.png?width=274&height=200&scale=canvas)
.png?width=500?width=274&height=200&scale=canvas)
.png?width=274&height=200&scale=canvas)
.png?width=274&height=200&scale=canvas){kind=icon}
AI Ready PC Components
Designed to harness AI for faster workflows, smarter multitasking, and more powerful everyday performance.
AI workloads expose every weakness in standard hardware.
Standard hardware buckles fast under real AI workloads - undersized memory, narrow bandwidth, thermal throttling under sustained load. Push it hard enough and the weak point always shows first. These are the five specs that actually hold up a serious build - get any one of them wrong, and the rest of the system is compensating for it.
RAM and VRAM Capacity
Load larger models fully into memory to avoid the slowdowns caused by offloading to storage.
Maximum Bandwidth
Increase data speeds between the CPU, GPU, and memory to keep inference times low.
Sustained performance
Avoid thermal throttling with hardware designed for long-term endurance, not just short benchmarks.
Multi-component scalability
Run multiple high-demand components simultaneously without hitting PCIe or power bottlenecks.
Validated Stability
Ensure predictable, consistent output with components tested for professional AI pipelines.
Build your AI workstation
Answer a few quick questions and we'll map out exactly what you need - or read on for the logic behind each decision.
Looking for a complete solution?
If you prefer to skip the assembly and go straight to deployment, the NVIDIA DGX Spark box provides a fully integrated, enterprise-grade ecosystem ready for immediate production.
Best for: developers and teams who want data-centre-class local AI on a desk, without building a workstation or renting cloud space.
Why choose DGX Spark over a custom build?
- Validated stability: Unlike custom-sourced parts, the DGX Spark is a factory-validated system, ensuring predictable and consistent output for professional AI pipelines .
- Zero configuration friction: Bypass the challenges of matching PCIe lanes, selecting ECC RDIMM memory, and sizing dual-PSU setups. The system arrives pre-configured for frontier models and massive datasets.
- Industrial-scale performance: Designed for sustained throughput rather than short benchmarks, the DGX Spark is built to handle the most demanding AI workloads without thermal throttling or power bottlenecks.
Why each component depends on the last
Every part on this page is chosen because of the one before it - GPU count sets the CPU, the CPU sets the motherboard, the motherboard sets your memory ceiling. AI workloads quickly push standard hardware to its limit, and you need a system built for sustained throughput and massive memory capacity, rather than components designed for short bursts of speed.
Here's how that chain works, whether you're double-checking the wizard's answer or working through a build by hand.
1 - Define your workload
What are you running, and how demanding is it?
Local LLM inference, model fine-tuning, rendering, simulation, CAD?
This sets your scale and, crucially, your GPU count - the single decision everything else flows from.
2 - Graphics card
It's the heart of an AI build and the biggest cost, so everything else is sized to serve it.
The decision comes down to two questions:
- How much VRAM your workload needs (16GB for entry CAD and light AI, 48GB for serious rendering and local LLMs, 96GB for frontier models), and whether it's going in a desktop or a server.
- Pick a Workstation card for a desktop tower, a Max-Q if you're stacking several cards in one chassis, and a Server Edition for a rack.
3 - Processor
This is where GPU count forces the platform.
Choose by how many GPUs you're feeding and how big your datasets are:
- Threadripper 9000 is the cost-effective pick for one or two GPUs
- Threadripper PRO 9000WX is the platform for three-to-four-GPU AI thanks to its 128 PCIe lanes and 8-channel ECC memory.
- Go Intel Xeon W only if your professional software is specifically certified on Intel.
4 - Motherboard
The CPU dictates the socket/chipset.
The chipset is the whole decision, because it sets your CPU, GPU and memory ceilings at once:
- TRX50 suits 1–2 GPU Threadripper builds
- WRX90 unlocks the full Threadripper PRO platform for multi-GPU AI (seven x16 slots, 8-channel ECC, dual-PSU)
- W790 is the Intel Xeon equivalent.
