16 ms·
M1 Pro First Impressions: Core Management and CPU Performance
- Aeolun 5y agoI’m not sure how to interpret this, even after reading the article. Is any reason given for why the last 4 Performance cores get only a limited amount of work?
- wmf 5y agoI assume that a core cluster can power off if it's completely idle; this should be more energy-efficient than running both clusters at low utilization.
- masklinn 5y agoIt can also allow for boosting a core more: the “turbo” ramping is on a per-cluster basis so if only one core is loaded it’ll work at 3.2, two of them at 3.1, and 3 or 4 lowers that to 3.0.
- wmf 5y agoThat sounds like you'd want to keep both clusters on and spread the work so cores could clock higher. By packing everything onto one cluster, macOS is reducing clocks.
- eyelidlessness 5y agoLike sibling comment says, it’s more efficient. It allows those cores to sort of pretend they’re ECs. It also reserves those PCs for bursty workloads which would otherwise be contentious for cache and memory bandwidth in an evenly balanced scheduler.
- yazaddaruvala 5y agoIs it basically giving programs some automatic core affinity or are the unused cores meant to be unused until the first set are saturated? I guess also, are the 4 unused cores always the same?
- eyelidlessness 5y agoMy guess is it’s treating the “extra” cores as a throttling mechanism without throttling. If your thermal budget is 9 (8 + 2*1/2), and you operate normally with a thermal budget of ~5, you can temporarily go to 9 without any noticeable difference in actual performance, or heat, or cooling. In another sense, they’re probably leaving a lot of performance potential on the table to achieve that perception. As far as which cores are demoted… I definitely don’t know enough to speculate. If favoring the same set would affect longevity of the device, they’d rotate. If not it’s simpler not to do that. I have no idea which is more likely.
- matthew-wegner 5y agoKeep in mind the only devices with M1 Pro/Max CPUs are currently battery-powered laptops (which in itself is quite a strange inversion!) Probably the "high power" mode available on an M1 Max in a 16" chassis is doing just that. And maybe Apple will decide to treat some kind of new iMac Pro as more of a pro machine, and make it chunky enough to dissipate that sort of heat without needing much fan noise...
- stephen_g 5y agoYeah, it was surprising that they didn't pop these SoCs into the Mac Mini but perhaps they are doing a full redesign of the Mini and we'll see a new version some time next year? Obviously a 'Pro' version of the iMac is coming eventually with a larger screen but I don't think many people expected that yet. Maybe February or so next year...
- iSnow 5y agoLaptops command higher margins. They will want to milk the market with their Pro/Max laptops and then release Minis and larger iMacs next year. Also, even a huge company like Apple would want to stagger releases so design teams can have some constant workload. Releasing laptops and a new Mini and a new iMac would put quite some work on the team and then it would be idle for a time. They are probably done with the Mini and working on a Mac Pro right now.
- christoph 5y agoUnless I’ve misunderstood it’s somewhat guessed at, at the very end - “I suspect that Apple has done this to further improve energy efficiency and ensure good responsiveness to new CPU-intensive tasks.”
- deagle50 5y agoThey are in a separate, power and frequency gated cluster with its own L2. You'd want to keep that cluster in a low power state as much as possible.
- TillE 5y ago"Surprisingly, these benchmarks seldom exceed 50% load on any of the ten cores, which raises the question of how accurately they represent maximum CPU performance." Eh, maybe. I'm not inclined to rely on the reported % CPU usage to represent anything useful. For example, it's very easy to be at 100% CPU when you're actually spending most of that time waiting for memory reads.
- Gigachad 5y agoCPU % is such a complex metric that it hardly means anything real other than an general feel for how busy the system is. I have seen on linux that I can be running something that uses 100% and the cpu is at 50c, then I run prime95 also at 100% and the cpu instantly hits 99c
- throwawaylinux 5y agoIn Linux, the actual metric is simple. The amount of time a "CPU" (what Linux calls an architectural thread of execution) is not in the idle routine, as a proportion of the total time elapsed, and average this over all "CPUs" in the system. It's a simple metric, but it can be complicated to relate it to what is going on with execution because the real system is very complicated. The CPU % metric does not take into account execution strength, of the CPU, big vs small, dynamic CPU frequency, or effect that one CPU might have on another (e.g., SMT or shared caches or memory controllers). Further complicating it is that the Linux CPU scheduler is not work-conserving. So you could have at least 100 application threads runnable at any given moment in your workload, but your 64 CPU system might only be hitting 80% CPU busy. Then you get to application and kernel effects of course, locking, blocking, etc. can mean you don't even have as many runnable threads as you might think. Then how all this actually relates to the work and heat the system creates is another matter again. Simple integer execution might only use half the CPU power of a vector heavy workload because you have fewer transistors clocked or switching (or even powered) in the core.
