Projection of M5 Max & M5 Ultra performance

[Synap]

Happy New Year folks. I was having fun over the holidays trying to project the performance of the M5 Max and M5 Ultra based on the various past performance ratios we have seen from Apple. Except for the AI race, Apple seems to be managing performance improvements on a rather consistent geometric pace.

I based my projections on geekbench.com and geekbench.ai published numbers. Thank you to Geekbench and all the folks that have uploaded their system benchmarks. BTW, don't forget to benchmark your laptop when it is plugged into a power source and otherwise idle for consistency.

I have read that the M5's 16 neural processors are much faster (~4x) than the M4 in a variety of operations. So my logic was to use scale up the M5 performance for the M5 Max and M5 Ultra based on the M4 to Max & Ultra ratios.

My thinking is that Apple will maintain for (i) M5 to M5 Max and (ii) M5 Max to M5 Ultra:

• the memory bandwidth improvement ratios
• the increase in the number of CPU and GPU processors

Below is the result of my work:

					CPU	CPU	CPU	GPU	GPU	GPU	Neural	Neural	Neural
		Single CPU	Multi CPU	Metal	Single Precision	Half Precision	Quantized	Single Precision	Half Precision	Quantized	Single Precision	Half Precision	Quantized
Computation:	Projection:
M5 x M4 Max / M4	M5 Max	4,551	27,375	245,912	6,496	10,448	8,015	36,142	63,798	58,160	6,432	42,280	56,610
M5 x M4 Max / M4 x M3 Ultra / M3 Max	M5 Ultra	4,636	37,431	394,211	7,460	11,831	9,101	38,939	68,456	62,522	7,202	43,048	57,094

Consistent baselines for all inputs:

• All AI benchmarks were using the Core ML
• All MacOS versions were 26.2 or 26.3.

Here is the input data that I used for my computations:

				CPU	CPU		CPU	CPU	CPU	GPU	GPU	GPU	Neural	Neural	Neural
Model	Model ID	CPU Model	O/S	Single CPU	Multi CPU	Metal	Single Precision	Half Precision	Quantized	Single Precision	Half Precision	Quantized	Single Precision	Half Precision	Quantized
Mac Studio (2025)	Mac15,14	M3 Ultra	MacOS 26.2 (Build 25C56)	3,250	28,359	230,364	5,389	8,641	6,782	21,401	23,835	22,262	5,258	30,181	32,784
MacBook Air (13-inch, 2025)	Mac16,12	M4	MacOS 26.2 (Build 25C56)	3,749	15,288	54,891	4,912	7,911	6,294	8,914	10,480	10,305	4,943	36,186	50,716
MacBook Pro (14-inch, Nov 2023)	Mac15,3	M3	MacOS 26.3 (Build 25D5087f)	3,169	11,532	46,347	2,611	4,223	3,259	8,382	10,055	8,636	5,728	37,080	50,210
MacBook Pro (14-inch, 2025)	Mac17,2	M5	MacOS 26.2 (Build 25C56)	4,319	17,871	75,243	5,328	8,699	6,905	13,619	25,484	24,415	5,350	41,599	57,122
MacBook Pro (14-inch, Nov 2023)	Mac15,10	M3 Max	MacOS 26.2 (Build 25C56)	3,190	20,740	143,703	4,693	7,631	5,973	19,864	22,213	20,709	4,696	29,642	32,506
MacBook Pro (16-inch, 2024)	Mac16,6	M4 Max	MacOS 26.2 (Build 25C56)	3,950	23,418	179,397	5,989	9,502	7,306	23,656	26,236	24,548	5,943	36,778	50,261
Mac Studio (2025)	Mac16,9	M4 Max	MacOS 26.2 (Build 25C56)	4,089	26,652	179,397	6,028	9,606	7,449	21,428	24,038	22,365	6,017	36,469	49,202

Hope you find this interesting and potentially useful.

 

[smash1]

For readers that cannot read the charts- change your browser to night mode (black background). The text of the charts is rendered in white text which is not visible if your browser is using a white background.

 

[Synap]

For readers that cannot read the charts- change your browser to night mode (black background). The text of the charts is rendered in white text which is not visible if your browser is using a white background.

Thank you Smash1!

I hate light mode so I didn’t notice that the paste from Numbers was white text and invisible in Light mode. I put my system in Light mode and then edited the text in Pages.

I can now read in both Light and Dark mode.

 

[lapstags]

These seem underwhelming but we will see

 

[Synap]

Predicted:

CPU

CPU

CPU

GPU

GPU

GPU

Neural

Neural

Neural

Single CPU

Multi CPU

Metal

Single Precision

Half Precision

Quantized

Single Precision

Half Precision

Quantized

Single Precision

Half Precision

Quantized

Computation:

Projection:

M5 x M4 Max / M4

M5 Max

4,551

27,375

245,912

Based on early results it looks like the M5 Max scores are:
Single CPU = ~4,300
Multi CPU = ~29,000
Metal = ~230,000

So the M5 Max appears less improved over the M5 than the M4 Max is improved over the M4 except for the Multi CPU performance.

 

[MayaUser]

 

[1d1otic]

M3 Ultra and M4 Max performance is similar to desktop RTX 4060 based on Cyberpunk 2077 max options so we'll see how M5 Max performs.

But I wouldn't expect high performance as long as Apple limited memory bandwidth, TBDR rendering, and limited power consumption.