Match the chipset to the CPU you've chosen and the number of GPUs you need to run.
5 - Memory
The platform fixes the type and channel count.
Your platform dictates the type, not your preference - every workstation board here needs DDR5 ECC RDIMM, and ordinary UDIMM simply won't work.
From there it's just channel count:
- Four matched modules for a TRX50 Threadripper build scaling capacity (128GB up to 256gb) to how much data you need to hold in memory.
- Eight for an 8-channel WRX90 or W790 Threadripper PRO build scaling capacity (256GB up to 2TB) to how much data you need to hold in memory.
6 - Storage
Internal storage serves two distinct roles, separate from your NAS.
- A fast PCIe NVMe SSD should be used for the operating system and to load model weights into VRAM quickly; for this drive, prioritise sequential read speeds (Gen4 is typically sufficient, while Gen5 offers the fastest possible loading times).
- For active datasets, checkpoints, and local scratch space, prioritise capacity and drive endurance. Use the NAS as your shared, backed-up library for long-term storage, keeping your day-to-day working data on these high-performance internal drives.
7 - Power supply
You can only size this correctly after you know the GPUs and CPU, because it's the sum of their power plus headroom.
Size it to your hardware and don't skimp on the standard:
- Add up your GPU and CPU wattage
- Add about 30% headroom
- Always use a native ATX 3.1 / 12V-2x6 unit
One flagship GPU build lands around 1300–1600W, a dual-GPU or 96-core build wants 2000–2200W, and three-to-four-GPU rigs need a dual-PSU setup.
8 - Cooling
Ensure your cooling solution is matched to the CPU socket and GPU layout.
- High-core Threadripper or Xeon processors (typically 280–350W) require a socket-rated air tower or an AIO liquid cooler compatible with sTR5, sWRX8, or LGA 4677.
- For multi-GPU configurations, heat is primarily managed at the card level (via Max-Q blowers or server-grade coolers) and through robust case airflow, rather than the CPU cooler.
9 - Case
Select a chassis that accommodates all physical components.
- The case must support the motherboard form factor (often E-ATX or larger), provide sufficient clearance for full-height GPUs, and offer the necessary expansion slots and airflow for multiple cards.
- If the build requires a second PSU, ensure the chassis has the space to house it. Depending on the environment, choose a full-tower for desktop use or a 2U/4U rack chassis for server-room integration.
10 - Uninterruptible power supplies (UPS)
Protect your hardware with a UPS sized to your system's total power draw.
- For single-GPU workstations, a line-interactive pure-sine tower UPS is generally sufficient.
- For always-on systems or multi-GPU builds, an online (double-conversion) unit is recommended.
Then select your form factor - rackmount for servers or freestanding for a desktop
High-specification rigs typically require 2200VA+ with true sine-wave output to support active-PFC power supplies, ensuring the system can ride through power fluctuations and trigger a clean automatic shutdown.
11 - Network attached storage (NAS)
Implement a NAS to provide shared, protected storage for your data. While fast internal NVMe SSDs should be used to load models onto the GPU machine, the NAS acts as the central library for datasets and checkpoints.
Network speeds should be matched to the workload:
- 2.5GbE is suitable for a single machine
- 10GbE (or all-flash arrays) is recommended when streaming large datasets to multiple systems.
We recommend RAID 6 for redundancy, though please remember that RAID is not a substitute for a dedicated backup strategy.
Frequently asked questions
What makes a component "AI ready"?
It comes down to the demands of sustained AI and professional workloads, which differ from gaming. AI requires massive memory pools, ECC (Error Correction Code) for stability during long processing jobs, a high number of PCIe lanes to feed multiple GPUs, and stable, high-wattage power. A setup with a GPU featuring 48–96GB of VRAM, a CPU with 128 PCIe lanes, ECC RDIMM memory, and an ATX 3.1 PSU is "AI ready" in a way that consumer-grade hardware simply isn't.
