Most active commenters
  • aurareturn(7)

←back to thread

Ask HN: Why hasn't x86 caught up with Apple M series?

447 points stephenheron | 15 comments | 25 Aug 25 21:50 UTC | HN request time: 0.001s | source | bottom

Hi,

My daily workhorse is a M1 Pro that I purchased on release date, It has been one of the best tech purchases I have made, even now it really deals with anything I throw at it. My daily work load is regularly having a Android emulator, iOS simulator and a number of Dockers containers running simultaneously and I never hear the fans, battery life has taken a bit of a hit but it is still very respectable.

I wanted a new personal laptop, and I was debating between a MacBook Air or going for a Framework 13 with Linux. I wanted to lean into learning something new so went with the Framework and I must admit I am regretting it a bit.

The M1 was released back in 2020 and I bought the Ryzen AI 340 which is one of the newest 2025 chips from AMD, so AMD has 5 years of extra development and I had expected them to get close to the M1 in terms of battery efficiency and thermals.

The Ryzen is using a TSMC N4P process compared to the older N5 process, I managed to find a TSMC press release showing the performance/efficiency gains from the newer process: “When compared to N5, N4P offers users a reported +11% performance boost or a 22% reduction in power consumption. Beyond that, N4P can offer users a 6% increase in transistor density over N5”

I am sorely disappointed, using the Framework feels like using an older Intel based Mac. If I open too many tabs in Chrome I can feel the bottom of the laptop getting hot, open a YouTube video and the fans will often spin up.

Why haven’t AMD/Intel been able to catch up? Is x86 just not able to keep up with the ARM architecture? When can we expect a x86 laptop chip to match the M1 in efficiency/thermals?!

To be fair I haven’t tried Windows on the Framework yet it might be my Linux setup being inefficient.

Cheers, Stephen

Show context
ben-schaaf ◴[26 Aug 25 07:03 UTC] No.45023206[source]▶
Battery efficiency comes from a million little optimizations in the technology stack, most of which comes down to using the CPU as little as possible. As such the instruction set architecture and process node aren't usually that important when it comes to your battery life.

If you fully load the CPU and calculate how much energy a AI340 needs to perform a fixed workload and compare that to a M1 you'll probably find similar results, but that only matters for your battery life if you're doing things like blender renders, big compiles or gaming.

Take for example this battery life gaming benchmark for an M1 Air: https://www.youtube.com/watch?v=jYSMfRKsmOU. 2.5 hours is about what you'd expect from an x86 laptop, possibly even worse than the fw13 you're comparing here. But turn down the settings so that the M1 CPU and GPU are mostly idle, and bam you get 10+ hours.

Another example would be a ~5 year old mobile qualcomm chip. It's a worse process node than an AMD AI340, much much slower and significantly worse performance per watt, and yet it barely gets hot and sips power.

All that to say: M1 is pretty fast, but the reason the battery life is better has to do with everything other than the CPU cores. That's what AMD and Intel are missing.

> If I open too many tabs in Chrome I can feel the bottom of the laptop getting hot, open a YouTube video and the fans will often spin up.

It's a fairly common issue on Linux to be missing hardware acceleration, especially for video decoding. I've had to enable gpu video decoding on my fw16 and haven't noticed the fans on youtube.

replies(14): >>45023243 #>>45023603 #>>45023693 #>>45023904 #>>45023939 #>>45023972 #>>45024390 #>>45024405 #>>45024494 #>>45025515 #>>45026011 #>>45026727 #>>45026857 #>>45027696 #
1. aurareturn ◴[26 Aug 25 08:48 UTC] No.45023972[source]▶

  All that to say: M1 is pretty fast, but the reason the battery life is better has to do with everything other than the CPU cores. That's what AMD and Intel are missing.
This isn't true. Yes, uncore power consumption is very important but so is CPU load efficiency. The faster the CPU can finish a task, the faster it can go back to sleep, aka race to sleep.

Apple Silicon is 2-4x more efficient than AMD and Intel CPUs during load while also having higher top end speed.

Another thing that makes Apple laptops feel way more efficient is that they use a true big.Little design while AMD and Intel's little cores are actually designed for area efficiency and not power efficiency. In the case of Intel, they stuff as many little cores as possible to win MT benchmarks. In real world applications, the little cores are next to useless because most applications prefer a few fast cores over many slow cores.

replies(2): >>45024922 #>>45027555 #
2. yaro330 ◴[26 Aug 25 11:05 UTC] No.45024922[source]▶
> Apple Silicon is 2-4x more efficient than AMD and Intel CPUs during load while also having higher top end speed.

