> And it's doing this while using more than an order of magnitude less power (10W vs. a TDP of 125W for that Intel part).

That's the wrong conclusion to make. For instance, the Lenovo ThinkBook 14s (with a Ryzen 4800u) with a 15W TDP posts the same Geekbench multicore scores [1] as the M1 Macbook. But the ThinkBook isn't in any way faster than the top-end iMac for real world compute intensive tasks.

The M1 certainly looks efficient, but there's little you can conclude from a single benchmark running for a very short period of time.

[1]: https://browser.geekbench.com/v5/cpu/4642736

Maybe Geekbench is kinda useless as a benchmark suite then? I only see it used by Mac fans.

>There’s been a lot of criticism about more common benchmark suites such as GeekBench, but frankly I've found these concerns or arguments to be quite unfounded. The only factual differences between workloads in SPEC and workloads in GB5 is that the latter has less outlier tests which are memory-heavy, meaning it’s more of a CPU benchmark whereas SPEC has more tendency towards CPU+DRAM.

https://www.anandtech.com/show/16226/apple-silicon-m1-a14-de...

And yet here we have the M1 MacBook Air apparently beating the M1 MacBook pro, by a large margin.

It's thought that the MBP score is due to it being ran possibly during indexing on setup. The score difference is too big, it's a single sample, and the pro still fries it at single core.

Based on a sample size of 1 or 2, most likely. It could be due to something as stupid as Spotlight running in the background on the MBP but not the Air.

The silicon lottery is still a thing.

I imagine they will eventually have enough chips to start binning for different performance levels like AMD and Intel do.

They are already binning. The air is supposed to be less powerful, with cheaper variants even having one core less.

The benchmark is just ... that accurate.

They are binning for functional GPU cores and allow chips with only 7 functional GPU cores instead of 8 to go into the lower priced Air.

They are not binning for how high the cores will clock, which is just how business is done with Intel and AMD.

It's not measuring sustained performance. The fanless MacBook Air is going to throttle much sooner than a desktop iMac with proper cooling and unlimited power.

A Ryzen 4800u actually uses up to 25W TDP, depending upon implementation.

And it’s 45% slower in single core.

Most importantly, the M1 is estimated to cost Apple $65, the 4800u is a $300+ part.

That price is a meaningless comparison, you can't buy the Apple processor in retail. What's the cost to procedure the AMD part? Something similar I'd guess.

AMD on it’s current hot steak has gotten its gross profit margins to 43%, which would make the cost to manufacture a $300 part around $171.

Two and a half times higher cost to build a slower, more power hungry CPU is not actually very similar.

Yeah using the gross margin of the whole company and applying it to a single product is going to provide a reliable figure...

It’s how companies price their products. Is it perfect? No, but that’s why I used the word “around”.

You can argue that this particular Ryzen has a higher gross margin, say 50%, and lower ASP than $300, but that only gets your cost down to what, $140? And with RAM costing extra.

how much of the $300 goes to the retailer? how much goes to the distributor?

How much goes to R&D?

Not to mention I believe the RAM is included on the M1 SOC.

Unlikely OEM's are paying 300+ for a 4800u. Certainly more than $65, though.

That makes sense, AMD is selling to OEMs for a profit (over cost) while Apple is its own OEM, if any charging is done it's purely internal and for accounting purposes.

This comparison looks at different segments of the fab<>manufacturer<>OEM relationship. Add the user in there and you might say that you can buy an AMD CPU for $100 but an Apple CPU will cost you $1000. Not very meaningful as a comparison.

Certainly more than $200, so what’s your point?

While it sounds promising, I'm going to wait for some additional benchmarks and real world usage scenarios; factors like cooling, multi-process work, and of course suboptimal applications (browsers, Electron apps, stuff compiled for Intel) will be a big factor as well.

That said, it's promising and I'm really curious to see where this development will lead to in a few years' time.

> Electron apps

iOS/iPhones/iPads already smoke every other device in running javascript. Apple Silicon may end up being the best thing to ever happen for Electron apps.

Me too - let's see what the sustained performance is like. That said, with this much headroom, I'm cautiously optimistic that even with some throttling going on, it'll still be plenty fast for anything I'm likely to throw at it.

The data we don’t have is for sustained use over time. An Intel iMac Pro can sustain max performance far longer than an Intel laptop, as it has a far higher thermal exhaust capacity.

Does the M1 performance have to be ramped down during sustained use due to exceeding thermal envelope of the fanless MBA? Of the fan’d MBP?

We’ll know soon enough!

Without doubt, the Apple M1 has the highest single-threaded performance of any non-overclocked CPU, being a little faster than AMD Zen 3 and Intel Tiger Lake.

Nevertheless, because Apple has chosen to not increase their manufacturing costs by including more big cores, the multi-threaded performance is not at all impressive, being lower than that of many much cheaper laptops using AMD Ryzen 7 4800U CPUs.

So for any professional applications, like software development, these new Apple computers will certainly not blow away their competition performance-wise, and that before taking into account their severe limitations in memory capacity and peripheral ports.

But given that M1 is clearly the basic CPU for low cost, thin-and-light devices, we can strongly infer that Apple’s next M chip will be significantly more capable. Chips with eight or more performance cores would be a certainty for the upper tier of laptops and iMacs.

Given that the M1 is a full node ahead of Zen 3 and two nodes ahead of whatever Intel has to offer, one would think that when on same node, Intel and AMD will be just as capable.

But the truth is comparing to future offerings is bullshit, and we have to stick to what's available today. Impressive power/performance and all that, I have to say. We will see how sustained load looks like and how it runs non-optimized software. But to put in perspective 1 CCX of zen 3 performs better on 7nm (but draws up to 65W). With approximately the same die size (although w/o GPU and other things, the M1 has).

TSMC’s 5nm node is exceeding everyone’s expectations.

On the non-Apple side, it will be interesting what AMD does with the 5nm node.

Yes I’m guessing we can expect a die shrink of Zen 3 at least next year meaning 10-15% additional performance with no architectural changes. Crazy.

5nm is not design compatible with 7nm, but 6nm is, so there might be a die shrink to that.

Interesting! What makes the design incompatible I don’t know enough about it to know...

It’s also outperforming the 5950X in single core. Incredible!

Yeah, and doing this without a fan. It's almost like Apple is rubbing Intel's face in it for sport. It's not even fair.

But did the test run long enough to need the fan, and what was the ambient temperature?

The fanless Intel Core-M CPUs could post excellent benchmark scores (for its time). But if you give it a lengthy compile task, it'll slow down dramatically.

Aren’t they both running the same test?

Or are you saying that the test needs to run for way longer to be fair?

> Or are you saying that the test needs to run for way longer to be fair?

Yes, the main computing constraint of mobile devices is heat management (This doesn't really reflect the CPU but the complete device. Putting the CPU in a more ideal setup like a traditional desktop or water cooling will improve the CPU's performance in longer tasks)

The 5950X in a single core load only uses around 18w. Drop 200mhz off of the top end boost frequency and it drops to 11w.

Short duration single core workloads workout a fan is trivial even for CPUs that aren't trying to do so.