Can I use normal gaming components for AI?
For light, single-GPU local AI, yes—a high-end gaming GPU and a standard desktop will run smaller models. However, you'll quickly hit a ceiling. Consumer CPUs lack the PCIe lanes to run several GPUs at full speed, consumer boards don't support ECC RDIMM, and gaming GPUs have limited VRAM. Workstation components are designed specifically to remove these bottlenecks.
Do I need workstation parts, or will a powerful desktop do?
If you're running a single GPU and don't require ECC or certified drivers, a powerful desktop is usually fine. But the moment you need multiple GPUs, massive memory capacity, 24/7 reliability, or software that requires certified professional drivers, the investment in a workstation platform becomes necessary.
Graphics cards
How much VRAM do I need for local AI?
As a rough guide for 4-bit quantised LLMs, a 7–13B model fits in 16GB, a ~30B model needs around 24–32GB, and a 70B model requires roughly 48GB. Frontier models, long context windows, or unquantised models will push you toward 96GB. For rendering and simulation, it depends on your dataset or scene size, but the principle is the same: check your peak usage and aim for the tier above it.
What’s the difference between Workstation, Max-Q and Server editions?
They often share the same silicon but differ in cooling and power. Workstation editions are full-power with active cooling for desktop use, while Max-Q runs at lower power with a blower cooler, allowing you to stack several cards in one chassis. Server editions are passively cooled and rely on the high-pressure airflow of a rack server. In short, use Workstation for desktop towers, Max-Q for multiple cards in one box, and Server for rackmounts.
Can I game on a workstation graphics card?
You can, but it isn't cost-effective. You're paying a premium for VRAM, ECC, and certified drivers that don't benefit gaming. A gaming card will give you significantly more frames per pound. Workstation drivers are tuned for stability in CAD, rendering, and AI rather than peak gaming performance.
Can I run multiple GPUs together?
Yes, provided you have the right platform. You'll need sufficient PCIe lanes, such as a Threadripper PRO / WRX90 board, so each card has full bandwidth, enough power, which often requires a dual-PSU setup, and appropriate cooling, such as Max-Q blower or Server cards. Note that current RTX PRO Blackwell cards do not use NVLink, so multi-GPU scaling is handled over PCIe.
Why choose a workstation graphics card over a GeForce RTX 5090?
It comes down to memory, reliability and support. The 5090 is a fast 32GB gaming card, but workstation cards go up to 96GB with ECC (error-correcting) memory, ISV-certified professional drivers, and blower or passive cooling built for stacking several cards — none of which the 5090 offers. If your model or scene fits in 32GB and you don’t need ECC, certified software support or multi-card density, the 5090 wins on raw value; the moment you need more VRAM, error correction, certified support or several GPUs in one chassis, only a workstation card will do.
Processors and motherboards
Why a Threadripper instead of a Ryzen or Core CPU?
It boils down to PCIe lanes and memory. Mainstream CPUs typically have 20–24 usable PCIe lanes, which is only enough for one GPU. Threadripper offers 80 or more, with Threadripper PRO providing 128, allowing them to feed several GPUs at full x16 speed while supporting much larger amounts of ECC memory. For multi-GPU AI, that lane count is essential.
What’s the difference between Threadripper and Threadripper PRO?
Threadripper (HEDT) is the cost-effective choice for one or two GPUs, offering quad-channel memory and approximately 80 lanes. Threadripper PRO is designed for the heaviest professional work, adding 8-channel ECC memory up to 2TB, the full 128 PCIe 5.0 lanes, and advanced management and security features. It is the platform of choice for three-to-four GPU AI builds.
AMD or Intel for a workstation?
