Unified memory or VRAM for local LLMs?
Last updated
Pick VRAM when your models fit in it, up to about 32B on 24 GB, because it's fast and CUDA tools run unchanged. Pick unified memory for 70B and larger in one box. It holds far more, but those models generate fewer tokens per second.
Olares One → Local LLM hardware → VRAM · quantization · tokens per second → The right machine for your model size
Unified memory or VRAM for local LLMs: which should you pick?
Unified memory or VRAM for local LLMs comes down to model size. Pick VRAM when your models fit in it, and unified memory when they don't. VRAM sits beside a discrete GPU and tops out at 32 GB on consumer cards in our data. Unified memory, shared by CPU and GPU, reaches 128 GB or more.
The Mac Studio M5 Ultra goes up to 512 GB in Apple's largest configuration.
Unified memory appears on Apple silicon, NVIDIA's GB10 machines and AMD's Ryzen AI Max+ 395. VRAM appears on any discrete NVIDIA or AMD card, including the RTX 5090 Mobile inside Olares One.
How do capacity and bandwidth compare?
Unified-memory machines hold far more than any VRAM option, but their bandwidth varies widely. NVIDIA lists 273 GB/s for the DGX Spark. Apple lists 460 GB/s for the base M5 Max chip and 1.2 TB/s for the M5 Ultra. A used RTX 3090's GDDR6X runs at 936 GB/s, per the build guide we cite.
| Machine | Memory for models | Type and bandwidth | CUDA |
|---|---|---|---|
| Olares One | 24 GB | GDDR7 VRAM (bandwidth not on Olares' spec page) | Yes |
| Used RTX 3090 tower | 24 GB | GDDR6X VRAM, 936 GB/s | Yes |
| RTX 5090 desktop card | 32 GB | GDDR7 VRAM, 512-bit bus | Yes |
| DGX Spark / ASUS Ascent GX10 | 128 GB | Unified LPDDR5x, 273 GB/s (DGX Spark) | Yes |
| Mac mini M5 Pro | 64 GB | Unified, 307 GB/s listed for M5 Pro | No |
| Mac Studio M5 Max | 36 to 128 GB | Unified, 460 GB/s on the base chip | No |
| Mac Studio M5 Ultra | 96 to 512 GB | Unified, 1.2 TB/s | No |
| Ryzen AI Max+ 395 mini PCs | 128 GB total, GPU share lower | Unified LPDDR5X-8000 | No |
What does bandwidth mean for speed?
Generation speed is bounded by memory bandwidth divided by the bytes each token reads. A larger model reads more bytes per token, so on any machine it runs slower than a smaller one. Unified memory lets you load a 70B or 120B model, and each of its tokens costs far more reads than a 32B model's.
Among unified machines, the M5 Ultra's bandwidth sits far above the DGX Spark's, so on the same model the Mac should generate faster. The DGX Spark answers with CUDA.
Why do prompt processing and software matter?
Prompt processing and software can decide the choice even when memory doesn't. Prefill, the step that reads your prompt, depends on GPU compute more than bandwidth, and NVIDIA GPUs with tensor cores tend to be strong at it. Software decides which engines run. vLLM targets CUDA first, while Macs lean on llama.cpp, Ollama and MLX.
Ryzen AI Max+ boxes rely on ROCm or Vulkan builds of the same engines. Measure prefill with your own long prompts before you buy. The GPU options are compared in best GPU for local LLMs.
When does unified memory win?
Unified memory wins only once a 70B or larger model is a must. Below that, VRAM is our recommendation. Olares One ran Qwen3-30B-A3B at 157 tokens per second with one request on vLLM, in Olares' own test, and its CUDA stack runs vLLM unchanged. See hardware for a 70B model for the bigger machines.
Questions & answers
Is a Mac Studio better than an NVIDIA GPU for local LLMs?
For very large models in one box, the Mac Studio M5 Ultra holds far more, up to 512 GB in Apple's largest configuration. For models that fit in 24 GB or 32 GB and for CUDA tools such as vLLM, an NVIDIA GPU is the simpler choice.
Can the GPU on a Ryzen AI Max+ 395 use all 128 GB?
No. The 128 GB is shared with the CPU, and the GPU can use a smaller share. GMKtec doesn't state how much.
Sources for this entry
- Olares One: maker's page (Maker-published)
- Used RTX 3090 tower build: maker's page (Maker-published)
- NVIDIA GeForce RTX 5090 (desktop card): maker's page (Maker-published)
- NVIDIA DGX Spark: maker's page (Maker-published)
- ASUS Ascent GX10: maker's page (Maker-published)
- Apple Mac mini (M5 Pro, 64GB): maker's page (Maker-published)
- Apple Mac Studio (M5 Max): maker's page (Maker-published)
- GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB): maker's page (Maker-published)
- Olares blog: local AI hardware benchmarking (Oct 2025) (Company-published)
Continue this reading path
How the related websites connect
Cite this entry
Best Local LLM Hardware. “Unified memory or VRAM for local LLMs?.” Content reviewed 13 November 2026.
https://bestlocalllmhardware.com/unified-memory-vs-vram/