- sounds 5y agoThe way CPU usage is measured includes time waiting for memory reads. Only a hyperthreaded core would be able to schedule a different task on such a core. For the M1 Pro, which is not hyperthreaded, if all dispatched instructions are waiting for memory reads, then the retired instruction counter simply does not get incremented for that cycle.
- TheTon 5y agoI have noticed that threads running at the Utility QoS level perform a lot more consistently on the M1 than they do on Intel. Utility being the lowest libdispatch QoS level that still gets scheduled on a P core. For example, a piece of code I was optimizing recently ran with very little variation in time between runs on my M1 system, but had ~30% variation on a similarly loaded Intel Mac. Changing the QoS level to Default instead of Utility allowed the Intel Mac to perform much more consistently on the benchmark, but made no difference on the M1 (I have good reason to leave it at Utility in my actual application, but for benchmarking it made sense to try other levels). I found it hard to tell from the article, but it sounds like on the M1 Pro maybe the two highest QoS levels (interactive and user initiated) now preferably map to p cores 0-3, and maybe now default and utility map to p cores 4-7, and then background gets the e cores?
- masklinn 5y agoHave you noticed whether `nice(1)` also triggers these QoS change, or is it limited to libdispatch? > it sounds like on the M1 Pro maybe the two highest QoS levels (interactive and user initiated) now preferably map to p cores 0-3, and maybe now default and utility map to p cores 4-7 That seems unlikely, though I don't have an M1* device what feels more likely is that the two clusters are independently powered (the turbo-ing works on a cluster basis so that's almost certain), therefore the machine favors fully loading the first one before it starts "spilling" to the second (and thus has to power it up).
- uudecoded 5y agoI am curious, since Intel is relying more and more on P and E cores as well, is there any reference or research available for optimizing multithreaded userland process tasks with varying QoS? A lot of the pthreads books I see are from the late 90s. Is there a more recent reference? What's the best way to write cross-platform (e.g. not Grand Central Dispatch) multithreaded apps with these new chip architectures?
- wmf 5y agoI think people are just now starting that research and blog posts like this one are all we have so far.
- dev_tty01 5y agoAsymmetric multiprocessing has been a big topic of research for many, many years. https://scholar.google.com/scholar?hl=en&as_sdt=0%2C33&q=asymmetric+multiprocessing&oq=asymmetric+mul https://scholar.google.com/scholar?hl=en&as_sdt=0%2C33&q=asy...
- wmf 5y agoYeah, but I'd bet 90% of that research makes wacky assumptions that don't apply to real processors. When real hardware becomes available you start over from scratch. (Source: I am a former CS researcher.)
- seniorivn 5y agoisn't arm big little architecture the norm in a wildly used processors for a decade or so?
- sydthrowaway 5y agoMy understanding is as long as you specify the QoS currently, GCD takes care of it (as it has done for Apple Ax SoCs on iPhone).
- dwaite 5y ago
- yeldarb 5y agoMine just arrived today and I'm blown away. After upgrading our app's dependency chain to run native arm64 builds (which took a bit of googling) webpack's incremental build time on our app is down to 116ms from 983ms on my 2016 Macbook Pro. Over 8x faster. Our Tensorflow.js (webgl) models that previously ran about 8fps in the browser are now running at the camera's native 30fps framerate. (Similar to my desktop 2080ti.)
- Gigachad 5y agoAnd if its anything like my regular M1, it does that while remaining cold. I used a 2019 macbook pro and the thing would be roasting hot if you had docker running.
- skavi 5y agoThe regular M1 uses far less peak power than the M1 Pro/Max. They’re definitely still far more efficient than competitors though.
- asdff 5y agoCan you overclock the regular m1 then?
