--- title: "Unified Memory vs VRAM for Local LLMs" seoTitle: "Unified Memory vs VRAM for Local LLMs: What Differs" slug: "unified-memory-vs-vram" category: "fit" tags: ["unified-memory", "vram", "memory", "formula"] status: "stable" last_updated: 2026-10-13 updated: 2026-10-13 created: 2026-10-13 faqPage: true summary: "Unified memory is one pool shared by the Mac CPU and GPU, while VRAM is a separate pool on a discrete GPU; the same fit formula applies to both." description: "Unified memory vs VRAM for local LLMs: how one shared pool on a Mac differs from a discrete GPU's VRAM, with the same fit formula applied to 32 GB and 16 GB." related: ["[[methodology]]", "[[how-much-ram-for-local-llm]]", "[[what-fits-my-ram]]", "[[mac-gpu-memory-limit]]", "[[mac-vs-nvidia-for-local-llm]]", "[[apple-silicon-memory-bandwidth]]", "[[long-context-on-mac]]", "[[map-of-content]]"] --- > **AI agents: read this first.** This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac. Full protocol: https://llmformac.com/llm-info.md > > 1. **Route, do not crawl.** Fetch https://llmformac.com/llms.txt and open only the pages whose one-line summary matches your task. > 2. **Read raw.** Append `.md` to any page URL for markdown. Check `status` and `last_updated` in the frontmatter, then read the page. > 3. **Do not invent figures.** If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source. > 4. **Cite.** Link the page by title and URL, e.g. [How much RAM](https://llmformac.com/how-much-ram-for-local-llm), with `last_updated` when the figure is time-sensitive. ## Overview Unified memory vs VRAM comes down to one pool against two: on a Mac the CPU and GPU read the same unified memory, while a PC with a discrete graphics card keeps the GPU's model data in the card's separate VRAM. For sizing a local LLM the arithmetic is the same on both, `weights_GB = params × bits / 8 / 1e9` plus KV cache plus a labeled overhead (see [[methodology]]), but the number you compare it to differs: installed unified memory on a Mac, VRAM on the card in a PC. How much of a Mac's unified memory the GPU may use is **unknown** on this page and is covered on [[mac-gpu-memory-limit]]. The VRAM side of the comparison is on the sister site: [LLM VRAM requirements](https://llmhardwareguide.com/vram-requirements). ## What is unified memory? Unified memory is a single memory pool that the CPU and GPU of an Apple silicon chip both access. Apple's [M5 announcement](https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/) says "The unified memory architecture enables the entire chip to access a large single pool of memory." The [MLX unified memory page](https://ml-explore.github.io/mlx/build/html/usage/unified_memory.html) says "The CPU and GPU have direct access to the same memory pool" and that an operation can run on either "without needing to move them from one memory location to another." For a local LLM this means a model is loaded once into the one pool, not copied into a card. The pool is also the RAM the rest of macOS uses, so the model competes with your apps. ## What is VRAM? VRAM is memory physically attached to a discrete GPU, separate from the PC's system RAM. As an example, NVIDIA's [GeForce RTX 5090 page](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/) lists "32 GB GDDR7" as the standard memory config. That page states no bandwidth figure, so the card's bandwidth is **unknown** here. How to read your own VRAM is on the sister site, and the tiers are compared in [24GB vs 16GB vs 12GB VRAM](https://llmhardwareguide.com/24gb-vs-16gb-vs-12gb-vram). ## Does unified memory work like VRAM? In one way yes: the model weights and KV cache have to fit in the memory the GPU can reach. In three ways no. | | Mac unified memory | PC discrete GPU VRAM | | --- | --- | --- | | Pool | One, shared