I'm in Australia and have a total budget of A$6,000 for a Linux workstation that will run 24/7. I already own a Cooler Master HAF X case and a Super Flower Leadex 7 Pro 1.2kW Platinum PSU, although the case still needs suitable fans and dust filters. I also have roughly 4TB of NVMe and SATA storage.
The parts I still need are the CPU, motherboard, CPU cooler, RAM, and GPUs. The system will be used for containers, virtual machines, video, photo and audio production, CAD, 3D modelling, gaming, and local AI training, tuning, and inference. I'm aiming for around 64GB of total VRAM and would ideally like space for two to four GPUs.
The motherboard must include onboard IPMI or a BMC for remote management, support ECC memory, and provide enough PCIe lanes and physical slots for multiple GPUs. I'm considering Threadripper, EPYC, used server hardware, professional cards, or other workstation platforms rather than a typical consumer desktop.
The machine will run an Arch-based Linux distribution and needs to support significant undervolting so it can operate continuously at reasonable temperatures. It will be connected to a 230V circuit protected by a 10A breaker that also powers other equipment, so power draw matters. What parts or platform would make sense in Australia, and is this realistically achievable for A$6,000?
3 Answers
This combination is possible, but it is not really a normal desktop build. IPMI, ECC support, lots of PCIe lanes, four GPU slots, and 64GB of VRAM point toward used EPYC or workstation/server hardware. A used EPYC platform with a board that has a BMC may give you the best balance of cores, memory capacity, and PCIe lanes. Threadripper Pro is another option, but Australian pricing for the CPU and motherboard can consume most of the budget before adding GPUs and ECC RAM.
Be especially careful with the physical slot layout. Four x16-length slots do not necessarily mean four slots with useful electrical bandwidth; check the actual lane allocation and whether the cards fit without blocking airflow. Used enterprise GPUs could make the VRAM target more affordable, but Linux drivers, virtualization support, power connectors, and software compatibility need to be checked for each model.
Your existing 1.2kW PSU and a 10A circuit are important constraints. At 230V, a 10A circuit has about 2.3kW available in theory, but the other equipment on that circuit reduces the practical headroom. A multi-GPU workstation can also create large startup and sustained loads, so measure the actual circuit usage rather than relying only on the PSU rating.
I would avoid building this around a 5090, since its Australian price would leave too little for the workstation platform. A used server board, CPU, ECC RDIMM kit, and supported professional GPUs may fit the budget better, but the HAF X may need airflow modifications and GPU support brackets. Do not run a high-power multi-GPU system directly on carpet; raise it off the floor and provide strong front-to-back airflow. Also confirm that the motherboard, GPUs, and required virtualization features work properly under your chosen Linux distribution before buying.
That makes sense. The goal is a machine that can run continuously rather than a maximum-gaming build, so lower power limits and undervolting are more important to me than chasing peak benchmark performance.
Before choosing hardware, make sure 64GB of VRAM is actually required as one combined resource. Multiple GPUs generally do not behave like a single large pool of memory; many applications need to split the workload between cards, and some cannot use separate GPUs efficiently at all. For AI, CAD, and production software, the exact application and GPU ecosystem matter more than the total VRAM number. A pair of cards with supported drivers may be much more useful than four mismatched cards.
Cloud compute is not attractive if the workload runs constantly, but it may still be worth using temporarily for workloads that need a large unified-memory GPU. Buying older professional or data-centre cards locally could reduce the upfront cost, although they may need special cooling and can be difficult to integrate into a tower case.
The system will be used for local AI as well as CAD, 3D, and media work, so I do want substantial VRAM locally. I’m not expecting separate cards to magically become one pool, but I’d rather buy a system that can distribute workloads across multiple GPUs than be limited to a single consumer card.
I’ve also been checking second-hand EPYC and Threadripper systems, but the local prices are much higher than expected. The main problem is finding something with both a BMC and enough usable PCIe slots without spending A$1,000 or more on the CPU and another A$1,300 on the motherboard.