GPU memory matters more. RAM is secondary.

Yes, GPU memory is the bottleneck for training large models. Without enough VRAM, the model won’t even load. System RAM helps with data loading, caching, and offloading parameters via CPU (which is slow), but it’s not where the actual training math happens.

For LLM training, you’re fighting VRAM limits constantly. Batch size, sequence length, model size—all VRAM constraints. 24GB on a consumer card gets you started with small models or heavy quantization. 48GB+ (A6000, A100, etc.) is where real training lives.

64GB system RAM is nice to have if you’re loading large datasets or doing multi-GPU setups where CPU-side memory needs grow. But if you’re choosing between more system RAM and a better GPU, get the better GPU every time. Even 32GB system RAM is fine for most training workflows. The GPU is the engine; the rest is just support.

Don’t overthink this one: VRAM is the scarce resource, RAM is cheap insurance.