So M1 Max with 32 Core GPU matches RTX 3080 Mobile.

[JimmyjamesEU]

Again though, not what is happening here. The algorithms and the overall GB score should be scaling linearly across the M1/Pro/Max GPUs. The Max isn't powerful enough to hit the areas where the linear scaling stops.

Given this, I don’t see how we are disagreeing.

My point is, whatever the workload, real world or synthetic, if the benchmark isn’t revealing the full performance of the system, then it‘s an issue. Stating that real world applications also suffer scaling problems, and therefore it’s ok, is not true.

 

[nerdout86]

I would like to see some cross platform tests instead of Metal based. Can't compare it to a 2080 without doing so.

For instance, I find it highly unlikely the M1 GPU is north of mid-high Quadro mobile graphics seen in Dell and ThinkPad mobile workstations.

Right now, on the Windows side. You can't get any north of an A5000 (115W mobile variant).

I am a cross platform user and do have apple products, however aside from video editing and a couple mediocre 3d rendering programs I just do not see where the GPU can be utilized.

I understand this is not on Apple's shoulders but rather software companies [for some odd reason] not seeing the direction of hardware moving forward and failing to put the resources in to bring their software to ARM. With that said, still does not change the fact that right now this is a very handicapped system for anyone outside the creative industry.

Hell, even Excel isn't fully featured in MacOS.

 

[crazy dave]

Given this, I don’t see how we are disagreeing.

My point is, whatever the workload, real world or synthetic, if the benchmark isn’t revealing the full performance of the system, then it‘s an issue. Stating that real world applications also suffer scaling problems, and therefore it’s ok, is not true.

Well eventually the linear scaling stops even on Nvidia cards. That may indeed be reflective of the real workload’s behavior. I don’t know in every case. In which case, that is indeed okay! That’s what you want them to do because that does reveal the full performance of the system *for that workload*.

The issue is that the GB algorithms shouldn’t have stopped linear scaling for the M1 GPUs. The real life algorithms don’t stop scaling there, GB doesn’t stop scaling there either for Nvidia GPUs. So why does it for the M1 Max? For me gotta be thermals: Apple hasn’t reduced max clocks (ie peak performance), but it throttles the GPU aggressively to keep noise and heat down. This is why shorter benchmarks are scaling linearly (a few of the short graphics ones did which is why I was highly confused by that). For the 16” you can turn that throttling off but that’s not enabled yet. This is my guess.

 

[JimmyjamesEU]

Well eventually the linear scaling stops even on Nvidia cards. That may indeed be reflective of the real workload’s behavior. I don’t know in every case. In which case, that is indeed okay! That’s what you want them to do because that does reveal the full performance of the system *for that workload*.

The issue is that the GB algorithms shouldn’t have stopped linear scaling for the M1 GPUs. The real life algorithms don’t stop scaling there, GB doesn’t stop scaling there either for Nvidia GPUs. So why does it for the M1 Max? For me gotta be thermals: Apple hasn’t reduced max clocks (ie peak performance), but it throttles the GPU aggressively to keep noise and heat down. This is why shorter benchmarks are scaling linearly (a few of the short graphics ones did which is why I was highly confused by that). For the 16” you can turn that throttling off but that’s not enabled yet. This is my guess.

Makes sense. It will be interesting to see how the High performance changes results.

 

[EugW]

Well eventually the linear scaling stops even on Nvidia cards. That may indeed be reflective of the real workload’s behavior. I don’t know in every case. In which case, that is indeed okay! That’s what you want them to do because that does reveal the full performance of the system *for that workload*.

The issue is that the GB algorithms shouldn’t have stopped linear scaling for the M1 GPUs. The real life algorithms don’t stop scaling there, GB doesn’t stop scaling there either for Nvidia GPUs. So why does it for the M1 Max? For me gotta be thermals: Apple hasn’t reduced max clocks (ie peak performance), but it throttles the GPU aggressively to keep noise and heat down. This is why shorter benchmarks are scaling linearly (a few of the short graphics ones did which is why I was highly confused by that). For the 16” you can turn that throttling off but that’s not enabled yet. This is my guess.

