Alleged M4 benchmarks verify Apple's iPad Pro performance claims

AppleInsiderAppleInsider
Posted:
in iPad edited May 13

The new iPad Pros skipped straight to M4, and initial benchmarks show the potential performance improvements match Apple's claims.

A hand seen drawing with an Apple Pencil on an iPad Pro
Apple's new iPad Pros run the M4 processor



Apple's M4 processor won't be available in the wild until devices arrive in customers hands on May 15. But, that may not have stopped some reviewers with pre-release units from performing benchmarks with Geekbench.

A device called iPad16,6, which appears to be the latest iPad Pro with M4 with 10 cores shows a 3,767 single core and 14,677 multi core score. For context, that's almost an exact 1.5x CPU performance boost from a 12.9-inch iPad Pro with M2, which scored 2,590 single core and 10,019 multi core.

Geekbench scores for the iPad16,6
Geekbench scores for the iPad16,6
Another benchmark

for iPad16,6 showed up with a 53,792 metal score for the GPU. That's a more modest improvement from the M2 iPad Pro scoring 46,575.

The new benchmarks arrived a few hours after an ML score for the M4's Neural Engine was discovered with a 9,234. It beat out the M3's Neural Engine, but not the M3 Max.

The Neural Engine test is suspect because it shows iOS 18, which hasn't leaked. The other new benchmarks all show iOS 17.5, which is in beta.

The data is fresh and unverifiable at this time, but it appears to indicate Apple's claims made during the "Let Loose" announcement are accurate. The baseline M-series processor makes small gains with each generation, but the more important upgrades aren't tested by Geekbench.

Real world gains will be easier to determine when users are able to compare workflows like exporting a file or running a game. Benchmark tools provide a decent baseline for comparison, but are not meant to be definitive.

More tests will have to be run once the M4 iPad Pro models launch in a week. Expect early reviews to touch on workflows and benchmarks as well.



Read on AppleInsider

Comments

  • Reply 1 of 19
    thttht Posts: 5,494member
    3767? Holy Shit!
    cpsroMisterKitwatto_cobra
  • Reply 2 of 19
    JustSomeGuy1JustSomeGuy1 Posts: 322member
    The baseline M-series processor makes small gains with each generation, but the more important upgrades aren't tested by Geekbench.
    Small gains?!? Are you nuts? It's looking like 15-20%. That a *monster* number to put up. It instantly makes the M4 the fastest consumer CPU in the world, single core. (To the extent you accept GB6 as reasonable, which... is reasonable.)

    tht said:
    3767? Holy Shit!
    You think that's holy? Wait until it ships in the M4Max, or the Studio M4whatever. If they are consistent those have a good shot at passing 4000.
    edited May 8 watto_cobra
  • Reply 3 of 19
    thttht Posts: 5,494member
    JustSomeGuy1 said:
    tht said:
    3767? Holy Shit!
    You think that's holy? Wait until it ships in the M4Max, or the Studio M4whatever. If they are consistent those have a good shot at passing 4000.
    Well, if this level of single core performance holds up across most of the GB sub-benches, that is very very very good score. Very good score. I thought 3400 was going to be the target, but 3800? Wow. There is a submission at 3810 now. They had to bump the clock to 4.4 GHz to do it, so it is not entirely IPC improvements. Wonder how much power it is using.

    Yes. 4000 may be possible in a MBP or Studio if they give it more power, or more cache, or more memory performance. I do think having this level of performance in a 5.1 mm tablet is more impressive than in a laptop.
    watto_cobra
  • Reply 4 of 19
    KITAKITA Posts: 397member
    Apple has done some impressive things on the machine learning front it seems for the M4 CPU.

    Apple M3 (Mac) vs Apple M4 (iPad Pro):

    GBM4-2

    https://browser.geekbench.com/v6/cpu/compare/6017019?baseline=6016039

    Essentially double for Object Detection has massively boosted their overall score. Background Blur also seeing significant gains:

    GBM4-1

    GB-1

    GB-2


    edited May 9 muthuk_vanalingamblastdoorwilliamlondon
  • Reply 5 of 19
    michelb76michelb76 Posts: 636member
    Finally some decent single-core improvements! Can't wait to see how it performs in the mac versions.
    watto_cobra
  • Reply 6 of 19
    sloth77sloth77 Posts: 28member
    I'll be interested to see the multithreaded score for the 3/6 configuration M4 (for <1TB storage) .  Especially how it compares to the new M2 air, which is fully enabled (4/4).
    watto_cobra
  • Reply 7 of 19
    IermodeIermode Posts: 2member
    The benchmark for the iPad is pretty insane. I think it is 20% for both single- and multicore faster for the M4 from the M3 Macbook Air. The main concern is its limited OS. If iPad can run Xcode app then I'll buy it. Sadly to say. 
    williamlondonwatto_cobra
  • Reply 8 of 19
    blastdoorblastdoor Posts: 3,337member
    KITA said:
    Apple has done some impressive things on the machine learning front it seems for the M4 CPU.

