June 2026 LRC/PS updates ... faster Denoise measurements for Mac M with Neural Engine
Here’s the updated table and a couple of graphs. (I will redirect the links to post #9 and just update this page from now on).
In the past the denoise time has correlated well with # of GPU cores. I had used data provided by Artisright some time ago and I may do that again if he updates it. There seemed no point in plotting vs Neural Engine cores as, except for Ultra chips, the core count is the same. So I stuck with GPU and got a reasonably good correlation (1st graph below), so it appears that the NE is accelerating the GPU and more (GPU) is better, no shock. It makes sense that there are more NE cores for the Ultra, to ensure full acceleration. I guess that 16 is enough for all other configs and simplifies chip design.
While the correlation is OK I wanted to take into account that each generation, M1-> M2 etc gets faster, either by design or by process shrink (which speeds up transistors). I could have tried scouring the web to get speed tests for each device, GPU or CPU as a stand-in, but I took the easy way out and just used a 10% multiplier for each new rev. So M2 runs 10% faster than M1 (x 1.10), M3 would run 21% faster than M1 (1.10 x 1.10) etc. Rather than divide these values into the denoise time I multiplied them by the GPU core count and came up with what I called Virtual Cores on the horizontal axis (2nd graph). As you can see it tightens up the correlation a little though there are a couple of outliers.
So how useful is this? Not very .. just playing with numbers here. The table might be of some use if you plan to upgrade and AI/Denoise is important to your PP workflow/time. Set your pain threshold for Denoise time then look at configurations (GPU core count especially) that have better results. Discard older revs if you don’t want to buy used.
PS. I used M5 as the chip rev (and multiplier) for the device in the Neo. I read that is it on the 3nm process at TSMC.
June 2026 LRC/PS updates ... faster Denoise measurements for Mac M with Neural Engine
Here’s the updated table and a couple of graphs. (I will redirect the links to post #9 and just update this page from now on).
In the past the denoise time has correlated well with # of GPU cores. I had used data provided by Artisright some time ago and I may do that again if he updates it. There seemed no point in plotting vs Neural Engine cores as, except for Ultra chips, the core count is the same. So I stuck with GPU and got a reasonably good correlation (1st graph below), so it appears that the NE is accelerating the GPU and more (GPU) is better, no shock. It makes sense that there are more NE cores for the Ultra, to ensure full acceleration. I guess that 16 is enough for all other configs and simplifies chip design.
While the correlation is OK I wanted to take into account that each generation, M1-> M2 etc gets faster, either by design or by process shrink (which speeds up transistors). I could have tried scouring the web to get speed tests for each device, GPU or CPU as a stand-in, but I took the easy way out and just used a 10% multiplier for each new rev. So M2 runs 10% faster than M1 (x 1.10), M3 would run 21% faster than M1 (1.10 x 1.10) etc. Rather than divide these values into the denoise time I multiplied them by the GPU core count and came up with what I called Virtual Cores on the horizontal axis (2nd graph). As you can see it tightens up the correlation a little though there are a couple of outliers.
So how useful is this? Not very .. just playing with numbers here. The table might be of some use if you plan to upgrade and AI/Denoise is important to your PP workflow/time. Set your pain threshold for Denoise time then look at configurations (GPU core count especially) that have better results. Discard older revs if you don’t want to buy used.
PS. I used M5 as the chip rev (and multiplier) for the device in the Neo. I read that is it on the 3nm process at TSMC.
June 2026 LRC/PS updates ... faster Denoise measurements for Mac M with Neural Engine
Here’s the updated table and a couple of graphs. (I will redirect the links to post #9 and just update this page from now on).
In the past the denoise time has correlated well with # of GPU cores. I had used data provided by Artisright some time ago and I may do that again if he updates it. There seemed no point in plotting vs Neural Engine cores as, except for Ultra chips, the core count is the same. So I stuck with GPU and got a reasonably good correlation (1st graph below), so it appears that the NE is accelerating the GPU and more (GPU) is better, no shock. It makes sense that there are more NE cores for the Ultra, to ensure full acceleration. I guess that 16 is enough for all other configs and simplifies chip design.
While the correlation is OK I wanted to take into account that each generation, M1-> M2 etc gets faster, either by design or by process shrink (which speeds up transistors). I could have tried scouring the web to get speed tests for each device, GPU or CPU as a stand-in, but I took the easy way out and just used a 10% multiplier for each new rev. So M2 runs 10% faster than M1 (x 1.10), M3 would run 21% faster than M1 (1.10 x 1.10) etc. Rather than divide these values into the denoise time I multiplied them by the GPU core count and came up with what I called Virtual Cores on the horizontal axis (2nd graph). As you can see it tightens up the correlation a little though there are a couple of outliers.
So how useful is this? Not very .. just playing with numbers here. The table might be of some use if you plan to upgrade and AI/Denoise is important to your PP workflow/time. Set your pain threshold for Denoise time then look at configurations (GPU core count especially) that have better results. Discard older revs if you don’t want to buy used.
PS. I used M5 as the chip rev (and multiplier) for the device in the Neo. I read that is it on the 3nm process at TSMC.
June 2026 LRC/PS updates ... faster Denoise measurements for Mac M with Neural Engine
Here’s the updated table and a couple of graphs. (I will redirect the links to post #9 and just update this page from now on).
In the past the denoise time has correlated well with # of GPU cores. I had used data provided by Artisright some time ago and I may do that again if he updates it. There seemed no point in plotting vs Neural Engine cores as, except for Ultra chips, the core count is the same. So I stuck with GPU and got a reasonably good correlation (1st graph below), so it appears that the NE is accelerating the GPU and more (GPU) is better, no shock. It makes sense that there are more NE cores for the Ultra, to ensure full acceleration. I guess that 16 is enough for all other configs and simplifies chip design.
While the correlation is OK I wanted to take into account that each generation, M1-> M2 etc gets faster, either by design or by process shrink (which speeds up transistors). I could have tried scouring the web to get speed tests for each device, GPU or CPU as a stand-in, but I took the easy way out and just used a 10% multiplier for each new rev. So M2 runs 10% faster than M1 (x 1.10), M3 would run 21% faster than M1 (1.10 x 1.10) etc. Rather than divide these values into the denoise time I multiplied them by the GPU core count and came up with what I called Virtual Cores on the horizontal axis (2nd graph). As you can see it tightens up the correlation a little though there are a couple of outliers.
So how useful is this? Not very .. just playing with numbers here. The table might be of some use if you plan to upgrade and AI/Denoise is important to your PP workflow/time. Set your pain threshold for Denoise time then look at configurations (GPU core count especially) that have better results. Discard older revs if you don’t want to buy used.
PS. I used M5 as the chip rev (and multiplier) for the device in the Neo. I read that is it on the 3nm process at TSMC.
