M5 Ultra is the chip Apple built for the people whose work stopped fitting on a laptop several years ago. Announced on Tuesday 25 August 2026 alongside a redesigned Mac Studio, it pairs an up-to-80-core GPU with up to 512GB of unified memory and a media engine Apple says can play 33 simultaneous streams of 8K ProRes 422 at 30 frames per second.
It is also the first Ultra-class Apple silicon with a Neural Accelerator inside every GPU core. That single architectural change is what turns a graphics part into an inference engine, and it is the reason Apple is willing to talk about running AI models with hundreds of billions of parameters on a desktop that sits under a monitor rather than in a rack.
The framing matters. Nobody is going to pre-train a frontier model on this machine, and reinforcement learning runs at scale still belong in a data centre. What the M5 Ultra is aimed at is the other 90% of the work: serving a large model locally, iterating on fine-tunes, grading outputs, and doing it all without shipping proprietary data to somebody else’s GPU cluster.
This article breaks down exactly what Apple disclosed, which numbers are peak-theoretical and which are measured workloads, what 512GB of unified memory actually buys you, whether the 8K video claim survives contact with a real timeline, what the new Mac Studio costs, and the questions Apple pointedly left unanswered.
Table of contents
- What Apple Actually Announced About the M5 Ultra
- The 80-Core GPU and What Neural Accelerators Change
- Why 512GB of Unified Memory Is the Real M5 Ultra Story
- The Media Engine and the 8K Video Claim
- M5 Ultra Versus M5 Max Versus M3 Ultra
- What the New Mac Studio With M5 Ultra Costs
- Who the M5 Ultra Is Actually For
- What Apple Did Not Say About the M5 Ultra
- Frequently Asked Questions About the M5 Ultra
- References
What Apple Actually Announced About the M5 Ultra
Apple published two separate press releases on 25 August 2026: one introducing the M6 and the M5 Ultra as silicon, and one introducing the Mac Studio that carries them. Read together, they are unusually specific about architecture and unusually vague about independent verification.
The headline specification
The M5 Ultra tops out at a 36-core CPU built from 12 super cores and 24 performance cores, an 80-core GPU, and a 32-core Neural Engine. Memory runs to 512GB of unified memory at 1.2TB/s of bandwidth, which Apple states is 50 percent higher than the M3 Ultra it replaces.
Storage moves to PCIe Gen 6, which Apple claims is up to twice as fast as the previous generation. The base Mac Studio configuration built on this chip starts at 96GB of memory and 1TB of storage.
Quad-die, and why that is new
The M5 Ultra is a quad-die part. Apple builds it by fusing two dual-die M5 Max chips using a next-generation version of UltraFusion, its silicon interposer. Inter-die bandwidth exceeds 4.4TB/s, and Apple says the connection density has increased by more than six times.
Every previous Ultra chip was two dies. Four is a first for Apple silicon, and it is the structural reason the core counts double rather than merely increase.
What the executives said
“For the ultimate desktop performance and the ability to run massive AI models, M5 Ultra features a massive GPU, now with Neural Accelerators, and more unified memory bandwidth, pushing the boundaries of what a desktop can do,” said Sri Santhanam, Apple’s vice president of Silicon Engineering Group.
That sentence is the entire product strategy. Note the ordering: the GPU first, memory bandwidth second, and no mention of graphics at all.
| Specification | M5 Ultra as disclosed by Apple |
|---|---|
| CPU | Up to 36 cores (12 super, 24 performance) |
| GPU | Up to 80 cores, Neural Accelerator in each core |
| Neural Engine | 32 cores |
| Unified memory | 96GB standard, up to 512GB |
| Memory bandwidth | Up to 1.2TB/s |
| Packaging | Quad-die, next-generation UltraFusion |
| Inter-die bandwidth | Over 4.4TB/s |
| Media engine | Four ProRes encode and decode engines, hardware AV1 decode |
| Transistor count | Not disclosed |
The 80-Core GPU and What Neural Accelerators Change
The 80-core GPU is the number in every headline, but the core count is the least interesting half of the story. What changed is what sits inside each core.
