New Apple MacBook M5 has tensor cores. LC0 will be a lot faster :-)

[Hai]

Every GPU core has also a tensor core.

[smatovic]

GPU

The M5 introduces a next-generation 10-core GPU architecture featuring a dedicated Neural Accelerator integrated into each core.[4] This architecture delivers over 4x the peak GPU compute performance for AI compared to M4, and over 6x peak GPU compute for AI performance compared to M1.

The GPU includes enhanced shader cores and Apple's third-generation ray-tracing engine, providing up to 45 percent graphics uplift in applications using ray tracing compared to M4. Overall graphics performance is up to 30 percent faster than M4 and up to 2.5x faster than M1. The GPU also features rearchitected second-generation dynamic caching.

The Neural Accelerators in each GPU core can be programmed directly by developers using Tensor APIs in Metal 4. The GPU architecture is designed for seamless integration with Apple's software frameworks, including Core ML, Metal Performance Shaders, and Metal 4.

Neural Engine

The M5 includes an improved 16-core Neural Engine that works alongside the Neural Accelerators in the GPU cores to optimize the chip for AI workloads.[5]

https://en.wikipedia.org/wiki/Apple_M5

--
Srdja

[Werewolf]

This will be trounced by Nvidia.

As an experiment I decided to try Lc0 training on a M2 Ultra and it was 1/7 the speed of a 5090 card. Even an eventual M5 Ultra will still be a long way behind

[towforce]

This will be trounced by Nvidia.

As an experiment I decided to try Lc0 training on a M2 Ultra and it was 1/7 the speed of a 5090 card. Even an eventual M5 Ultra will still be a long way behind

Presumably the M5 is good for portable devices, whereas anything containing a 5090 is not going to convenient for carrying. Why compare?

[Werewolf]

This will be trounced by Nvidia.

As an experiment I decided to try Lc0 training on a M2 Ultra and it was 1/7 the speed of a 5090 card. Even an eventual M5 Ultra will still be a long way behind

Presumably the M5 is good for portable devices, whereas anything containing a 5090 is not going to convenient for carrying. Why compare?

The M2/M5 Ultra won’t be portable.
The M5 is portable but so are much faster Nvidia options.

[syzygy]

This will be trounced by Nvidia.

As an experiment I decided to try Lc0 training on a M2 Ultra and it was 1/7 the speed of a 5090 card. Even an eventual M5 Ultra will still be a long way behind

Presumably the M5 is good for portable devices, whereas anything containing a 5090 is not going to convenient for carrying. Why compare?

The M2/M5 Ultra won’t be portable.
The M5 is portable but so are much faster Nvidia options.

This is about the M5, a SoC which will power iPads. A comparison with a 5090 is hardly appropriate. Of course there is more powerful hardware around than you can find in a tablet.

[Werewolf]

This will be trounced by Nvidia.

As an experiment I decided to try Lc0 training on a M2 Ultra and it was 1/7 the speed of a 5090 card. Even an eventual M5 Ultra will still be a long way behind

Presumably the M5 is good for portable devices, whereas anything containing a 5090 is not going to convenient for carrying. Why compare?

The M2/M5 Ultra won’t be portable.
The M5 is portable but so are much faster Nvidia options.

This is about the M5, a SoC which will power iPads. A comparison with a 5090 is hardly appropriate. Of course there is more powerful hardware around than you can find in a tablet.

So you quoted me but didn't read the words? The M5 was not being compared with the 5090. I was comparing two desktop solutions and was surprised that for all its hype Apple is slower than I imagined. Even my LAPTOP 3050 card beat the M2 Ultra DESKTOP processor, despite being 2 years older.

M5 is new, it's in the iPad but also the MacBook Pro...
Anything tensor-like is welcome, but even when the heavier M5 pro / max / ultras (desktop) come out I think they'll be way behind Nvidia for Lc0.

On my net I got:
M2 Ultra: 30 games / hour (after much optimisation, started at 10 games/hour)
3050 card in old laptop: 50 g/h
desktop cards obviously much higher

[syzygy]

So you quoted me but didn't read the words? The M5 was not being compared with the 5090.

Then start another thread. This is about M5, which you say will be trounced and then you start about the 5090.

[Werewolf]

So you quoted me but didn't read the words? The M5 was not being compared with the 5090.

Then start another thread. This is about M5, which you say will be trounced and then you start about the 5090.

I think you're missing my point. Apple themselves are making performance comparisons to their previous chips - claiming "4x the peak GPU compute performance for AI compared to M4" and "6x compared to M1." They're setting expectations with these numbers.
So it's entirely reasonable to test whether Apple silicon lives up to the hype for actual AI workloads like LC0. My M2 Ultra experience is a data point showing that despite impressive-sounding specs, real-world AI performance has been disappointing compared to alternatives.

The thread title says "LC0 will be a lot faster" - my testing suggests we should temper expectations.

I'm not saying the M5 won't be an improvement - clearly it will be. I'm saying that for serious LC0 training or high-performance inference, Nvidia is still the way to go, and the gap is larger than many people realize.

[Eelco de Groot]

For M5 Ultra you only can get AI projections so nobody knows. 190 W or something. The M5 is much lighter:

If there’s a downside to the M5 in our testing, it’s that its performance improvements seem to come with increased power draw relative to the M4 when all the CPU cores are engaged in heavy lifting. According to macOS built-in powermetrics tool, the M5 drew an average 28 W of power in our Handbrake video encoding test, compared to around 17 W for the M4 running the same test.

Source is quote by the AI from https://arstechnica.com/gadgets/2025/10 ... r-wrapper/

But if you ask about the 5090 they say the TDP is 575 W and under load then it uses a bit more. Apart from things like memory bandwidth (says the above article too), GPU is parallel performance so you can just add more processors. (Like in a CRAY the first example I know of of tensor processing?) I know that is a gross oversimplification but that seems where most of the performance increase will come from, now, any other choice like increasing clockspeed will have too much impact on heat production (and power consumption) which you do not want if such a GPU card (which you can not add to an Apple Macbook or Mac desktop anymore, important distinction) has just a single chip with GPU (I don't know, I am just guessing that) with memory banks probably on the rest of the card circuit board. For professional uses AI training, supercomputing radio telescope data processing and such power consumption in a supercomputer or in a data center is extremely important. In a laptop or tablet portable etc too. Adding more processors but with low clockspeed is then logical and you can cool with less power. But this is my 1990s knowledge like about Cray technology reading from Bob here in the forum not really up to date with the facts sorry