AMD Threadripper and Threadripper PRO generally lead on core count and PCIe lanes, making them ideal for multi-GPU AI and heavily threaded rendering. Intel Xeon W is the preferred choice when your professional software is specifically certified for Intel, or for single-GPU, core-intensive workflows where ecosystem stability is the priority.
How do I know which motherboard fits my CPU?
The chipset is tied to the CPU. Threadripper 9000 requires a TRX50 board, Threadripper PRO needs WRX90, and Intel Xeon W needs W790. Once you've selected the CPU, you can choose a board within that chipset based on your requirements for slots, M.2 storage, and networking.
Memory
Why can’t I use normal RAM in these systems?
Threadripper and Xeon W workstation boards require DDR5 ECC RDIMM registered memory. Ordinary UDIMM, even ECC UDIMM, will not boot. ECC is standard on these platforms because it protects long AI and rendering jobs from silent bit-flips that can crash a project hours into a run.
How much memory should I get, and how many sticks?
Capacity should match your dataset, with 128–256GB suiting most HEDT builds, while PRO workloads often scale to 512GB–2TB. Module count is equally important; you should populate one stick per channel—four on TRX50 and eight on WRX90 or W790—to achieve full memory bandwidth. You should always use a matched kit.
Power supplies
What wattage PSU do I need?
Total your GPU and CPU TDP and add roughly 30% headroom. A single flagship-GPU build usually lands between 1300–1600W, while a dual-GPU or 96-core build typically needs 2000–2200W. Systems with three or four GPUs generally exceed the capacity of a single PSU and require a dual-PSU configuration.
Do I need an ATX 3.1 power supply?
This is strongly recommended. Modern high-end GPUs can have large, brief power spikes that older ATX 2.x standards struggle to handle. ATX 3.1 units provide the native 12V-2x6 connector, removing the need for adapters and improving stability under heavy load.
Can one PSU run three or four GPUs?
Usually not. A multi-GPU build drawing 600W per card exceeds both the output of a single PSU and the capacity of a standard UK wall socket. These rigs typically use a dual-PSU setup, consisting of a primary ATX unit plus a secondary PSU, which WRX90 boards support natively.
Does PSU efficiency rating matter?
Yes, especially for machines running long-term jobs. 80 Plus Platinum or Titanium units waste less power as heat, which reduces running costs and helps manage thermals in a dense build. Full modularity is also highly beneficial for cable management and airflow.
Cooling
How do I cool a multi-GPU build?
Your choice of card is critical here. Passively cooled Server cards require the high-pressure airflow of a server chassis. Blower-style Max-Q cards exhaust heat directly out the back, making them far better for stacking in a single chassis than flow-through cards, which dump heat into the card above.
Do I need liquid or air cooling for the CPU?
High-core Threadripper and Xeon W chips can draw 280–350W and require substantial cooling, either a large air cooler rated for the socket or an all-in-one liquid cooler. Ensure the cooler is rated for the specific socket, such as sTR5, sWRX8, or LGA 4677, as these mounts are larger than those on mainstream CPUs.
Will a multi-GPU workstation be loud?
Under sustained AI load, yes. Moving the amount of air required for multiple GPUs and high-wattage CPUs creates significant noise. While Platinum PSUs and a spacious case help, a heavily loaded multi-GPU machine is best placed where noise isn't an issue or managed remotely.
Cases
What case do I need for a workstation build?
You need to check three things. First, it must support the board form factor, as workstation boards are often E-ATX or larger. Second, it must have enough PCIe slots and internal clearance for full-height GPUs. Finally, it must provide strong front-to-back airflow. If you are running multiple GPUs or dual PSUs, ensure the case has dedicated mounting space for the second power supply.
Will my graphics card physically fit?
Not always. Flagship Workstation cards are full-height and won't fit a 2U chassis, whereas Max-Q and Server cards are shorter and will. Always check the card's length and slot width against the case's clearance, ensuring there is enough spacing between cards for airflow.
Can I put workstation parts in a rack instead of a tower?
