Memory for an AI development workstation is chosen in three steps: enough capacity for the GPUs and data pipeline, enough channels to keep the CPU side fed, and ECC for long runs. Puget Systems recommends at least twice as much system memory as total GPU memory. From there, the GPU count and the capacity target decide whether a desktop is enough or a workstation platform with registered ECC memory is needed.
Step 1: capacity from the GPUs
| Total GPU memory | Minimum system RAM (2x) | Practical kit | Platform |
|---|---|---|---|
| 32GB (two 16GB GPUs) | 64GB | 64 to 128GB | Desktop or TRX50 |
| 48GB | 96GB | 128GB | Desktop (2 x 64GB) or TRX50 |
| 96GB | 192GB | 256GB | TRX50 or W790 |
| 192GB (four 48GB GPUs) | 384GB | 512GB | TRX50, WRX90 or W790 |
| 384GB (four 96GB GPUs) | 768GB | 768GB to 1TB | WRX90 or W790 (W-3400/W-3500) |
System memory stages datasets, holds the framework and tools, and receives anything offloaded from GPU memory. Running out shows up as swapping and stalled data loaders, not a clean error. For sizing by use case, see how much RAM an AI workstation needs.
Step 2: channels and bandwidth
- Data loading and preprocessing run on the CPU and stream through memory. More channels keep more data loader workers busy.
- CPU inference and offload are bound by bandwidth: tokens per second scale with it. See capacity vs speed vs bandwidth.
- Fill every channel. Two on desktops, four on TRX50 and W-2400/W-2500, eight on WRX90 and W-3400/W-3500. Half-filled channels give half the bandwidth.
Step 3: ECC and module type
Training and fine-tuning runs last hours or days, and a memory error can crash a job or corrupt a checkpoint. Workstation platforms use ECC registered memory (RDIMM) by design. Desktops use unbuffered memory, and ECC there depends on the processor and board. See the ECC vs non-ECC guide.
Platform at a glance
| Platform | Memory | Practical ceiling | GPUs at full width |
|---|---|---|---|
| Desktop (AM5, LGA1851) | DDR5 UDIMM, 2 channels | 192 to 256GB | One |
| AMD TRX50 | DDR5 ECC RDIMM, 4 channels | Up to 1TB | Two to three |
| Intel W790 | DDR5 ECC RDIMM, 4 or 8 channels | Up to 2TB or 4TB | Up to four, board dependent |
| AMD WRX90 | DDR5 ECC RDIMM, 8 channels | Up to 2TB | Four or more |
For the AMD choice in detail, see when to choose WRX90 instead of TRX50; for Intel, what is Intel W790.
Memory for AI workstations
NEMIX RAM lists matched DDR5 ECC RDIMM kits for WRX90 (ASUS Pro WS WRX90E-SAGE SE, ASRock WRX90 WS EVO), TRX50 (ASUS Pro WS TRX50-SAGE WIFI, ASRock TRX50 WS), W790 (ASUS Pro WS W790-ACE and W790E-SAGE SE) and complete workstations such as the Dell Precision 7960 and Lenovo ThinkStation P7. Browse DDR5 RDIMM, or data center server memory when the work moves to servers.
Frequently asked questions
How much system RAM does an AI development workstation need?
Puget Systems recommends at least double the total GPU memory. Two GPUs with 16GB each (32GB of VRAM) point to at least 64GB of system RAM; four 48GB GPUs point to at least 384GB.
Do AI workstations need ECC memory?
For long training and fine-tuning runs, yes: a memory error can crash a job hours in or corrupt a checkpoint. RDIMM workstation platforms use ECC by design.
RDIMM or UDIMM for AI?
Desktop platforms take UDIMM and top out around 192 to 256GB. Past that, or with three or more GPUs, a workstation platform with RDIMM (TRX50, WRX90, W790) is the practical choice. See memory types explained.
Does memory speed matter for AI?
It matters most when the CPU does the work: data loading, preprocessing and CPU inference are bandwidth-bound. Filling every channel matters more than the module's rated speed, which runs at the platform's speed anyway.
Can I run large models on the CPU instead of the GPU?
Yes, with enough memory, but it is slower and bound by memory bandwidth. See WRX90 vs EPYC for AI for the numbers.
Buying for more than one system
IT teams and integrators that maintain AI development workstations for research and engineering teams can standardize on one kit per model. Request a quote with your models and quantities, or send us your system model and we will confirm the right module before you order.
Steps and rules are summarized from the vendors' published documentation; follow the guide for your exact model.