You'll note that Geekbench CPU scores for the iPad Air 3 are fairly consistent, but the iPhones that use the same A12 are all over the map.

I can get my wife's iPhone XR to score as well as the iPad Air 3, but only can get that consistently if I remove the case, keep it unplugged (so it's not warming up from charging), make sure it's cool at the beginning of the test, and keep it propped up for maximal air flow.

 

[crazy dave]

You'll note that Geekbench CPU scores for the iPad Air 3 are fairly consistent, but the iPhones that use the same A12 are all over the map.

I can get my wife's iPhone XR to score as well as the iPad Air 3, but only can get that consistently if I remove the case, keep it unplugged (so it's not warming up from charging), make sure it's cool at the beginning of the test, and keep it propped up for maximal air flow.

Aye … i think a GPU version of this is going on. A priori I wouldn’t have GB GPU compute would throttle that much, but looking at some of the scores from actual devices, especially laptops, … maybe it is. It would fit all the various disparate benchmark behaviors we’re seeing with the M1 Max and Apple’s TFLOPs claims and explain the “high performance mode” without affecting peak performance. And even the description for the high performance mode seems to be about sustained performance.

I don’t know if it’s right but I like it as an explanation! ?

 

[crazy dave]

Yup! I think we’ve hit a winner! Compare Geekbench in the air to the Pro! It’s throttling. An extreme case but in general laptops may throttle in GB.

https://browser.geekbench.com/v5/compute/search?utf8=✓&q=m1

For some reason it’s not defaulting to a link but you can copy and paste.

Again hypothesis is: peak performance of M1 Max is indeed 4x the M1 as seen in a couple of benchmarks, but Apple wants to keep the 14” and 16” MBP cool and quiet so sustained performance is less. Hence why GB and some others don’t scale like they should. They’re running long enough to test both peak and sustained. For the 16”, it has the thermal capacity for greater sustained performance at the cost of acoustics, so Apple has a “high performance mode” for it. This doesn’t change max clocks (or peak performance) but keeps the laptop running faster linger. But this mode isn’t enabled yet so we’re not seeing it reflected in benchmarks.

@leman, @jmho, others … thoughts?

 

[JimmyjamesEU]

Yup! I think we’ve hit a winner! Compare Geekbench in the air to the Pro! It’s throttling. An extreme case but in general laptops may throttle in GB.

https://browser.geekbench.com/v5/compute/search?utf8=✓&q=m1

For some reason it’s not defaulting to a link but you can copy and paste.

Again hypothesis is: peak performance of M1 Max is indeed 4x the M1 as seen in a couple of benchmarks, but Apple wants to keep the 14” and 16” MBP cool and quiet so sustained performance is less. Hence why GB and some others don’t scale like they should. They’re running long enough to test both peak and sustained. For the 16”, it has the thermal capacity for greater sustained performance at the cost of acoustics, so Apple has a “high performance mode” for it. This doesn’t change max clocks (or peak performance) but keeps the laptop running faster linger. But this mode isn’t enabled yet so we’re not seeing it reflected in benchmarks.

@leman, @jmho, others … thoughts?

Seems plausible and makes the most sense.

 

[diamond.g]

Yup! I think we’ve hit a winner! Compare Geekbench in the air to the Pro! It’s throttling. An extreme case but in general laptops may throttle in GB.

https://browser.geekbench.com/v5/compute/search?utf8=✓&q=m1

For some reason it’s not defaulting to a link but you can copy and paste.

Again hypothesis is: peak performance of M1 Max is indeed 4x the M1 as seen in a couple of benchmarks, but Apple wants to keep the 14” and 16” MBP cool and quiet so sustained performance is less. Hence why GB and some others don’t scale like they should. They’re running long enough to test both peak and sustained. For the 16”, it has the thermal capacity for greater sustained performance at the cost of acoustics, so Apple has a “high performance mode” for it. This doesn’t change max clocks (or peak performance) but keeps the laptop running faster linger. But this mode isn’t enabled yet so we’re not seeing it reflected in benchmarks.