Yes, this is exactly what Server-edition GPUs and boards with BMC or IPMI are designed for. A rack chassis, such as 2U or 4U and above, handles the cooling for passive cards and supports remote management, which is the ideal route for unattended, 24/7 operation.
How much space do I need for memory and dual PSUs?
Eight-channel boards require eight memory modules, and dual-PSU builds need extra mounting space. Both take up significant internal room and can obstruct airflow. A large E-ATX full-tower or a proper rack chassis is recommended to avoid choking the system.
Uninterruptible power supplies (UPS)
Do I need a UPS for an AI workstation?
While not strictly required, we highly recommend one. A sudden power cut can corrupt your filesystem or wipe out a multi-hour training run. If you're running an always-on inference or agent box, a UPS ensures it rides through brief outages. At the very least, it gives your system enough time to perform a clean, automatic shutdown.
What’s the difference between line-interactive and online (double-conversion) UPS?
A line-interactive UPS passes mains power through normally and switches to battery in a few milliseconds if the power fails—this is usually perfectly fine for a standard desktop PSU. An online UPS continuously runs your gear off its own regenerated power. This means there is zero transfer gap and your equipment receives fully conditioned power at all times. It’s the best choice for critical, always-on systems, though it comes at a higher price point.
What’s the difference between an online UPS and a rackmount UPS?
These actually refer to two different things. "Online" describes the topology (how the UPS delivers power), while "rackmount" describes the form factor (a unit designed to bolt into a 19-inch server rack). You can have a line-interactive tower, an online rackmount, or any other combination. We suggest picking your topology based on how critical your uptime is, and your form factor based on where the unit will sit.
What size UPS do I need?
Your UPS’s VA and watt rating must exceed your system’s actual power draw, with some headroom to spare. Keep in mind that the more load you put on the battery, the shorter the runtime will be. A single-GPU workstation is typically fine on 1000–1500VA, but a multi-GPU rig drawing 2000W+ will need a larger unit (2200VA+) and may require its own dedicated circuit.
Does the UPS need a pure sine-wave output?
Yes, for these specific builds. Most modern high-end PSUs use active PFC, which can struggle or fail when paired with the "stepped" or simulated sine waves found in cheaper UPS units. To be safe, always choose a true (pure) sine-wave model.
Will a UPS let me keep working through a power cut?
Generally, not for long. Large multi-GPU rigs drain batteries very quickly. Rather than thinking of it as a way to work for hours, it's better to view it as a safety net to bridge brief blips and trigger a graceful automatic shutdown (via USB or network cable) before the battery runs out.
Network attached storage (NAS)
Do I need a NAS for AI work?
If you have a single workstation and only do occasional work, you probably don't. However, a NAS becomes essential once your datasets grow, when multiple machines need access to the same data, or when you need reliable backups of checkpoints and results. It provides a central, protected library rather than having data scattered across various local drives.
NAS or a fast SSD—which one do I need?
Actually, you likely need both, as they do different jobs. A fast PCIe NVMe SSD inside your GPU machine is what you use to load models quickly for inference. The NAS is your shared, backed-up library where the bulk of your datasets and checkpoints live. They aren't interchangeable, and buying one when you actually need the other is a common mistake.
What network speed does a NAS need for AI?
2.5GbE is the practical minimum, but 10GbE is preferred if you're streaming large datasets or checkpoints to your GPU machines—standard 1GbE will likely bottleneck your training pipeline. Remember to match your NAS speed to your fastest client; a 10GbE NAS talking to a 2.5GbE workstation will still be capped at 2.5GbE.
How many bays and what RAID should I use?
Four or more bays are sensible to allow for growing datasets. For production environments, we suggest RAID 6; it can survive two simultaneous drive failures. This is important because rebuilding very large drives in RAID 5 can be stressful on the hardware, risking a second failure during the process. Just remember that RAID protects against hardware failure, not accidental deletion or ransomware, so please follow the 3-2-1 backup rule.