A matrix unit in every core
Apple has put a Neural Accelerator into each of the 80 GPU cores. These are dedicated matrix-multiply units, which is the operation that dominates transformer inference. Until now, an Ultra chip pushed that work either onto general shader cores or onto the separate Neural Engine.
Doing it inside the GPU matters because the GPU is where the memory bandwidth already is. A neural network running on the M5 Ultra no longer has to choose between compute density and memory access.
4.5x peak, and what that figure is not
Apple’s headline claim is up to 4.5x the peak GPU compute for AI compared with the M3 Ultra, and over six times the M1 Ultra. It also claims up to 40 percent faster graphics performance than the M3 Ultra.
Peak compute is a theoretical ceiling. It tells you what the silicon can do when every unit is fed perfectly, which no real workload achieves. The measured figures Apple published for the Mac Studio are lower and more honest, and they are the ones worth planning against.
The workload numbers Apple published
Against the M3 Ultra, Apple quotes 4x faster large language model prompt processing, 4.3x faster text-to-image generation, 3.3x faster CopyCat machine learning training in Nuke, and 1.7x faster scene rendering in Redshift.
Notice the spread. AI workloads land near the peak multiplier; a traditional renderer lands at 1.7x. If your bottleneck is ray tracing rather than matrix maths, the M5 Ultra is a modest upgrade, not a generational one.
Why 512GB of Unified Memory Is the Real M5 Ultra Story
Strip out the marketing and the most consequential number on the spec sheet is not 80 and not 4.5. It is 512.
Capacity decides what you can run at all
On a discrete graphics card, model weights have to fit in video memory or the model simply does not load. Apple’s unified memory architecture removes that wall: the GPU, the CPU and the Neural Engine all address the same pool. A 512GB M5 Ultra can hold a model that would need a small cluster of conventional accelerators.
This is why Apple keeps describing the machine in terms of parameters rather than teraflops. Capacity is the gate; throughput is what you optimise afterwards.
Rough arithmetic on what fits
At 4-bit quantisation, weights consume roughly half a gigabyte per billion parameters. On that basis a 512GB M5 Ultra has headroom for a model in the several-hundred-billion-parameter range, with room left for context and the operating system. At 8-bit the same machine halves that ceiling.
These are back-of-envelope figures, not Apple’s, and real memory use depends on context length and the runtime. But the order of magnitude is what makes the configuration interesting.
| Model size | Weights at 4-bit, roughly 0.5GB per billion parameters | Against a 512GB M5 Ultra |
|---|---|---|
| 70 billion parameters | About 35GB | Comfortable, even on the 96GB base configuration |
| 235 billion parameters | About 118GB | Comfortable, needs more than the M5 Max ceiling |
| 400 billion parameters | About 200GB | Fits with generous context headroom |
| 700 billion parameters | About 350GB | Fits, with headroom starting to matter |
| 1 trillion parameters | About 500GB | Borderline, weights alone fill the machine |
Bandwidth sets the tokens-per-second ceiling
Capacity decides whether a model loads. Bandwidth decides how fast it talks. At 1.2TB/s the M5 Ultra has 50 percent more memory bandwidth than the M3 Ultra and roughly double the M5 Max’s 614GB/s, which maps almost directly onto generation speed for memory-bound decoding.
The privacy argument nobody puts on a slide
There is a second reason organisations buy this class of hardware, and it has nothing to do with speed. A model that runs locally never sends a prompt anywhere. For legal, clinical and financial teams, that is often the only configuration compliance will approve, and it is a recurring theme across the AI models, tools and releases we track.
The Media Engine and the 8K Video Claim
The 8K half of the announcement is the more measurable one, because video throughput is far harder to overstate than AI performance.
