M1 Pro and M1 Max GPU performance versus Nvidia and AMD

Posted:
in Current Mac Hardware edited October 19
Apple's M1 Pro and M1 Max have GPU speeds competitive with new releases from AMD and Nvidia, with higher-end configurations expected to compete with gaming desktops and modern consoles




During Apple's keynote, the company boasted about the graphical performance of the M1 Pro and M1 Max, with each having considerably more cores than the M1 chip. It was said that the M1 Pro's 16-core GPU is seven-times faster than the integrated graphics on a modern "8-core PC laptop chip," and delivers more performance than a discrete notebook GPU while using 70% less power.

The M1 Max was said to have even more performance, with it apparently comparable to a high-end GPU in a compact pro PC laptop, while being similarly power efficient.

In estimates by NotebookCheck following Apple's release of details about its configurations, it is claimed the new chips may well be able to outpace modern notebook GPUs, and even some non-notebook devices.

The reference for the publication is the known quantity, namely the M1, which has an eight-core GPU that manages 2.6 teraflops of single-precision floating-point performance, also known as FP32 or float32.

Apple Silicon processorM1M1 ProM1 ProM1 MaxM1 Max
GPU core count814162432
Teraflops2.64.55.27.810.4
AMD equivalent GPURadeon RX 560
(2.6TF)
Radeon RX 5500M
(4.6TF)
Radeon RX 5500
(5.2TF)
Radeon RX 5700M
(7.9TF)
Radeon RX Vega 56
(10.5TF)
Nvidia equivalent GPUGeForce GTX 1650
(2.9TF)
GeForce GTX 1650 Super
(4.4TF)
GeForce GTX 1660 Ti
(5.4TF)
GeForce RTX 2070
(7.4TF)
GeForce RTX 2080
(10TF)


In the case of the M1 Pro, the 14-core variant is thought to run at up to 4.5 teraflops, while the advertised 16-core is believed to manage 5.2 teraflops. For the M1 Max, the 24-core version is expected to hit 7.8 teraflops, and the top 32-core variant could manage 10.4 teraflops.

If the estimates turn out to be accurate, it does put the new M1 chips in some esteemed company.

The 16-core GPU in the M1 Pro is thought to be 5.2 teraflops, which puts it in the same ballpark as the Radeon RX 5500 in terms of performance. The Nvidia equivalent would be the GeForce GTX 1660 Ti, which is slightly faster at peak performance with 5.4 teraflops.

At the high end, the M1 Max's 32-core GPU is at a par with the AMD Radeon RX Vega 56, a GPU that Apple used in the iMac Pro. Its Nvidia equivalent would be something like the GeForce RTX 2060.

Against game consoles, the 32-core GPU puts it at a par with the PlayStation 5's 10.28 teraflops of performance, while the Xbox Series X is capable of up to 12 teraflops.

For the moment, these are estimates based on what Apple said during its special event and in the following press releases and product pages, and therefore can't really be considered perfectly accurate, aside from the M1's performance.

Tflops are not the ultimate comparison of GPU performance. Real-world performance varies depending on if a task is CPU-bound, or if the GPU has a constant flow of data at the theoretical maximum data transfer rate. The performance estimates by the report also assume that the chips are running at the same clock speed as the M1.

No one outside of Apple will truly know the performance of the new chips until the latest 14-inch MacBook Pro and 16-inch MacBook Pro ship to consumers. At that time, benchmarks will reveal how powerful the new M1 chips truly are.

Read on AppleInsider
«1

Comments

  • Reply 1 of 29
    You can't compare Teraflops from one GPU architecture to the next.  That is not how it works.

    For example, the Radeon RX 5700 XT had 9.7 Tera flops for single, the previous generation the Radeon RX Vega 64 had a 12.6 Tera flops for single and yet in the benchmarks the Radeon RX 5700 XT was superior.

