cores, 128GB of quad-channel ECC memory and an NVIDIA RTX PRO 4000 Blackwell with 24GB of ECC GDDR7 — assembled, cable-managed and tested in Johannesburg. No cloud subscription, no per-token billing, and nothing leaving your premises.
What it will actually run
Two numbers decide this: 24GB of video memory, and 128GB of system memory running quad-channel at 6400MT/s — roughly 205 GB/s of bandwidth, which is what makes large models practical even when they don't fit on the card.
-
7B to 14B models — Llama, Mistral, Qwen, Gemma class — run entirely on the GPU at full speed. Chat assistants, RAG over your own documents, code completion, batch document processing.
-
32B models, quantised, sit in video memory with headroom. This is where local models start genuinely competing with hosted ones for reasoning and code.
-
70B models are properly usable here. They don't fit in 24GB alone, but with 128GB of fast quad-channel memory behind the card you can split the model across both rather than being stopped by it.
-
Very large mixture-of-experts models are on the table too, since 128GB holds far more than the card alone and only part of the model is active at any moment.
Beyond text: Stable Diffusion and Flux image generation, Whisper transcription, local embeddings and vector search, and LoRA fine-tuning of smaller models without renting an H100 by the hour.
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 have downloaded a second one. Gen 5 sequential speed means loading a model into memory takes seconds rather than minutes, which matters when you switch between them during a working day.
Built to grow
The graphics card is single slot, and that's deliberate. The board carries multiple PCIe Gen 5 x16 slots and the power supply is a 1600W unit, so this machine starts at one card and takes more as the workload grows — without replacing the platform. Blower cooling is what makes that work: cards exhausting out the back stay stable when stacked, where open-fan consumer cards choke on each other's hot air.
ECC on both system and video memory means a bit-flip halfway through a long run gets corrected rather than silently poisoning the result.
Specification
-
Processor — AMD Ryzen Threadripper PRO 9965WX, 24 cores / 48 threads, up to 5.4 GHz
-
Graphics — NVIDIA RTX PRO 4000 Blackwell, 24GB GDDR7 ECC, 672 GB/s, PCIe Gen 5, single slot
-
Motherboard — ASRock TRX50 WS, workstation series, PCIe Gen 5, Wi-Fi 6E
-
Memory — 128GB G.SKILL 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, fully modular
-
Chassis — In Win 915 full tower, tempered glass, aluminium