@leman, @jmho, others … thoughts?

GB should only take a minute or two to run if that. Is that long enough to thermal throttle on MacOS?

 

[crazy dave]

GB should only take a minute or two to run if that. Is that long enough to thermal throttle on MacOS?

I wouldn’t have expected this either but it is throttling for the Air. Compare OpenCL and Metal for the Air wrt to 13” Pro. Both scores are lower. Still not by quite enough but Apple may be very aggressive with the throttling. They have a reputation of doing that.

Now granted that’s no fan in Air vs an actual cooling system in the big Pros but I’ve seen other cases where it makes no sense for the two laptops to differ by as much as they do. I think it’s throttling the GPU.

 

[ElfinHilon]

Yup! I think we’ve hit a winner! Compare Geekbench in the air to the Pro! It’s throttling. An extreme case but in general laptops may throttle in GB.

https://browser.geekbench.com/v5/compute/search?utf8=✓&q=m1

For some reason it’s not defaulting to a link but you can copy and paste.

Again hypothesis is: peak performance of M1 Max is indeed 4x the M1 as seen in a couple of benchmarks, but Apple wants to keep the 14” and 16” MBP cool and quiet so sustained performance is less. Hence why GB and some others don’t scale like they should. They’re running long enough to test both peak and sustained. For the 16”, it has the thermal capacity for greater sustained performance at the cost of acoustics, so Apple has a “high performance mode” for it. This doesn’t change max clocks (or peak performance) but keeps the laptop running faster linger. But this mode isn’t enabled yet so we’re not seeing it reflected in benchmarks.

@leman, @jmho, others … thoughts?

Haven't been keeping up with the thread. We think the 16 inch is throttling??

 

[crazy dave]

Haven't been keeping up with the thread. We think the 16 inch is throttling??

*We* may be a bit broad. ? I think it might be.

GB GPU compute can cause a laptop to throttle. However even I have to admit this would be extreme for something like the 16” pro unless Apple is being super aggressive with thermals and acoustics.

 

[ElfinHilon]

*We* may be a bit broad. ? I think it might be.

GB GPU compute can cause a laptop to throttle. However even I have to admit this would be extreme for something like the 16” pro unless Apple is being super aggressive with thermals and acoustics.

Using the royal we here.

 

[dugbug]

So you are confident that will be around 3070, and probably close to the 3080 mobile
Still a big jump no matter how you see it...the biggest jump from 1 generation to the new one in history i think
Did you order something?

Fyi gfxbench shows m1max slightly ahead of the 3080 laptop chip

 

[crazy dave]

Using the royal we here.

Here’s the thing Aztec High only takes a minute to run, stresses the GPU the most out of all the gfxbench tests and shows scaling of 1x M1 -> 2.14x Pro -> 1.87x Max for a total of 4x between M1 and M1 Max. So not completely linear with the pro doing better than expected, but close enough.

GFXBench - Unified cross-platform 3D graphics benchmark database

The first unified cross-platform 3D graphics benchmark database for comparing Android, iOS, Windows 8, Windows Phone 8 and Windows RT capable devices based on graphics processing power.

gfxbench.com

Combine that with the fact that GB scales fine for Nvidia desktop chips around the same performance regime as the M1’s … this is my new pet theory.

 

[nerdout86]

Manhattan looks to be the only cross platform benchmark? I have a Quadro RTX 5000 mobile (P17 w/110W dGPU) so I wanted to run the test and see how is contrasts to the M1 Max.

 

[thunng8]

Manhattan looks to be the only cross platform benchmark? I have a Quadro RTX 5000 mobile (P17 w/110W dGPU) so I wanted to run the test and see how is contrasts to the M1 Max.

All are able to run on the Quadro RTX including Aztec high tier.