    The idea that a Vega 56 is as fast as a GeForce RTX 2080 is just laughable.  There is not a single benchmark review that puts the Vega 56 matching or beating the GeForce RTX 2080.
    edited October 19 prismaticsentropyscgWerksFileMakerFellerbeowulfschmidtwatto_cobrajony0
  • Reply 2 of 29
    I’m assuming that, as many other times, the real-world performance will exceed the expectations built on the announcement. Remember what happened with the original M1 machines? “Oh, it’s going to be bad with only 16GB of memory”, and look at what was actually delivered. I believe it will be the same with these new machines.
    aderutterwilliamlondoncgWerksdanoxlkruppwatto_cobrajony0
  • Reply 3 of 29
    I'm waiting for someone to overclock the M1 Max and put watercooling in the Macbook Pro to squeeze ridiculous amounts of power in it ("just because it is fun"). Not needed at all, but it would get people's attention.
    watto_cobra
  • Reply 4 of 29
    The M1 Pro and M1 Max are extremely impressive processors. Not only are the CPUs among the best in computer the market, the GPUs are the best in the laptop market for most tasks of professional users. A minor concern is that the Apple Silicon GPUs currently lack hardware ray tracing which is at least five times faster than software ray tracing on a GPU. Apple is likely working on hardware ray tracing as evidenced by the design of the SDK they released this year which closely matches that of NVIDIA's. Hopefully it will appear in the M2. The one area where the M1 Pro and Max are way ahead of anything else is in the fact that they are integrated GPUs with discrete GPU performance and also their power demand and heat generation are far lower. Not only does this mean that the best laptop you can buy today at any price is now a MacBook Pro it also means that there is considerable performance head room for the Mac Pro to use with a full powered M2 Pro Max GPU. That one could very well be the most disruptive processor to hit the market. It will be interesting to see how NVIDIA and AMD rise to the challenge.

    Also note the 64 GB of vRam is unheard of in the GPU industry for pro consumer products. For some tasks, the new MacBook Pros will be the best graphics processor on the market. Better even than desktop computers.
    edited October 19 FileMakerFellerwatto_cobrajony0
  • Reply 5 of 29
    HrebHreb Posts: 18member
    So the M1 Max, announced yesterday, deployed in a laptop, has floating-point compute performance (but not any other metric) comparable to a 3 year old nvidia chipset or a 4 year old AMD chipset.  Somehow I don't think this comparison is going to be useful to anybody.
    williamlondoncgWerks
  • Reply 6 of 29
    cgWerkscgWerks Posts: 2,752member
    OK, that isn't nearly as positive as I had hoped (and not as good as some previous estimates I've seen).

    That means the Pro is more entry-level dGPU, and the Max is kind of middle-of-the-road dGPU. I guess that's fine in laptops, but then hopefully we'll see more cores in the desktops that are coming (mini 'Pro', bigger iMac, etc.). I'm sure we'll see another 'Uber' version for the new Mac Pro, but I'm disappointed in terms of anything below that level if these same chips will go in them. Maybe there is still hope for eGPUs?

    And, then there is still the issue Metal (vs OpenGL, etc.) and apps kind of emulating until they get ported (if they ever do), which further degrades performance.
    williamlondon
  • Reply 7 of 29
    cgWerkscgWerks Posts: 2,752member
    Blizzard said:
    You can't compare Teraflops from one GPU architecture to the next.  That is not how it works.
    ...
    The idea that a Vega 56 is as fast as a GeForce RTX 2080 is just laughable.  There is not a single benchmark review that puts the Vega 56 matching or beating the GeForce RTX 2080.
    Depends on what you're doing. For example, they are quite close on crypto-mining. I'm not sure about 3D apps, but the software makes so much difference there (ie. CUDA). But, you're right that it is hard to directly compare, unless you have some particular application in mind.

    I’m assuming that, as many other times, the real-world performance will exceed the expectations built on the announcement. Remember what happened with the original M1 machines? “Oh, it’s going to be bad with only 16GB of memory”, and look at what was actually delivered. I believe it will be the same with these new machines.
    True, but this will, IMO, largely come down to software where we're already at a disadvantage. But, yes, for apps like were shown (C4D, Resolve, etc.) it will probably be pretty good. Will it match high end GPUs though? I missed that 'compact' laptop thing when watching the event. You can get laptops with an Nvidia 3080, and I think I even saw a 3090 the other day, which are considerably faster than a Vega 56/64.

