Unpacking quantised weights IS NOT FREE you can check it using coreml-quant and compare it to native weights. Ofc. i don't say it is main bottleneck as it can be as low as x-10% performance drop depending on method of compression and used coprocessor.
hey you got it last November. Just got my M3 MBA a few weeks ago 😂
I think at this time we can be certain this means new AMX units with SME/streaming mode SVE support. Better late than never
That sounds like a conservative take. They are comparing a preliminary M4 benchmark in a 5mm thick iPad to a 1% top score of an actively cooled M3 Max. Of course such comparison would penalize the M4. If we look at the typical high scores of the M3 Air instead we consistently see 10-20% improvements in pretty much every subtest (Object Detection notwithstanding). The IPC improvement is there, and it's quite sizable.
It really looks like the M4 CPU is the "real deal" while M3 CPU is more of an alpha version.
I couldn’t believe my eyes when I saw that 253 Watt max multi-core spec for Intel’s flagship chip. And that will be reigned in to 188 Watts soon too.
Seriously? We just did this a couple weeks ago. QC's new chip doesn't even stack up to the M3. It loses badly in SC, and depending on the power envelope, loses to either the M3 or the M3 Pro in MC. The only way it can win is if they price it at or lower than M3 - if it comes in a few hundred $ less than the M3 Air, it will certainly find a warm welcome. Even priced the same as the Air, if it comes in a similar package, it could do well in the Windows world (though it will lose in perf to the M3 Air in all ways, in that configuration).
Almost no chance of that.
It would be unprecedented for Apple to release new M4 MacBook Airs at WWDC. I got my first day order M3 MBA on March 11, 2024. That was less than 60 days ago. That would be less than 100 days at WWDC. I'm nearly certain that that's not going to happen. I do believe that it is possible that we will see mini, studio, and mac pro at WWDC though.
I agree, and I would bet against it. But not a lot! The M4 shipping in the new iPad is unprecedented (or as nearly as the "unprecedented" you mentioned) in at least two different ways.
Or they skip it. Apple likely has a commitment to TSMC to consume a certain number of wafers. They might just keep them in the Air for the rest of the year, maybe update the iPad Air to M3 sometime in the future, and go to M5 when it's here.
Fair enough, but I think it's reasonably clear I spent quite a bit of time looking at various sources, I just didn't realize that GB6 had a recent update.Geekbench 6.3 - Geekbench Blog
www.geekbench.com
Come on, can you just look at the change log?
>
- Introduce support for Arm Scalable Matrix Extensions (SME) instructions.Geekbench 6.3 includes SME implementations of the matrix multiplication kernels used by the Geekbench 6 machine learning workloads.
...and AMD is likely to announce Zen 5 shipping at Computex.
You know, that's true right now. But...
I think that's extraordinarily unlikely. If I were Apple I'd have been damn sure to negotiate that my commitment would be portable from N3B to N3E. And I doubt TSMC would care. They use the same machines and raw materials to produce both. I suspect that both Apple and TSMC would prefer to move everything to N3E as soon as possible. It improves everyone's logistics a lot.
Yes about the tandem display. As for the rest, I bet you a virtual dollar you're wrong.
I was going to say that that seems unlikely - if PL had the iPad, they'd know what to call it instead of iPad 16,x. But... perhaps they got an early M4 Mac. Or else, perhaps Apple offered patches? Though that seems unlikely, just because I imagine PL being leery of accepting code from a vendor they test.
Doesn't even have to be anything Apple ever ships to consumers, Apple could have offered Primate Labs a VM running on one of the silicon bringup / validation boards Apple keeps in their labs:
Not necessarily. SME has apparently been part of XCode for the past year (at least since October or so). It's been part of LLVM for longer, but was rolled into XCode at around that time.
I'm guessing Primate tested their code on a simulator?
This is a very interesting table!Here's a fun table I made:
Rather than continually deal with IPC averages, I thought it might be fun to look at each individual test and see how they changed since the M1. I chose a GB 6.3 result from each processor (so run-to-run variation comes into play I'm not trying to get exact numbers from averages) and compare the change with clock speed. Given my results I can say that IPC increases depends strongly on the workload. For instance since the M1, IPC for GB's HTML 5 and Background Blur tests have increased roughly 40%, IPC for PDF Render and Photo Library and Object Remover and Ray Tracer have increased ~20%, IPC for Clang, HDR, Photo filter by 11-15%. IPC Text Processing, Asset Compression, Structure from motion, are about 7%, File Compression and Navigation are about 3-4%, while Horizon Detection is completely flat. Object Detection prior to SME went up by 18% between the M1 and M2 but was flat between the M2 and M3. Obviously unknown what it would've done in the M4 without SME. Now if you want you could create "an average" of those by taking the geometric mean for FP and INT tests and weighted arithmetic mean over the two but that would conceal everything that's interesting and why I don't like averages. This shows in fact Apple that is iterating quite strongly in areas they care about for the CPU's performance and they are leaving to clock speed those that they don't. The average is brought down by the latter. I would argue rather than the criticism that Apple "studies for the test" it would appear that they have their own design priorities for what's most important to improve for their users and those are different from GB's.
I can't attach the full spreadsheet to check it for errors, but I think it's right. It only took me a bit to do this so anyone could do the same (and please do point out any errors beyond run-to-run stuff if you get very different results).
Also I should point out that clocks have increased by 38% since the M1. Often times big clock speed increases like we've been getting necessitate microarchitecture changes just to keep up as otherwise IPC falls as clocks rise (especially if you increase them by nearly 40%). Thus, part of this is that Apple has been so aggressive with clocks, particularly M2->M3 was a huge gain in clocks (I know the M2 Max could clock higher, but this is base M2) and there was even a pretty sizable jump in M3->M4. These have "eaten" IPC gains if you will. That they've also increased clocks by this much without spiking power to ungodly amounts is a testament to both their and TSMC's engineering efforts as well.
@name99 @mr_roboto @leman
Absolutely just to reiterate though the exact numbers are subject to run to run variation and looking at the data on GB's website I may have picked a particularly good M4 result.
But's it's more about the overall relationships.