Just download here: https://gfxbench.com/result.jsp

 

[jmho]

Geekbench is so quick and given the wattage of the M1 I would hope that throttling wasn't a huge issue unless Apple have set the thermal limits incredibly low. But I guess that's possible as one of the biggest selling points of these machines is that they're near silent.

Another possibility is that Amdahl's law states that the speedup of parallel computing is 1 / ((1 - p) + p/N) where p is the amount of the task that can be parallelized and N is the number of cores.

If you have a task that can be 100% done in parallel, then this equation simplifies to literally just N where adding twice the cores will give twice the performance.

Geekbench's whole thing of using "real world" tasks means that that number is not going to be 100%. If you have a task that is only 95% parallelizable, then even if you somehow had an infinite number of cores, the maximum speedup going from 1 core to infinity cores is 20x (because the equation will simplify to 1 / ((1 - 0.95) + 0)). Amdahl's law is incredibly sensitive to values of p, and the lower this value gets, the worse the scaling becomes. If p fell to 50%, 1 core to infinity cores would only see a 2x boost.

If we assume 95% parallizability and plug that into Amdahl's equation:
M1 with it's 8 core as N = 1 that gives us 1x performance baseline as expected
M1 Pro with 16 as N = 2 that gives us 1.90x, close enough that it seems like it doubled
M1 Max with 32 as N = 4 that gives us 3.48x, which suddenly seems very low

The problem with Geekbench as a metric is that from architecture to architecture and API to API these values of p can vary, which can lead to a massive falloff in the effectiveness of additional cores. There could be an issue with the new architecture or Geekbench's metal port that is causing the M1's p number to be lower than it is for say NVIDIA cards (or even AMD cards using metal in way that is more suited to discrete GPUs).

 

[crazy dave]

Geekbench is so quick and given the wattage of the M1 I would hope that throttling wasn't a huge issue unless Apple have set the thermal limits incredibly low. But I guess that's possible as one of the biggest selling points of these machines is that they're near silent.

Another possibility is that Amdahl's law states that the speedup of parallel computing is 1 / ((1 - p) + p/N) where p is the amount of the task that can be parallelized and N is the number of cores.

If you have a task that can be 100% done in parallel, then this equation simplifies to literally just N where adding twice the cores will give twice the performance.

Geekbench's whole thing of using "real world" tasks means that that number is not going to be 100%. If you have a task that is only 95% parallelizable, then even if you somehow had an infinite number of cores, the maximum speedup going from 1 core to infinity cores is 20x (because the equation will simplify to 1 / ((1 - 0.95) + 0)). Amdahl's law is incredibly sensitive to values of p, and the lower this value gets, the worse the scaling becomes. If p fell to 50%, 1 core to infinity cores would only see a 2x boost.

If we assume 95% parallizability and plug that into Amdahl's equation:
M1 with it's 8 core as N = 1 that gives us 1x performance baseline as expected
M1 Pro with 16 as N = 2 that gives us 1.90x, close enough that it seems like it doubled
M1 Max with 32 as N = 4 that gives us 3.48x, which suddenly seems very low

The problem with Geekbench as a metric is that from architecture to architecture and API to API these values of p can vary, which can lead to a massive falloff in the effectiveness of additional cores. There could be an issue with the new architecture or Geekbench's metal port that is causing the M1's p number to be lower than it is for say NVIDIA cards (or even AMD cards using metal in way that is more suited to discrete GPUs).

It could be … but I don’t feel like p should vary between architectures to such a huge extent for such a wide bevy of algorithms without compromising the ability of Apple to claim 10 TFlops of 32 bit compute. We also know these algorithms intrinsically are capable of scaling at the TFlops these GPUs are running at.

That leaves the API/programming optimization issue but other benchmarks, like Redshift, also didn’t quite see the expected scaling. But then again big Apple GPUs are new, so maybe none of them are optimized yet - we know redshift explicitly says this is a work in progress.

So far the only benchmark that seems to scale linearly (that Ive seen) is Aztec Ruins High. But that shows it can be done. And we’ve seen that GB can throttle the GPU in laptops. But this I have to admit is bigger than I would’ve expected … well bigger even after i realized it wasn’t going to be 0! Which is what I thought at first.