This is false, in cross platform tasks it's on par if not worse than latest X86 arches. As others pointed out: 2.5h in gaming is about what you'd expect from a similarly built X86 machine.

They are willing due to lower idle and low load consumption, which they achieve by integrating everything as much as possible - something that's basically impossible for AMD and Intel.

> The faster the CPU can finish a task, the faster it can go back to sleep, aka race to sleep.

May have been true when CPU manufacturers left a ton of headroom on the V/F curve, but not really true anymore. Zen 4 core's power draw shoots up sharply pass 4.6 GHz and nearly triples when you approach 5.5 GHz (compared to 4.6), are you gonna complete the task 3 times faster at 5.5 GHz?

replies(2): >>45025105 #>>45036647 #
3. aurareturn ◴[26 Aug 25 11:33 UTC] No.45025105[source]▶

  This is false, in cross platform tasks it's on par if not worse than latest X86 arches.
This is Cinebench 2024, a cross platform application: https://imgur.com/a/yvpEpKF

  They are willing due to lower idle and low load consumption, which they achieve by integrating everything as much as possible - something that's basically impossible for AMD and Intel.
Weird because LNL achieved similar idle wattage as Apple Silicon.[0] Why do you say it's impossible?

  May have been true when CPU manufacturers left a ton of headroom on the V/F curve, but not really true anymore. Zen 4 core's power draw shoots up sharply pass 4.6 GHz and nearly triples when you approach 5.5 GHz (compared to 4.6), are you gonna complete the task 3 times faster at 5.5 GHz?
Honestly not sure how your statement is relevant.

[0]https://www.notebookcheck.net/Dell-XPS-13-9350-laptop-review...

replies(3): >>45025359 #>>45025599 #>>45036255 #
4. yaro330 ◴[26 Aug 25 12:03 UTC] No.45025359{3}[source]▶
> Weird because LNL achieved similar idle wattage as Apple Silicon.[0] Why do you say it's impossible?

And where is LNL now? How's the company that produced it? Even under Pat Gelsinger they said that LNL is a one off and they're not gonna make any more of them. It's commercially infeasible.

> Honestly not sure how your statement is relevant.

How is you bringing up synthetics relevant to race to idle?

Regardless, a number of things can be done on Strix Halo to improve the performance, first would be switching to some optimized Linux distro, or at least the kernel. That would claw back 5-20% depending on the task. It would also improve single core efficiency, I've seen my 7945hx drop from 14-15w idle on Windows to about 7-8 on Linux, because Windows likes to jerk off the CCDs non stop and throw the tasks around willy nilly which causes the second CCD and I/O die to never properly idle.

replies(1): >>45025969 #
5. atwrk ◴[26 Aug 25 12:27 UTC] No.45025599{3}[source]▶
This is Cinebench 2025, a cross platform application: https://imgur.com/a/yvpEpKF

You sure like that table, don't you? Trying to find the source of that blender numbers, I came across many reddit posts of you with that exact same table. Sadly those also don't have a source - the are not from the notebookcheck source.

replies(1): >>45025872 #
6. aurareturn ◴[26 Aug 25 12:53 UTC] No.45025872{4}[source]▶
The reason why I keep reposting this table is because people post incorrect statements about AMD/Apple so often, often with zero data backing.

For Blender numbers, M4 Pro numbers came from Max Tech's review.[0] I don't remember where I got the Strix Halo numbers from. Could have been from another Youtube video or some old Notebookcheck article.

Anyway, Blender has official GPU benchmark numbers now:

M4 Pro: 2497 [1]

Strix Halo: 1304 [2]

So M4 Pro is roughly 90% faster in the latest Blender. The most likely reason for why Blender's official numbers favors M4 Pro even more is because of more recent optimizations.

Sources:

[0]https://youtu.be/0aLg_a9yrZk?si=NKcx3cl0NVdn4bwk&t=325

[1] https://opendata.blender.org/devices/Apple%20M4%20Pro%20(GPU...

