That’s fair. I also am interested to see the result with High Power mode.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
So M1 Max with 32 Core GPU matches RTX 3080 Mobile.
- Thread starter iBug2
- Start date
- Sort by reaction score
no i doesnt..apple never posted any graphs with geekbench results or even alluded to geekbench in any way. for some reason in this thread its the end all be all..its a synthetic benchmark!
Well Apple claimed a 4x improvement from m1 to Max. It scales from 8 to 16 perfectly and then much less from 16 to 32. That is strange. So much so that we’ll respectEd and knowledgeable people have questioned the results.
Do we think it is a thermal throttle that the high performance mode unlocks? Seems weird, cause GB doesn't run long enough to thermally throttle (at least not that I have seen).
If you actually read Apple's test description for individual GPU tests in their press release, the biggest scaling they themselves show for M1 Max is 1.7X as compared to M1 Pro. (The range was 1.4X to 1.7X.) AFAIK, nowhere do they show a real world test with 2X scaling.
Na still if the opinion it it either a geekbench issue or bad scaling from 16 to 32 cores.
If it is throttling that their cooling is not as good as they make us believe.
Yep, I think it's likely this is the 32 core now machine my self. I was thinking about this when I woke up.
I HIGHLY doubt they are testing these machines with High Performance Mode on. Doesn't Apple typically ship reviewer devices with the current RELEASED OS? Not the beta OS? If that's the case, there's no way reviewers are testing with high performance mode. I'm spectulating out loud here, but I wonder if the "regular" mode normally runs at around 30-40W, with high performance mode running at the 50-60Ws they showed when comparing it to the laptop 3080's. I have no evidence to suggest this is actually the case, but could be really interesting if it was true!
Yep, there's no doubt this will happen. M1 Max 32 core GPU has less compute execution units than the 3080 laptop.
This primarily sums up my feelings on the matter. That all being said, we still haven't seen the full results, so I'm going to wait and reserve judgement until we actually see it all.
Like you said, Geekbench doesn't run long enough to thermal throttle (in 90% of cases). There's NO chance this is happening to the new thermal design on either the 14 or 16. I'm also doubtful that these benchmarks are using High Performance Mode, for the reasons I stated above.
what are you not understanding about "synthetic benchmark"
...there are many variables that could explain why this isnt scaling properly on geekbench..BUT WHO CARES, ITS GEEKBENCH. Seriously why is this whole thread hung up on this one stupid benchmark number.
We’re having a discussion about tech. Moderate your tone.
Higher benchmark numbers = bigger e-peen.
Until your own numbers are lower, then “benchmarks don’t matter”
This is a universal law of tech forums
Yes, and you can even compare the M1 Max to the M1 Pro. One subtest does not show any improvement.
Pretty much this. Benchmarks are weird. They are an ok approximation in absence of real data, and some of them are even good approximations. But I would still like to know why AMD GPUs are slower in Geekbench compute than Nvidia GPUs with the same nominal peak throughput...
There are also a huge difference between Vulkan and openCL versions of the compute test on Windows. It makes me wonder if it performs the same task with two different APIs, or if it's just two different tests.
No. It appears all review units are running Monterey. This makes sense, because I would expect Apple won't ship any of these machines with Big Sur. Remember, Monterey comes out officially next week anyway.
Which really suggests to me that not all of the resources are duplicated and these tests must rely on some subsystem that wasn't duplicated.
Remember: Geekbench GPU is not a graphics rendering benchmark. If we don't see (and it remains to be tested stringently) the expected scaling from doubling GPU resources and bandwidth, it could be connected to other limitations, such as SLC bandwidth not doubling with capacity (since the benchmark predominantly runs from cache), thermal/power throttling, et cetera. We simply can't tell from this data.
That’s not entirely true. I have no experience with Octane Rendering, but I do use Blender 3D and it is very much cross-platform, and it supports M1.
The Blender Foundation is working on a native M1 version, that’s in beta.
Just a reminder Geekbench is just 1 tool and as other have mentioned, Geekbench is more tuned for Nvidia GPUs.
The GFXBench numbers are showing the M1 Max closer to (and in some things actually surpassing) the 3080m. Still no way of telling if this is in high power mode so maybe these numbers will be even better.
The GFXBench numbers are showing the M1 Max closer to (and in some things actually surpassing) the 3080m. Still no way of telling if this is in high power mode so maybe these numbers will be even better.
GFXBench shows M1Max neck and neck with the RTX 3080 laptop GPU. Its actually a bit ahead.
If you filter on GFXBench to only show DESKTOP, and MacOS-Metal and cpu=ARM you get all the M1 family there now. still no breakout by model or GPU counts.
-d
Serious question, because I don't actually know, but what workloads are compute intensive that requires the GPU? Can people provide examples for me please?
To answer the question “why does geekbench matter?”. Simply because it’s pretty much all we have right now.
These are new computers with new processors and gpus. None of us have used them and there aren’t any reviews. All we can do is “read the tea leaves” about performance. Especially gpu performance. Some of us are trying to understand what will be a good value purchase. It’s strange how the scores are scaling, and that could help people decide what to buy.
Once we know more about real world performance, then benchmarks are less important, but for now they are the best approximation of real world performance.
Finally, for many of us, discovering and pondering about this stuff is fun.
These are new computers with new processors and gpus. None of us have used them and there aren’t any reviews. All we can do is “read the tea leaves” about performance. Especially gpu performance. Some of us are trying to understand what will be a good value purchase. It’s strange how the scores are scaling, and that could help people decide what to buy.
Once we know more about real world performance, then benchmarks are less important, but for now they are the best approximation of real world performance.
Finally, for many of us, discovering and pondering about this stuff is fun.
Apple rates the GPU at 10.4 TFLOPS and explicitly says it’s 4x the M1 and 2x the M1Pro. No not every application will see that level of scaling, but in general compute should and does between the Pro and M1. So we’re trying to figure out where the bottleneck is.
For consumer/prosumer GPUs? There are scientific applications that use such GPUs for 32bit compute (many make heavy use of 64-bit compute on professional GPUs but not all), some cryptocurrency algorithms unfortunately, AI (Apple has a separate NPU in addition to AMX blocks on the CPU, but most AI research is GPU accelerated with the tensor cores built into the GPU), and rendering. Probably more but those are off the top of my head. Basically anything compute intensive where the algorithm ranges from “really” parallel to “embarrassingly” parallel (especially the latter) is a good fit for the GPU with caveats.
Last edited:
More specifically, could you tell me tasks that people actually do, i.e, export a video, apply effects to a video, uhhh... dataset number crunching or something?