    Very interesting! So from the description of the components you highlighted, I'm inferring that it doesn't use the NPU, but rather uses the SIMD units in the CPUs, which suggests those SIMD units got some attention in this update. Following the helpful link you provided, I see that there are some bigger-than-clockspeed gains in other components, too, like html5 and pdf (but not nearly as large as the two you highlight). But gains are smaller for other things, like clang. 

    So if you're doing work that benefits from SIMD, you're going to love the M4. 
    watto_cobra
  • Reply 9 of 19
    danoxdanox Posts: 2,953member
    Iermode said:
    The benchmark for the iPad is pretty insane. I think it is 20% for both single- and multicore faster for the M4 from the M3 Macbook Air. The main concern is its limited OS. If iPad can run Xcode app then I'll buy it. Sadly to say. 

    Get a Mac Laptop....
    edited May 9 williamlondon
  • Reply 10 of 19
    mikethemartianmikethemartian Posts: 1,361member
    I assume they are going to launch M4 Macs soon. M3 might have just been a stopgap.
    edited May 9 watto_cobra
  • Reply 11 of 19
    netroxnetrox Posts: 1,436member
     It makes it more likely that Mac Studio will use M4, not M3. 


    watto_cobra
  • Reply 12 of 19
    JustSomeGuy1JustSomeGuy1 Posts: 322member
    netrox said:
     It makes it more likely that Mac Studio will use M4, not M3.
    That's a near certainty at this point.
  • Reply 13 of 19
    danoxdanox Posts: 2,953member
    Does that mean the Mac Studio M4 Ultra will have 320 gigs of UMA memory? For running AI models and does that mean there might be a possibility that Apple may design an extreme Mac Pro M4 with 640 Gigs of UMA memory? With each running under 200 watts for everything? :smile: 
    edited May 10
  • Reply 14 of 19
    ApplejacsApplejacs Posts: 38member
    Just remember when you spec your order that you MUST choose either 1tb or the 2tb option if you expect to realize these nice improvements. The only way you get 16gb of RAM, 256 or 512 only get 8gb.
    williamlondon
  • Reply 15 of 19
    mjtomlinmjtomlin Posts: 2,677member
    Applejacs said:
    Just remember when you spec your order that you MUST choose either 1tb or the 2tb option if you expect to realize these nice improvements. The only way you get 16gb of RAM, 256 or 512 only get 8gb.

    Also remember this is a passively cooled, extremely thin, mobile device... you will never see sustained performance at these levels regardless which model you get. Even the MacBook Air throttles after several minutes of "full speed" use.
    williamlondon
  • Reply 16 of 19
    mjtomlinmjtomlin Posts: 2,677member

    mikethemartian said:
    I assume they are going to launch M4 Macs soon. M3 might have just been a stopgap.

    I think WWDC... Mac Studio M4 Max/Ultra (possibly new Mac Pro - although I believe they're developing a new, separate SoC for it (M4 Extreme?) ... definitely need to add more PCI buses to make better use of all those slots.).
  • Reply 17 of 19
    KITAKITA Posts: 397member
    I missed something in my earlier post:

    Geekbench 6.3 now supports Arm SME - hence the huge boost in the machine learning workloads (https://www.geekbench.com/blog/2024/04/geekbench-63/).
    edited May 11
  • Reply 18 of 19
    blastdoorblastdoor Posts: 3,337member
    KITA said:
    I missed something in my earlier post:

    Geekbench 6.3 now supports Arm SME - hence the huge boost in the machine learning workloads (https://www.geekbench.com/blog/2024/04/geekbench-63/).
    Do we know if any other M chips support SME?
  • Reply 19 of 19
    MarvinMarvin Posts: 15,354moderator
    blastdoor said:
    KITA said:
    I missed something in my earlier post:

    Geekbench 6.3 now supports Arm SME - hence the huge boost in the machine learning workloads (https://www.geekbench.com/blog/2024/04/geekbench-63/).
    Do we know if any other M chips support SME?
    It says here M4 uses the ARM v9 architecture:

    https://wccftech.com/apple-m4-adopts-armv9-run-complex-workloads-efficiently/
    https://newsroom.arm.com/blog/armv9
    https://www.anandtech.com/show/16584/arm-announces-armv9-architecture
    https://www.engadget.com/arm-armv9-architecture-180043435.html

    This page lists which ARM versions Apple has used before:

    https://en.wikipedia.org/wiki/ARM_architecture_family#Cores

    Accelerating vector calculations will help with AI and some 3D rendering. Libraries like Embree used in V-Ray are using AVX:

    https://www.willusher.io/graphics/2020/12/20/rt-dive-m1
    https://github.com/lighttransport/embree-aarch64/
    williamlondon
Sign In or Register to comment.