Can a NAS run AI models itself?
Some can, but in a modest capacity. x86 NAS units with a decent CPU and 16GB+ RAM can run small quantised models or containers (like Ollama or Jupyter) via Docker. Some newer 2026 models even include NPUs for photo and video recognition. However, a NAS is primarily a storage node; for serious inference, you'll still want a GPU workstation or a DGX Spark, which the NAS then feeds with data.
Which NAS brand should I choose?
Synology (DSM) is the most polished and best-supported option with long update cycles, though some newer models are more restrictive regarding third-party drives. QNAP often offers more hardware for your money and supports ZFS (QuTS Hero), but it has a steeper learning curve and requires more active security patching. UGREEN and TerraMaster offer the most hardware per pound, though their software is less mature. We recommend picking the OS you're happiest to use, then choosing the hardware that runs it.
Building
In what order should I choose the components?
You should work from the workload outwards. First, determine the workload, which sets the GPU type and count. The GPU count then determines the required CPU for PCIe lanes, and the CPU dictates the motherboard. The motherboard fixes the memory type, and the total power draw determines the PSU. The most common mistake is choosing a CPU or PSU first and finding they cannot support the GPUs needed for the task.
Can I build a complete AI workstation from these parts?
Yes. They are essentially divided into three matched lanes. AMD HEDT, using Threadripper and TRX50, is for one to two GPUs. AMD PRO, using Threadripper PRO and WRX90, is for multi-GPU AI. Intel, using Xeon W and W790, is for certified professional workflows. Pick the lane that matches your workload, and the rest of the components will follow.
DGX Spark
What is the DGX Spark?
The DGX Spark is a compact, desktop "personal AI supercomputer." It's built on NVIDIA’s GB10 Grace Blackwell Superchip, featuring 128GB of unified memory and up to 1 petaFLOP of FP4 AI compute. It’s a turnkey alternative to building a custom workstation, designed to run large models locally right out of the box.
DGX Spark or a Threadripper + RTX PRO build—which should I choose?
Go for the Spark if you want a sealed, NVIDIA-supported appliance for local AI with minimal setup and total on-prem privacy. Choose a custom Threadripper/RTX PRO build if you need Windows, upgradeable memory, multiple large GPUs, or if you plan on doing heavy sustained training or gaming. The Spark is primarily a development and inference tool, not a full-scale training cluster.
What can it run?
It can run models up to around 200B parameters locally using its 128GB of unified memory. If you link two units over their 200GbE ports, you can pool the memory to 256GB and run inference on models up to roughly 405B. It excels at prototyping, LoRA fine-tuning, and inference.
Can it run Windows?
No. It runs DGX OS (based on Ubuntu 24.04) exclusively, as there is no Windows driver support for the GB10 chip. You can use it as a Linux desktop, but most users run it "headless" as a remote AI server.
Can I upgrade the memory or storage?
The 128GB of unified memory is integrated into the superchip and cannot be upgraded. Storage does vary by model; the Founders Edition typically ships with 4TB, while OEM versions (from the likes of ASUS, Dell, Acer, or MSI) often come with 1TB.
Is it worth it compared to cloud GPUs?
If you're running models regularly, the Spark can pay for itself in a few months by removing hourly rental bills and latency, all while keeping your data on-premises. However, if your AI needs are only occasional or if you require massive peak throughput, the cloud or a discrete-GPU workstation might be more cost-effective.
Can I link multiple Sparks together?
Yes, you can. The DGX Spark features 200GbE ports specifically for this purpose, allowing you to link up to four units together. Linking two units pools their unified memory to 256GB, which enables you to run inference on much larger models—up to roughly 405B parameters. You can link up to four units together which expands your networking and compute capacity, this is primarily designed to scale memory for larger model inference and prototyping rather than creating a massive, high-end training cluster..