Discussion M5 series GPU discussion

[poke01]

Apple finally made a decent GPU architecture for once.
So I made this thread to post proper benchmarks related to the M5 GPU.

Will be used for LLMs tests, gaming and rendering.

 

[marokra]

For those of us who don't pay deep attention to GPU architectures, what sets M5 GPU apart from earlier ones? What did the earlier GPUs lack?

 

[poke01]

For those of us who don't pay deep attention to GPU architectures, what sets M5 GPU apart from earlier ones? What did the earlier GPUs lack?

The M5 GPU architecture “Apple10” is a much more area performant architecture and each GPU core is properly utilised leading to huge uplifts in rasterisation and in ray traced workloads even more than M3 -> M4 which introduced RT cores. It also includes tensor cores useful for prefil and Topez etc.

The performance per watt is also very good. Previous architectures from Apple lacked good PPA and had weak performance for the price you paid relative to dGPUs from Nvidia. The Pro and Max variants should scale well and offer a better overall package now.

I could go further into detail but it’s better to wait till next week where we can see actual tests.

 

[marokra]

Thanks, you've given me much to google

 

[mvprod123]

Dynamic caching is an impressive tech. It is surprising that the industry has not paid much attention to it.

 

[poke01]

Dynamic caching is an impressive tech. It is surprising that the industry has not paid much attention to it.

amd has something simliar coming

 

[mvprod123]

amd has something simliar coming

Where can I read about this?

 

[mvprod123]

In 2022, Apple hired a leading engineer from Nvidia, who now heads the development of Apple's GPU architectures. We are now seeing the results of that work.

 

[poke01]

Where can I read about this?

I think it was talked about in the RDNA5 thread couple weeks back

 

[mvprod123]

I think it was talked about in the RDNA5 thread couple weeks back

So, is this just rumours and speculation, or are there already patents?

 

[poke01]

So, is this just rumours and speculation, or are there already patents?

Post in thread 'RDNA 5 / UDNA (CDNA Next) speculation'
https://forums.anandtech.com/threads/rdna-5-udna-cdna-next-speculation.2624468/post-41497631

It looks like speculation

 

[Mopetar]

I haven't had time to do much more than look at the highlights for new new M5 MBP from threads here, but I was intrigued by the reported uplifts in graphics improvements.

The posts here have certainly helped to explain why it's such a large jump despite the paper specs not suggesting as much of an uplift.

 

[poke01]

I haven't had time to do much more than look at the highlights for new new M5 MBP from threads here, but I was intrigued by the reported uplifts in graphics improvements.

The posts here have certainly helped to explain why it's such a large jump despite the paper specs not suggesting as much of an uplift.

yea it’s smart design. Increasing performance without increasing GPU area/adding more cores so the cost doesn’t balloon.

My expectation is that the 10 core M5 GPU should be on par with the 14 core M3 Pro in GPU bound tasks.

 

[poke01]

Investigating the GPU Neural Accelerators on Apple A19/M5

tzakharko.github.io

This is a break down of the new GPU core in M5/A19 by “leman”.

 

[mvprod123]

Blender GPU score (62% faster than M4, on par with M3 Pro (18-core GPU), faster than 3050)

 

[mvprod123]

Blender GPU score (62% faster than M4, on par with M3 Pro (18-core GPU), faster than 3050)
View attachment 132197

If this increase extends to the entire M5 chip line, there is a chance that the M5 Max could match or even surpass the 5090 laptop.

 

[regen1]

Blender GPU score (62% faster than M4, on par with M3 Pro (18-core GPU), faster than 3050)
View attachment 132197

That's still only a single sample but that score is ahead a 9060XT and below B570 for Blender 4.5.

 

[mvprod123]

I wonder if Apple's GPU architecture roadmap is consistent with what Imagination is doing with its PowerVR GPU architecture? By a strange coincidence, in May of this year, Imagination presented a new architecture that offers a 35% improvement in energy efficiency (which almost matches the energy efficiency curves in Geekerwan's A19 Pro review) and has a built-in Neural Core that provides 4 times more performance in AI computing, which also matches Apple's Neural Accelerator.

E-Series GPU - Imagination

E-Series GPU IP: fast, flexible parallel acceleration for graphics and AI workloads across automotive, mobile, consumer and desktop markets.

www.imaginationtech.com

"Imagination’s PowerVR GPU architecture is renowned for its energy efficiency and has been deployed in power-constrained devices for nearly twenty years. The E-Series’ new Burst Processors technology enhances power efficiency by a further 35% for AI workloads, games and user interfaces. This improvement is achieved by reducing pipeline depth and minimising data movement within the GPU."
"E-Series continues to offer the advanced graphics capabilities of previous generations of Imagination GPUs, including support for ray tracing. To this, it adds deeply integrated acceleration for power-efficient, low precision AI operations into every GPU core. This creates the compute-dense E-Series Neural Cores that scale up to 200 TOPS INT8 and unleash up to 400% the AI performance of the previous D-Series. The Neural Cores support a wide range of popular AI number formats, enabling developers to design networks that meet a broad spectrum of performance, accuracy and power demands. One of their many performance efficiency measures is an AI-friendly memory architecture that prioritises local memory for compute, greatly reducing the power and performance costs of going to external memory."

 

[soresu]

I wonder if Apple's GPU architecture roadmap is consistent with what Imagination is doing with its PowerVR GPU architecture? By a strange coincidence, in May of this year, Imagination presented a new architecture that offers a 35% improvement in energy efficiency (which almost matches the energy efficiency curves in Geekerwan's A19 Pro review) and has a built-in Neural Core that provides 4 times more performance in AI computing, which also matches Apple's Neural Accelerator.

E-Series GPU - Imagination

E-Series GPU IP: fast, flexible parallel acceleration for graphics and AI workloads across automotive, mobile, consumer and desktop markets.

www.imaginationtech.com

"Imagination’s PowerVR GPU architecture is renowned for its energy efficiency and has been deployed in power-constrained devices for nearly twenty years. The E-Series’ new Burst Processors technology enhances power efficiency by a further 35% for AI workloads, games and user interfaces. This improvement is achieved by reducing pipeline depth and minimising data movement within the GPU."
"E-Series continues to offer the advanced graphics capabilities of previous generations of Imagination GPUs, including support for ray tracing. To this, it adds deeply integrated acceleration for power-efficient, low precision AI operations into every GPU core. This creates the compute-dense E-Series Neural Cores that scale up to 200 TOPS INT8 and unleash up to 400% the AI performance of the previous D-Series. The Neural Cores support a wide range of popular AI number formats, enabling developers to design networks that meet a broad spectrum of performance, accuracy and power demands. One of their many performance efficiency measures is an AI-friendly memory architecture that prioritises local memory for compute, greatly reducing the power and performance costs of going to external memory."
View attachment 132315

OSS driver efforts of PowerVR have also picked up considerably.

The VK driver is still about 62 extensions behind PanVK (Mali VK), but they at least put it over the v1.2 compliance mark a week ago, a significant improvement from not even being v1.0 conformant for a very long time.

I'd say that both are still waaaaaaaaaaaaaaaaaaaaaaaayyyyy behind RADV in general levels of optimisation though, and they also lack a dedicated vendor specific shader compiler like RADV, NVK and ANV all have.