[2] https://opendata.blender.org/devices/AMD%20Radeon%208060S%20...

replies(1): >>45026094 #
7. aurareturn ◴[26 Aug 25 13:03 UTC] No.45025969{4}[source]▶

  And where is LNL now? How's the company that produced it? Even under Pat Gelsinger they said that LNL is a one off and they're not gonna make any more of them. It's commercially infeasible.
Why does it matter that LNL is bad economically? LNL shows that it's definitely possible to achieve same idle or even better idle wattage than Apple Silicon.

  How is you bringing up synthetics relevant to race to idle?
I truly don't understand what you mean.
8. vient ◴[26 Aug 25 13:15 UTC] No.45026094{5}[source]▶
Weren't we comparing CPUs though? Those Blender benchmarks are for GPUs.

Here is M4 Max CPU https://opendata.blender.org/devices/Apple%20M4%20Max/ - median score 475

Ryzen MAX+ PRO 395 shows median score 448 (can't link because the site does not seem to cope well with + or / in product names)

Resulting in M4 winning by 6%

replies(1): >>45026240 #
9. aurareturn ◴[26 Aug 25 13:26 UTC] No.45026240{6}[source]▶

  Weren't we comparing CPUs though? Those Blender benchmarks are for GPUs.
Yes, but I was asked about Blender GPU.

Blender CPU tasks are highly parallel. AMD's Ryzen Max 395 has great MT performance. It's generally 5-20% slower in CPU MT than the M4 Max depending on the application.

10. jandrewrogers ◴[26 Aug 25 15:01 UTC] No.45027555[source]▶
> Apple Silicon is 2-4x more efficient than AMD and Intel CPUs during load while also having higher top end speed.

This is not true. For high-throughput server software x86 is significantly more efficient than Apple Silicon. Apple Silicon optimizes for idle states and x86 optimizes for throughput, which assumes very different use cases. One of the challenges for using x86 in laptops is that the microarchitectures are server-optimized at their heart.

ARM in general does not have the top-end performance of x86 if you are doing any kind of performance engineering. I don't think that is controversial. I'd still much rather have Apple Silicon in my laptop.

replies(1): >>45028607 #
11. aurareturn ◴[26 Aug 25 16:16 UTC] No.45028607[source]▶

  For high-throughput server software x86 is significantly more efficient than Apple Silicon.
In the server space, x86 has the highest performance right now. Yes. That's true. That's also because Apple does not make server parts. Look for Qualcomm to try to win the server performance crown in the next few years with their Oryon cores.

That said, Graviton is at least 50% of all AWS deployments now. So it's winning vs x86.

  ARM in general does not have the top-end performance of x86 if you are doing any kind of performance engineering. I don't think that is controversial.
I think you'll have to define what top-end means and what performance engineering means.
replies(1): >>45034069 #
12. ksec ◴[27 Aug 25 00:36 UTC] No.45034069{3}[source]▶
I dont think the point Amazon uses ARM was about performance but purely cost optimisation. At one point, nearly 40% of Intel's server revenue was coming from Amazon. They just figure it out at their scale it would be cheaper to do it themselves.

But I am purely guessing ARM has risen their price per core so it makes less financial sense to do a yearly update on CPU. They are also going into Server CPU business meaning they now have some incentives to keep it all to themselves. Which makes the Nvidia moves really smart as they decided to go for the ISA licences and do it by themselves.

replies(1): >>45036937 #
13. ben-schaaf ◴[27 Aug 25 06:53 UTC] No.45036255{3}[source]▶
> This is Cinebench 2024, a cross platform application: https://imgur.com/a/yvpEpKF

Cool, now compare M1 to AI 340. The AI 340 has slightly better single core and better multi-core. If battery life was all about race to idle like you claim then the AI 340 should be better than the M1.

See also Snapdragon X Elite, which is significantly slower than the AI 340, uses more power under load, so in total has much less efficient cores, and yet still beats the AI 340 on battery life.

14. williamDafoe ◴[27 Aug 25 07:46 UTC] No.45036647[source]▶
I did mips per watt calculations in 2017 and Apple (A10 i think) was 2-3x better than intel. See "how to build a computer" by donald gillies (slideshare slides). I was shocked, i didnt expect this at all!
15. aurareturn ◴[27 Aug 25 08:38 UTC] No.45036937{4}[source]▶
Server CPUs do not win on performance alone. They win on performance/$, LTV/$, etc. That's why Graviton is winning on AWS.