Ram, Vram, And Unified Memory

RAM (system memory) is what your computer uses for normal operations: running programs, opening files, holding data. Your laptop has 8-64GB typically. Server-class machines go higher. RAM is connected to the CPU through a memory bus.

VRAM (video memory) is RAM attached directly to the GPU. It's faster than system RAM and physically closer to the chip that does the math. For traditional ML workloads, this is where the model lives. Consumer GPUs (RTX 4090, RTX 5090) have 24-32GB of VRAM. Data center GPUs (H100, H200, B200) have 80-192GB.

UNIFIED MEMORY is what Apple Silicon Macs use. The CPU, GPU, and Neural Engine all share a single pool of memory. There's no copy-from-RAM-to- VRAM step because there's no separate VRAM. A Mac Studio with 192GB of unified memory can run models that would require an enterprise GPU on the PC side.

REGISTERED VS. UNREGISTERED RAM (also called ECC vs. non-ECC). Server- grade RAM has error-correcting code circuits that detect and correct bit flips in memory. This is the "registered" or "ECC" RAM. Consumer RAM usually doesn't have ECC and is cheaper. For running models, you do not need ECC RAM. For 24/7 production servers handling money or medical records, you do.

The practical implications:

For LOCAL CODING ASSISTANCE, you want enough memory to load a 13B-30B model at Q4. That's 8-20GB of accessible memory. A Mac mini with 32GB of unified memory is enough. A PC with a 24GB-VRAM GPU is enough. Either is fine.

For LARGER MODELS (70B+), you want 48-80GB of accessible memory. Mac Studio (96GB or 192GB unified) is the cleanest path. On the PC side, you're looking at multi-GPU setups or buying a workstation card.

For VERY LARGE MODELS (200B+, even quantized), you're in specialized hardware territory: NVIDIA DGX systems (Spark, B200-based servers), Mac Pro with maxed-out unified memory, or you're renting cloud GPUs.

For RUNNING AGENTS 24/7, the memory budget needs to also fit the OS, your inference server, your agent framework, and any other software. Add 4-8GB on top of the model's memory requirement.

The decision framework:

1. What model do I want to run?
2. At what quantization?
3. That gives me my memory floor.
4. Add 4-8GB for system overhead.
5. That's the minimum hardware I should buy.