Unified memory: from Strix Halo to Mac Studio M5 Ultra — what fits and at what speed

Have you ever tried to load a model that's too big for your GPU? The program tells you there isn't enough memory, and that's where it stops. With a discrete GPU, the ceiling is fixed: the VRAM is either there or it isn't. Unified memory changes the question. CPU and GPU share a single pool of RAM, so the model can use nearly all of the machine's memory, not an isolated slice of a few GB.

The point, though, isn't only the "how much". The real bottleneck of a local AI isn't memory capacity: it's memory bandwidth, i.e. the speed at which the machine reads the model's weights. And here the numbers matter more than they look: 256 GB/s, 614 GB/s or 1.2 TB/s are not three technical specs, they are three completely different user experiences.

In this article we take three real platforms — AMD's Strix Halo, the Macs with M5 Max, and the Macs with M5 Ultra — and do the math: what fits, at what speed, at what cost, and for whom it makes sense. But first, two quick concepts.

What unified memory actually is

On a PC with a discrete GPU there are two separate memories: system RAM for the CPU and VRAM for the GPU, connected by a PCIe bus. If the model doesn't fit in VRAM, you reduce the quantization or buy a bigger GPU. On a unified-memory platform — Apple Silicon, Strix Halo, DGX Spark — there is a single pool, and the GPU reads from it directly. The allocation is tuned in the BIOS or the operating system: on the Strix Halo, up to 96GB of the 128GB total can be assigned to the GPU.

This means a model that is a luxury on a 16GB discrete card becomes ordinary on a machine with 128GB of unified memory. "What fits" stops being a limit and becomes a choice.

Why memory bandwidth determines the speed

When you generate text, the model is in the decode phase: it produces one token at a time, and for each token it re-reads the active weights from memory. Tokens per second therefore depend, as a first approximation, on the bandwidth divided by the bytes the model occupies. Simple arithmetic, with strong consequences.

A concrete example. A 27B-parameter model in Q4_K_M weighs about 16.4GB. A Strix Halo at 256 GB/s produces roughly 11-13 tokens per second: usable, but not fast. An M5 Max at 614 GB/s delivers 28-30. An M5 Ultra at 1.2 TB/s reaches 50-80. Same model, same quantization: only the memory changes the outcome.

The prefill — the phase where the model reads the whole prompt — works differently: there, raw compute matters most, and a chip with lower bandwidth but more cores can still process prompts at high speed. Bandwidth is what limits the phase you feel most, which is the text appearing.

The three platforms: 256, 614 and 1.200 GB/s

Platform	Bandwidth	Unified memory
AMD Ryzen AI Max+ 395 (Strix Halo)	256 GB/s	128GB
NVIDIA DGX Spark (GB10)	273 GB/s	128GB
Apple M5 Max (Mac Studio)	614 GB/s	up to 128GB
Apple M5 Ultra (Mac Studio)	1.2 TB/s	up to 512GB

The Strix Halo is the cheapest of the band and the most "consumer": 256 GB/s is roughly a quarter of an RTX 4090, so anything that fits in a discrete card's VRAM runs faster on that card. But it wins on capacity: it is the only one on this list, outside the Apple world, to run a 100B MoE locally without a discrete GPU, on a mini-PC that fits on a shelf.

The DGX Spark sits in between: 273 GB/s like the Strix Halo, but with full CUDA and about 1 petaFLOP in FP4. For developers, it's the "unified memory without giving up the NVIDIA ecosystem" option.

Apple's jump is the most marked of the generation: from the M5 Max's 614 GB/s to the M5 Ultra's 1.2 TB/s, with the bandwidth doubled versus the previous generation (the M3 Ultra was at 800 GB/s). The M5 Ultra is also the first quad-die chip: two M5 Max joined by UltraFusion, with up to 36 CPU cores and 80 GPU cores.

128GB vs 512GB: what actually fits

With 128GB of unified memory (and 96GB allocable to the GPU on the Strix Halo), in Q4_K_M:

• a 27B in high quantization, with generous context — there's room even for Q8;
• a 70B dense in Q4, with moderate context: the weights take about 40GB, the rest is KV cache;
• a 100B MoE in 4-bit, with wide context: this is where the MoE shines, because total parameters are 100B but active ones are few, and speed depends on the active count, not the total.

With 512GB on the M5 Ultra, the story changes:

• frontier-class MoE with tens or hundreds of billions of total parameters in 4 or 8-bit: models that until yesterday only ran in the cloud;
• RAG over large corpora: tens of thousands of local documents, with the embedding store and the context entirely in memory, no external service;
• a 27B with extended context (128K-262K tokens): the KV cache alone can weigh tens of GB, and at 512GB that's not a problem.

The practical rule: 128GB if your models stay under 120B in 4-bit and you use contexts up to 128K. 512GB if you want the biggest models locally, if you run RAG over a large corpus, or if long context is your main use case.

How much it costs, and for whom it makes sense

Current street prices: 128GB Strix Halo mini-PCs start around 1,500-1,800 euros; the Mac Studio with M5 Max and 128GB sits near 2,999 euros; the Mac Studio with M5 Ultra and 512GB is 6,699 euros; the DGX Spark, after the February 2026 price revision, costs about 4,800 euros.

But the comparison that counts isn't the purchase price: it's the cost per token. The cloud charges you for every token you generate, forever. Hardware is paid once, and from there the cost per token is nearly zero (electricity). A 128GB machine at 2,000 euros pays for itself in months against an API subscription for heavy use; a 512GB machine at 6,700 euros only makes sense if it actually replaces cloud usage that was expensive.

For whom it makes sense, in summary:

• 128GB, the 1,500-3,000 euro band — those who want an always-available AI at home or in the office, for chat, RAG over their own documents, agentic assistants. Typical model: a 27B or a 100B MoE.
• 512GB, the 6,500 euro band — those who work with frontier-class models locally, RAG over large corpora, or very long contexts. It's the choice for people who know exactly what they're putting inside.

One practical warning: the 512GB configurations of the Mac Studio M5 Ultra arrive at the end of October 2026; the M5 Max Macs have been available since late September. Strix Halo mini-PCs are already on the shelves, in various formats, from desktop mini-PCs to laptops.

Conclusion

Unified memory moved the question from hardware to the model. You no longer ask "how much VRAM do I need for this model?"; you ask "how many billions of parameters can I afford on my machine?". The answer depends on two numbers: capacity, which decides what fits; bandwidth, which decides the speed.

256 GB/s is a daily-use AI at a calm pace. 614 GB/s is a working AI. 1.2 TB/s is a frontier AI. And the difference between 128GB and 512GB is not a tier jump: it's the difference between running the models that ship this year and the ones that will ship in the next two.

When you buy hardware for local AI, look first at how much memory you need for your models, then at which bandwidth makes the experience comfortable, and finally at the total cost over time — not just the purchase price, because the cloud is never free, it's just billed. The memory you own. The rest you rent.

References

• Apple, "Apple introduces new Mac Studio with M5 Max and M5 Ultra" (official press release): https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
• Apple, "Mac Studio (M5 Max or M5 Ultra) — Tech Specs": https://support.apple.com/en-us/128107
• AMD, "Ryzen AI Halo Developer Platform with Ryzen AI Max+ 395" (official specifications): https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo/ryzen-ai-max-plus-395.html
• AMD, "Ryzen AI MAX+ 395: Breakthrough AI Performance in Thin and Light" (VRAM allocation up to 96GB, 256 GB/s): https://www.amd.com/en/blogs/2025/amd-ryzen-ai-max-395-processor-breakthrough-ai-.html

Article generated with AI