    ... Not only does this mean that the best laptop you can buy today at any price is now a MacBook Pro it also means that there is considerable performance head room for the Mac Pro to use with a full powered M2 Pro Max GPU. That one could very well be the most disruptive processor to hit the market. It will be interesting to see how NVIDIA and AMD rise to the challenge.

    Also note the 64 GB of vRam is unheard of in the GPU industry for pro consumer products. For some tasks, the new MacBook Pros will be the best graphics processor on the market. Better even than desktop computers.
    Except note the are talking about 'compact laptops'. As mentioned above, some of the bigger laptops have Nvidia 3080 (maybe 3090) in them, which I don't think this will compete with, at least in every way. That RAM thing could be a game-changer, true, but probably only for certain things.

    Of course we don't know how far Apple will advance in the next year, but now seeing this, I really, really hope we'll be getting eGPU support back, as this might not cut it for higher end pros until the Mac Pro. I was hoping we wouldn't be in that situation again (where the Mac Pro was the only option). I was really hoping this transition would enable a more prosumer machine like the Wintel market has in spades.

    But, as has been said, this (article subject matter) is only one (hopefully highly flawed) take on the situation.
  • Reply 8 of 29
    entropysentropys Posts: 3,200member
    I don’t believe this is a valid comparison, and not reflective of actual performance of any of these GPU.
    I have just replaced a Vega 64, the top Radeon in its days and higher spec, with a base 6600XT, and the 6600XT does higher frame rates at higher settings. Technically, the Vega 64 has double the bandwidth.
    who knows what the sum of the M1 and M1 pro max can do until real world testing.
    edited October 19 danoxFileMakerFellerwatto_cobracgWerks
  • Reply 9 of 29
    There's a reason you don't see GPUs directly compared using TFlops. This is a meaningless chart. 

    Should have a better idea Monday when the review embargo is over.
    watto_cobra
  • Reply 10 of 29
    Can we play Crysis?
    williamlondonbeowulfschmidtwatto_cobracgWerks
  • Reply 11 of 29
    danoxdanox Posts: 679member
    Hreb said:
    So the M1 Max, announced yesterday, deployed in a laptop, has floating-point compute performance (but not any other metric) comparable to a 3 year old nvidia chipset or a 4 year old AMD chipset.  Somehow I don't think this comparison is going to be useful to anybody.
    If those AMD and Nvidia chipsets are in laptops bet they don’t work too well went unplugged from a wall socket…
    williamlondonthtFileMakerFellerwatto_cobra
  • Reply 12 of 29
    Simply comparing GPU internal speeds also ignores the Wintel copy overhead.

    Using the discrete GPU model, requests and data are formatted in main memory by the CPU, compressed, transmitted by the CPU over PCIe, received by the GPU into GPU memory, decompressed, then executed by the GPU. All this results in realtime clock tick loss which has nothing to do with putting pixels on screen.

    Compute tasks returning results have to go through the same process in the opposite direction.

    All this overhead means that despite the possibly superior speed of a discrete GPU, the actual speed can come surprisingly close the the M1's 7 or 8 cores.

    This applies to all Wintel graphics workflows, no mater the internal speed of the GPU.

    For integrated GPUs, the same basic process involves copying data from main memory to the GPU's memory partition, replacing the compress, transfer over PCIe, and decompress portions of the cycle.
    jeroenhmgFileMakerFellerbadmonkwatto_cobra
  • Reply 13 of 29
    I won't make any conclusions until the first real-world tests and even after that, a good year is needed at least for developers to get to know with this new architecture, even if there are A* processors for a while.
    williamlondonwatto_cobra
  • Reply 14 of 29
    My experience on the M1 MBA 16GB make me very optimistic regarding the GPU performance of the M1Pro/M1Max.

    I play Elder Scrolls Online (still!) and on the M1 MBA it runs using Rosetta 2. I play it at 2560x1440 on an external 32" 4K screen with medium to high level graphics settings and it is pretty smooth most of the time. I usually play this on a 2018 MBP i9 with a Vega56 eGPU on the same 32" 4K screen and it is certainly a lot smoother with the eGPU (I use the same settings on both).