33 streams of 8K ProRes 422
Apple states the M5 Ultra can simultaneously play up to 33 streams of 8K ProRes 422 at 30 frames per second. That is a decode figure, and it is the practical ceiling for a multicam timeline rather than a benchmark score.
For context, a feature-film multicam setup rarely exceeds a dozen angles. Thirty-three is deliberate overkill, which is the point: the M5 Ultra is sized so that the editor never waits for a proxy render.
Four ProRes engines and hardware AV1
The media engine carries hardware-accelerated H.264 and HEVC encode and decode, four ProRes encode and decode engines, and hardware-accelerated AV1 decode. Apple states the M5 Ultra has twice the video encode and decode blocks of the M5 Max.
Apple did not publish an equivalent 8K stream count for the M5 Max, so the honest read is “roughly half” rather than a specific number.
The rest of the pipeline keeps up
The Mac Studio built around the M5 Ultra offers up to six Thunderbolt 5 ports at 120Gb/s, Wi-Fi 7 and Bluetooth 6 through Apple’s N1 networking chip, support for up to eight external displays, and Genlock over USB-C for synchronising with broadcast equipment.
That last item is small and telling. Genlock is not a consumer feature; it exists because Apple expects these machines on live production floors.
| Capability | M5 Ultra | M5 Max |
|---|---|---|
| GPU cores | Up to 80 | Up to 40 |
| CPU cores | Up to 36 (12 super, 24 performance) | 18 (6 super, 12 performance) |
| Maximum unified memory | 512GB | 128GB |
| Memory bandwidth | 1.2TB/s | 614GB/s |
| Video encode/decode blocks | Twice the M5 Max | Baseline |
| Standard memory | 96GB | 36GB |
| Standard storage | 1TB | 512GB |
| Starting price | $5,499 | $2,499 |
M5 Ultra Versus M5 Max Versus M3 Ultra
Choosing between the two new chips is mostly an exercise in working out whether you are memory-limited or compute-limited, and most buyers guess wrong.
The M5 Max is not a consolation prize
The M5 Max is a genuinely fast part. Apple quotes 3.9x faster prompt processing, 3.5x faster text-to-image and 3x faster Magic Mask in DaVinci Resolve against the M4 Max, plus GPU performance up to 50 percent higher than the previous generation.
For anyone whose models fit inside 128GB, the M5 Max delivers most of the experience for less than half the money. The M5 Ultra earns its premium only when 128GB is the thing standing in your way.
What the M3 Ultra comparison quietly omits
Every M5 Ultra number is measured against the M3 Ultra, because there was no M4 Ultra. That makes the multipliers look larger than a single generation would suggest, and it is worth adjusting for when you compare Apple’s figures against a competitor’s year-on-year claims.
Where the CPU lands
The CPU story is the modest one: up to 1.25x higher single-threaded and up to 1.3x higher multithreaded performance than the M3 Ultra. Anyone upgrading for compile times or simulation is buying a 25 to 30 percent improvement, not a 4.5x one.
What the New Mac Studio With M5 Ultra Costs
Pricing is where the positioning becomes unambiguous.
The configurations
The Mac Studio starts at $2,499 with the M5 Max, or $2,299 for education buyers. The M5 Ultra configuration starts at $5,499, or $5,099 for education. Apple also lists financing from $110.10 a month over 36 months on the Ultra tier.
The $3,000 gap between the two starting points is roughly the price of a second complete Mac Studio, which is the comparison every finance director will make.
When you can get one
Pre-orders opened on 25 August 2026 with general availability on 22 September 2026. The 512GB memory configuration slips to late October, which is the configuration most AI teams actually want.
That staggered date is a supply signal. The largest memory bin is the constrained part, and it is also the one that justifies the whole product.
The comparison that actually matters
$5,499 buys a complete, near-silent computer with 96GB of addressable memory for AI work. Nothing in the discrete-accelerator market puts that much memory behind a single chip at that price, and the machines that do are rack-mounted, loud and sold with a support contract attached.