 

[crazy dave]

M3 Ultra and M4 Max performance is similar to desktop RTX 4060 based on Cyberpunk 2077 max options so we'll see how M5 Max performs.

But I wouldn't expect high performance as long as Apple limited memory bandwidth, TBDR rendering, and limited power consumption.

Apple's memory bandwidth is actually quite good for its compute size and TBDR rendering is a performance multiplier rather than a hinderance for most applications, including games. The main issue for Mac vs PC desktops is the third one you list - these are mobile chips and while Apple's CPU is so good they can do that and still offer competitive desktop performance, the Apple GPU simply doesn't have that level of advantage over its peers. Add in that most games are ports and to get Apple's best GPUs you are tied into also getting its best CPUs and so forth (high base VRAM), then the value proposition for gaming isn't great.

 

[1d1otic]

Apple's memory bandwidth is actually quite good for its compute size and TBDR rendering is a performance multiplier rather than a hinderance for most applications, including games. The main issue for Mac vs PC desktops is the third one you list - these are mobile chips and while Apple's CPU is so good they can do that and still offer competitive desktop performance, the Apple GPU simply doesn't have that level of advantage over its peers. Add in that most games are ports and to get Apple's best GPUs you are tied into also getting its best CPUs and so forth (high base VRAM), then the value proposition for gaming isn't great.

Apple GPU still lacks 1TB/s grade and beyond like 80/90 series and workstation GPU, still using TBDR which isn't great for complexed 3D works, too limited power consumption, Metal optimization, and more.

The hardware itself is still limited so far compared to highend GPU.

 

[crazy dave]

Apple GPU still lacks 1TB/s grade

You have to compare bandwidth size to compute size. Apple's bandwidth is easily commensurate with the size of their GPUs.

and beyond like 80/90 series and workstation GPU,

Apple's high end Ultra GPUs are great at workstation tasks due to their high VRAM capacity which consumer Nvidia chips lack. Unfortunately this rarely comes up in standard benchmarks since standard benchmarks want to run on consumer hardware and simply having it not run isn't useful for most people. That said, Apple GPUs do top out at smaller core counts with lower clock speeds. They are still performant regardless, they just can't quite close the gap to their desktop Nvidia counterparts in all but a few applications.

still using TBDR which isn't great for complexed 3D works,

That's the opposite: TBDR is actually better for complex 3D workflows, it's simple 3D scenes that you can't get any benefit out of it - the fewer things on screen that are occluding each other, the less benefit you get.

too limited power consumption,

Yes, for desktops

Metal optimization,

No idea what this means, that most apps aren't optimized for Metal? If so, true.

and more.

The hardware itself is still limited so far compared to highend GPU.

Citation needed.

 

[1d1otic]

You have to compare bandwidth size to compute size. Apple's bandwidth is easily commensurate with the size of their GPUs.

Tell that to GPUs with faster than 1TB/s of bandwidth.

Apple's high end Ultra GPUs are great at workstation tasks due to their high VRAM capacity which consumer Nvidia chips lack. Unfortunately this rarely comes up in standard benchmarks since standard benchmarks want to run on consumer hardware and simply having it not run isn't useful for most people. That said, Apple GPUs do top out at smaller sizes with lower clock speeds.

VRAM is just a size of the memory, not the GPU performance itself. It only allows to work with big projects, nothing more. Even M3 Ultra and M4 Max are only as good as RTX 4060, that's all. Where is 5090 or workstation grade GPU?

That's the opposite: TBDR is actually better for complex 3D workflows, it's simple 3D scenes that you can't get any benefit out of it - the fewer things on screen that are occluding each other, the less benefit you get.

Wrong, TBDR is mobile based rendering because it can not handle complex and high performance. That's where IMR is needed and Nvidia has proven faster.

No idea what this means

CUDA is a standard while Metal is Apple's proprietary standard while only few are taking advantages. Besides, who even use Mac for 3D works?.

 

[crazy dave]

Tell that to GPUs with faster than 1TB/s of bandwidth.

Those GPUs are also bigger ... hence the bigger bandwidth ...

VRAM is just a size of the memory, not the GPU performance itself. It only allows to work with big projects, nothing more. Even M3 Ultra and M4 Max are only as good as RTX 4060, that's all. Where is 5090 or workstation grade GPU?

That's part of what makes a good workstation GPU ... if you can't work on big projects ... not a lot of point to having a 5090 for workstation tasks is there?

Apple's workstation GPU is the Ultra.

Wrong, TBDR is mobile based rendering because it can not handle complex and high performance. That's where IMR is needed and Nvidia has proven faster.

That's simply not true. Do you know what TBDR actually is?

CUDA is a standard while Metal is Apple's proprietary standard while only few are taking advantages. Besides, who even use Mac for 3D works?

If so, then why are you bothering to compare them?

 

[1d1otic]

Those GPUs are also bigger ... hence the bigger bandwidth ...

Then why do you bother to compare? You just admitted the problem then.

That's part of what makes a good workstation GPU ... if you can't work on big projects ... not a lot of point to having a 5090 for workstation tasks is there?

Apple's workstation GPU is the Ultra.

And Apple is the one who ditched workstation itself, Mac Pro. Mac Studio is far from being a workstation.

That's simply not true. Do you know what TBDR actually is?

TBDR has been used since PowerVR era and the main reason of using TBDR on mobile is due to the power consumption and limited GPU performance as a mobile device. Even Mac was using IMR until they got Apple Silicon.

If so, then why are you bothering to compare them?

You are not realizing the limitation after all.