 

[diamond.g]

All are able to run on the Quadro RTX including Aztec high tier.

Just download here: https://gfxbench.com/result.jsp

Aztec I think is the only test that works across all APIs. Vulkan gives the best performance then Metal, DX12, DX11, OpenGL.

 

[ElfinHilon]

It could be … but I don’t feel like p should vary between architectures to such a huge extent for such a wide bevy of algorithms without compromising the ability of Apple to claim 10 TFlops of 32 bit compute. We also know these algorithms intrinsically are capable of scaling at the TFlops these GPUs are running at.

That leaves the API/programming optimization issue but other benchmarks, like Redshift, also didn’t quite see the expected scaling. But then again big Apple GPUs are new, so maybe none of them are optimized yet - we know redshift explicitly says this is a work in progress.

So far the only benchmark that seems to scale linearly (that Ive seen) is Aztec Ruins High. But that shows it can be done. And we’ve seen that GB can throttle the GPU in laptops. But this I have to admit is bigger than I would’ve expected … well bigger even after i realized it wasn’t going to be 0! Which is what I thought at first.

I think we are just seeing something ****y with geekbench. I'd be rather surprised if its thermal throttling in that short of time.

 

[mi7chy]

GB should only take a minute or two to run if that. Is that long enough to thermal throttle on MacOS?

About 30 seconds. Doubtful it's thermal throttling.

 

[crazy dave]

About 30 seconds. Doubtful it's thermal throttling.

GB GPU throttles on the MB air. Granted just passive cooling on the air but I’ve been seeing large variances between laptop GPUs which are supposedly the same or larger variation between laptop GPUs.

I mean I admit I’m surprised too, but this is what I’m seeing.

 

[Boil]

How to fix thermal throttling...?

A larger chassis with a larger heat sink & larger fans...

I am now thinking maybe the best performing Mac, without being crazy stupid expensive, might be the Mac Pro Cube with Dual M1 Max SoCs...?

Mac Pro Cube

• 20-core CPU (16P/4E)
• 64-core GPU
• 32-core Neural Engine
• 128GB LPDDR5 RAM
• 400GB/s memory bandwidth
• 1TB NVMe SSD
• Dual 10Gb Ethernet ports
• (6) USB-C ports
• (4) USB-A ports
• HDMI port
• 3.5mm headphone jack

$4999

• 7.7" x 7.7" x 9.8" including feet
• same dims as G4 Cube (the reason the Mac mini has had the same footprint since forever)...
• Mac mini style PSU up one side of the chassis, just taller & more juice, 420W...?
• PSU is sitting behind the vertically mounted motherboard, all ports on rear panel...
• 2019 Mac Pro style venting top & bottom...
• 180mm fans top & bottom (bottom intake, top exhaust)...
• 2019 Mac Pro style heat sink (with a vapor chamber...?) on SoC, filling remaining inside volume...
• Illuminated Apple logo on front panel...
• Available in any color you want, as long as it's Space Black... ;^p

 

[treehuggerpro]

Take two. Based on the GB metal scores now available, the GFX in the M1 Max MacBooks are most likely down-clocked by about 20% from their full-clock potential and the full-clock GFX in the M1 Pro models, not 25%. Not that the exact numbers really matter in this discussion, but with High Power Mode, or full-clock GFX, in the 16" M1 Max MacBooks, the GB metal scores will probably land around:

M1 Pro (16 core) ~ 42000
M1 Max (32 core) ~ 67200
M1 Max (32 core)+HP ~ 84000

Apple’s likely having some fun with all the scrutiny they typically get, by treating High Power Mode as an easter egg of sorts in the internal testing they’ve provided on their MacBook Pro page. I say that, because there is no mention of High Power Mode in any of the fine print of their current testing details.

A bit of controversy that turns out false, is just going to emphasise the performance-per-watt advantage they will have established in the middle-tier market with the M1 Pro and M1 Max. Just one tier to go . . .