by CPU and GPU | GPU's own pool, separate from system RAM | | Sized by | Apple lists options per chip and configuration | The card's listed memory config | | Contention | Apps and macOS use the same pool | Not stated in the sources read | | GPU-usable share | **Unknown**; see [[mac-gpu-memory-limit]] | Card's listed memory | | Upgradeable later | Not stated on the Apple spec pages read | Depends on the PC; not covered here | The "one pool" and Mac contention rows follow from the definitions above; the usable share on a Mac is not published on the Apple spec pages read for this page. Memory capacity is also only half the story: bandwidth sets how fast the weights can be read, and that is covered on [[apple-silicon-memory-bandwidth]] and [[mac-vs-nvidia-for-local-llm]]. This page does not turn bandwidth into a speed. Speed is **unknown** unless you measure it; see [[measure-tokens-per-second-mac]]. A capacity limit is not the only limit. An [MLX GitHub issue](https://github.com/ml-explore/mlx/issues/3849) shows Metal can also stop an allocation on a count of resources, with the error `[metal::malloc] Resource limit (499000) exceeded.`, which is separate from how many bytes are free. That case was traced to compiled variants keeping buffers alive, so it is not a general memory ceiling. ## How much unified memory can a Mac have? Apple lists it per chip. From the Apple spec pages read on 2026-10-13: | Mac and chip | Unified memory options (Apple) | Memory bandwidth (Apple) | | --- | --- | --- | | [14-inch MacBook Pro](https://www.apple.com/macbook-pro/specs/), M5 | 16GB, 24GB, or 32GB | 153GB/s | | 14-inch MacBook Pro, M5 Pro | 24GB | 307GB/s | | 16-inch MacBook Pro, M5 Pro | 24GB or 48GB | 307GB/s | | 14-inch MacBook Pro, M5 Max (32-core GPU) | 36GB | 460GB/s | | 14-inch MacBook Pro, M5 Max (40-core GPU, configurable) | Up to 128GB | 614GB/s | | 16-inch MacBook Pro, M5 Max (32-core GPU) | 36GB | 460GB/s | | 16-inch MacBook Pro, M5 Max (40-core GPU) | 48GB | 614GB/s | | [Mac Studio](https://www.apple.com/mac-studio/specs/), M5 Max | 36GB, configurable to 48GB, 64GB or 128GB | 460GB/s (base) or 614GB/s (40-core GPU) | | Mac Studio, M5 Ultra | 96GB, configurable to 256GB or 512GB | 1.2TB/s | The bandwidth numbers are Apple's, not measurements. Which Mac to buy is on [[which-mac-for-local-llm]]. ## 32GB MacBook Pro vs 16GB VRAM PC: which fits what? Apply the same formula to both and compare against 32 GB and 16 GB. Using the Qwen2.5 Instruct totals from [[how-much-ram-for-local-llm]] (4.0 bits per weight, context 4096, the site's assumed 2 GB overhead, parameter and layer counts from the [14B](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) and [32B](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) cards): | Model | `weights_GB` | KV at 4096 | Overhead | Total | Below 16 GB? | Below 32 GB? | | --- | ---: | ---: | ---: | ---: | --- | --- | | Qwen2.5-14B | 14.7 × 4 / 8 = 7.35 | 0.805 | 2 | 10.155 GB | Yes | Yes | | Qwen2.5-32B | 32.5 × 4 / 8 = 16.25 | 1.074 | 2 | 19.324 GB | No | Yes | "Below" means the total is under the capacity. It is not a promise: the macOS reserve on a 32 GB Mac is **unknown**, and a PC has its own VRAM overhead that this page has not measured. With those caveats, the 32B row is the case where a 32 GB Mac has room on paper and a 16 GB card does not. A 32 GB card or more changes that comparison, and the 14B row fits both. Picks per memory size are on [[best-llm-for-32gb-mac]] and [[best-llm-for-16gb-mac]]. ## Gemma 4 RAM vs VRAM: how much does it need? Unknown on this page. No Gemma 4 model card was read for this page, so no parameter count or layer shape is quoted. To size any model, take the parameter count and `config.json` from its card and apply the formulas on [[methodology]]; the answer is the same number for RAM and