Who the M5 Ultra Is Actually For
Apple has built a narrow product with a wide marketing message. It is worth separating the two.
Teams running models on their own hardware
If you serve a large language model internally, the M5 Ultra is the most memory you can buy per dollar in a desktop form factor. That covers regulated industries, on-premise assistants, and anyone whose data cannot legally leave the building. Teams weighing that decision usually start with an honest assessment of their AI and machine learning readiness before they buy hardware.
Video and post-production
Thirty-three streams of 8K and four ProRes engines describe a colour suite, not a YouTube channel. If your timeline is 8K multicam and your deliverables are ProRes, this is the machine, and the M5 Ultra will spend most of its life waiting for you rather than the reverse.
Everyone who should buy the M5 Max instead
Most developers, most designers and most data teams will never touch 128GB, let alone 512GB. If your working set fits, the M5 Max is the better purchase and the difference funds several years of storage and data platform work.
The people this chip is not for
Gamers, general office users and anyone whose bottleneck is single-threaded software. A 25 percent single-core improvement does not justify a five-thousand-dollar desktop, and generative AI features in consumer apps run perfectly well on far smaller silicon.
What Apple Did Not Say About the M5 Ultra
The gaps in a launch are usually more informative than the claims, and this one has several.
No transistor count and no independent benchmarks
Apple did not publish a transistor count for the M5 Ultra, which is unusual for an Ultra-class part and makes cross-vendor density comparisons impossible. Every performance figure in circulation today is Apple’s own, measured on Apple’s chosen workloads. No third party has run the chip yet.
No Mac Pro
There is no Mac Pro configuration in this announcement. The most powerful chip Apple has ever shipped launched in the mid-range desktop enclosure, which leaves the tower without a current-generation part and its future unclear.
The software question
Hardware is rarely the constraint in local inference; runtime support usually is. Whether these Neural Accelerators are exposed cleanly to the frameworks people actually use, and how quickly popular AI tools adopt them, will decide how much of the 4.5x anyone sees. Apple said nothing about that timeline.
M6, and the odd ordering
Apple introduced the M6 in the same breath: a 12-core CPU on TSMC’s 2nm process, 32GB maximum memory, 170GB/s bandwidth and a 30 percent neural processing uplift. Launching a next-generation base chip alongside a previous-generation Ultra is an unusual sequence, and it suggests the Ultra tier now trails the rest of the family by a full node.
Frequently Asked Questions About the M5 Ultra
Can the M5 Ultra really run large models locally?
Yes, within the limits of its memory. A 512GB configuration can hold models that would otherwise require multiple discrete accelerators, and Apple’s own framing is explicitly about running massive AI models on device.
Is the 80-core GPU the most important number?
No. The 512GB memory ceiling and the 1.2TB/s of bandwidth determine what runs and how fast it responds. The core count matters mainly because each core now contains a Neural Accelerator.
How much faster is the M5 Ultra than the M3 Ultra in practice?
Apple’s measured workloads range from 1.7x for Redshift rendering to 4.3x for text-to-image. The 4.5x figure is peak theoretical GPU compute for AI, not a workload result.
Should I wait for an M6 Ultra?
Apple has not announced one. The M6 shipped as a base chip on a 2nm process, so an Ultra derivative would be some way out, and buying against an unannounced roadmap is rarely a good plan.
Does the M5 Ultra come in a Mac Pro?
Not as announced. It is currently a Mac Studio exclusive, and Apple gave no indication of a Mac Pro configuration.
References
Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute
Apple introduces new Mac Studio with M5 Max and M5 Ultra
Apple’s M5 Ultra with an 80-core GPU will power through your AI models and 8K video
Apple Debuts M5 Ultra as Most Powerful Chip Ever
Apple unveils next-generation Mac Studio with M5 Max and M5 Ultra
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.