    I recently, for fun, installed ARM Parallels with ARM Windows 10 and then installed Intel Far Cry 2 which is a bit of an old game but still great. It was properly playable and ran smoother at same settings than it does on the 2018 MBP i9 using Parallels and Windows 10, where it is not as playable due to choppy framerates. Sure, if I install it in Bootcamp it will be better but that was not the point of the exercise.

    CSGO (Rosetta2 I think) also plays wonderfully at 1680x1050 with everything turned on and set to high on the internal screen.


    Also Apple listed the laptops they used for the GPU performance comparisons, here they are (footnote 26 on the M1Pro/M1Max MBP info page):

    Discrete PC laptop graphics performance data from testing Lenovo Legion 5 (82JW0012US).

    https://www.newegg.com/p/1TS-000E-0SX24

    https://www.digitweek.com/lenovo-legion-5-15-82jw0012us-review/


    High-end discrete PC laptop graphics performance data from testing MSI GE76 Raider (11UH-053).

    https://www.newegg.com/titanium-blue-msi-ge-series-ge76-raider-11uh-053-gaming-entertainment/p/N82E16834155868

    https://www.tomshardware.com/reviews/msi-ge76-raider-review


    PC compact pro laptop performance data from testing Razer Blade 15 Advanced (RZ09-0409CE53-R3U1).

    https://www.razer.com/gaming-laptops/Razer-Blade-15/RZ09-0409CE53-R3U1


    edited October 20 FileMakerFellerwatto_cobra
  • Reply 15 of 29
    Hreb said:
    So the M1 Max, announced yesterday, deployed in a laptop, has floating-point compute performance (but not any other metric) comparable to a 3 year old nvidia chipset or a 4 year old AMD chipset.  Somehow I don't think this comparison is going to be useful to anybody.
    Indeed. I assume it will use much less power which is great, because pc laptops with those GPU's don't last long on a battery charge. So that's something. But yeah, it seems middle of the road, which is where highend Apple GPU performance always has been.

    Also, I wonder how we are going to benchmark them properly, most GPU heavy games don't run a mac, and the Pro Apps don't run on PC's..Maybe only Cinebench? Adobe apps skew heavily towards the Windows platform.

    In the end maybe it doesn't matter, since most people buying these will using the Pro Apps anyway. 3D and ML has moved to other platforms.
    williamlondonwatto_cobra
  • Reply 16 of 29
    What will such a processor do to me if I can't play FPS AAA games with it :( I Would be happy to test the M1 Max with some latest COD or similar game. Because it's fun not only work ;)
    williamlondonwatto_cobra
  • Reply 17 of 29
    mcdavemcdave Posts: 1,831member
    The compute looks disappointing though it is OpenCL so Metal should do better:

  • Reply 18 of 29
    mjtomlinmjtomlin Posts: 2,490member
    The problem with these types of comparisons (both CPU and GPU) is that a lot of the work those "main" units used to do will be offloaded to other processing units on Apple's SoC, which means in real world tests these SoC's are going to be a lot more capable then what the benchmarks score.
    watto_cobra
  • Reply 19 of 29
    KITAKITA Posts: 382member
    mcdave said:
    The compute looks disappointing though it is OpenCL so Metal should do better:


    The Apple M1 scores ~18,000 in OpenCL (1) and ~20,000 in Metal (2) on Geekbench.

    A 130W version of the RTX 3080 in a laptop scores ~120,000 in OpenCL (3) on Geekbench.

    Apple used a 165W version of the RTX 3080 in their comparison:


    cgWerks
  • Reply 20 of 29
    Can we play Crysis?
    Exactly.  Teraflops don't really matter much.  What matters for games is frames per second and average frame time.  What matters for productivity apps is time to task completion.  What matters for mining is the hash rate.  And heat generation and dissipation matter for all of them.

    I'm reasonably sure that the M1 Pro and Max will measure up favorably to comparable laptops in "real world" comparisons, but a teraflops comparison is really kind of secondary, or maybe even tertiary, where actual application benchmarks are concerned.
    watto_cobra
Sign In or Register to comment.