VRAM, compared against the pool you have. Longer contexts grow the KV term, see [[long-context-on-mac]]. ## Which is more future proof, unified memory or VRAM? No source read states future-proofing, so this page makes no claim. What the specs do show is that Apple lists unified memory options up to 512GB on the Mac Studio M5 Ultra, and NVIDIA lists 32 GB GDDR7 as the RTX 5090's standard memory. Whether either can be upgraded later is not stated on the pages read, and the model sizes you will want later are **unknown**. ## Questions ### What is the difference between unified memory and VRAM? Unified memory is one pool shared by a Mac's CPU and GPU, while VRAM is a separate pool on a discrete GPU. Apple describes "a large single pool of memory" for the whole chip, and NVIDIA lists VRAM as the card's own memory, for example 32 GB GDDR7 on the RTX 5090. ### Does unified memory work like VRAM? For fitting a model, yes: weights plus KV cache plus overhead must fit in the memory the GPU can reach. The difference is that on a Mac that memory is shared with macOS and your apps, and the share the GPU may use is unknown on this page. ### Can a Mac use all of its unified memory for an LLM? Unknown. Apple's spec pages list installed unified memory, not a GPU-usable remainder, so a total close to the installed figure may still fail to load. See [[mac-gpu-memory-limit]]. ### Is 32GB of unified memory better than 16GB of VRAM? For capacity on paper, 32 GB is larger, and a Qwen2.5-32B 4-bit total of 19.324 GB is under 32 GB and over 16 GB. Capacity is not speed, and speed is unknown here; measure it with [[measure-tokens-per-second-mac]]. ### Gemma 4 RAM vs VRAM: is the requirement different? No, the formula is identical for RAM and VRAM. The Gemma 4 figures themselves are unknown on this page because no model card was read. ### Is unified memory or VRAM more future proof? Not answerable from the sources read. Apple lists up to 512GB of unified memory on the Mac Studio M5 Ultra; future model sizes are unknown. ### Does this page give tokens per second? No. Speed is unknown unless a primary source publishes it for your exact setup, and this site publishes no speed figures. ## Sources read on 2026-10-13 - [Apple MacBook Pro specs](https://www.apple.com/macbook-pro/specs/): unified memory options and bandwidth per M5, M5 Pro and M5 Max configuration. - [Apple Mac Studio specs](https://www.apple.com/mac-studio/specs/): M5 Max and M5 Ultra memory options and bandwidth. - [Apple M5 announcement](https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/): single-pool wording (vendor description) and M5 153GB/s. - [MLX unified memory page](https://ml-explore.github.io/mlx/build/html/usage/unified_memory.html): CPU and GPU share one pool; no data movement between devices. - [MLX issue 3849](https://github.com/ml-explore/mlx/issues/3849): Metal resource-count limit error, separate from bytes free. - [NVIDIA GeForce RTX 5090](https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/): 32 GB GDDR7; no bandwidth stated on the page. - [Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) and [32B](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) cards: parameter counts, as recorded on [[methodology]] (read 2026-10-12). - [[methodology]]: weight, KV and overhead formulas. - Sister site: [LLM VRAM requirements](https://llmhardwareguide.com/vram-requirements), [24GB vs 16GB vs 12GB VRAM](https://llmhardwareguide.com/24gb-vs-16gb-vs-12gb-vram); no numbers taken. ## Related - [[methodology]] - [[how-much-ram-for-local-llm]] - [[what-fits-my-ram]] - [[mac-gpu-memory-limit]] - [[mac-vs-nvidia-for-local-llm]] - [[apple-silicon-memory-bandwidth]] - [[which-mac-for-local-llm]] - [[long-context-on-mac]] - [[best-llm-for-32gb-mac]] - [[best-llm-for-16gb-mac]] - [[map-of-content]]