- saagarjha 5y agoI've gotten M1 Pro to heat up and require active cooling, unlike M1. It's certainly no space heater like my Intel Core i9 work laptop, but you can push it pretty hard.
- lordnacho 5y agoGot mine last week and not only is it faster, the power usage is crazy good. I took it to town for the first time yesterday, thought I'd do some coding between meetings for a few hours. Battery is still over 90%. I actually don't know what the thing sounds like with fans on, because it hasn't happened yet. My old machine from 2019 would have been down to half power at least, and the fans used to come on with just two monitors plugged in and not much workload.
- aroman 5y agoIn my usage, which is game development with Unity, I haven't really been particularly impressed with the performance improvements on my 10-core M1 Pro machine. I'm sure most of the blame is due to Unity (the app is a buggy mess), but it definitely took the initial "wow this is magic" down to reality very quickly for me.
- zibzab 5y agoUnity is slow on any machine. With that said, there is a huge M1 hype on HN. It reminds me of the 64-bit ARM hype when someone benchmaked encryption (because ARMv8 has AES instructions) and everyone were made to belive that the new iPhone would be 3-4 times faster.
- BoorishBears 5y agoI've seen little to no hype based on synthetic benchmarks. M1 macs are just really damn fast for daily usage. My M1 Macbook Air blows my 16" out of the water in every conceivable way. I'd elaborate but there seriously isn't any dimension of it that isn't an improvement. Maybe the smaller display?
- perryizgr8 5y ago> My M1 Macbook Air blows my 16" That is the issue though. Intel macbooks were notoriously badly cooled. That hurts the performance a lot. If all you have used is that, of course you find the new one a lot better. However, compared to any other high end Intel/AMD laptop, you would not perceive this much difference in speed. Pick up a Macbook 16 and XPS 15 from 2019 with the same specs (CPU/RAM) and the XPS absolutely runs circles around the macbook when it comes to general responsiveness in UI. All because of cooling.
- kllrnohj 5y agoAnd weirdly one of the major design changes for the new M1 Pro & Max machines were significantly improved cooling. Weird how suddenly Apple cared about airflow when it was their 100w SoC in the machine instead of Intel's....
- sydthrowaway 5y agoThese laptops are too expensive. What the hell is the markup on these things
- sumnuyungi 5y agoCompared to what? Their pricing is in-line with the previous Intel models and are a huge upgrade in performance. For some tasks, you have to look at a desktop configuration for comparable performance (at a much higher price).
- BoorishBears 5y agoThese aren't too expensive, the M1 Macbooks are just really good values. Before the M1 Air an entry-level MBP was the only "serious" choice for productivity. Even if you weren't a professional, an Air could be limiting. Now we have an $800 laptop that can compete with the previous MBPs. The MBP can actually be a pro-only machine.
- DrBenCarson 5y agoI believe these are actually one of the lowest markups Apple has and in general a great deal. Consider this: 1. 14” or 16” custom miniLED display 2. Anodized one-piece aluminum clamshell 3. Completely custom SoC (much more expensive than buying Intel’s latest, at least in the short-term) 4. Completely custom OS with a full suite of native apps 5. Custom MagSafe adapter 6. Thunderbolt 4 (3 of them) Now consider that this laptop outperforms most desktop workstations that cost $4k+…and does it on 16h of battery life. If you’re in the market for a high-performance computer, you’d have to be dense to not buy one of these (or building x86 firmware)
- whalesalad 5y agoI can’t wait for the desktop-class Pro hardware. If you extrapolate on power consumption alone it would be a complete monster. They could do something gnarly like fit 8 sockets on a single board and send your compute ability to level 9000. I’m hoping for a mid-tier unit though - a rebirth of the G4 Cube as a “Mac Mini Pro” I daily drive an M1 mini now and abuse the hell out of it. I’ve never heard the fan. It’s a remarkable device. I was installing a Homebrew package the other day on my i7 MacBook Pro and the fans kicked on - it almost startled me - I had forgotten about active cooling.
- Groxx 5y agoI managed to get the M1 fans running pretty enthusiastically with my first attempt: doing a normal build in a monorepo :| But performance, even with half the ram, was still stellar. Everything remains speedy, even when it's doing far too much.
- DSingularity 5y agoThey will skip desktop and ship server chips. If the trends hold, they will have TCO savings which no cloud provider can resist.
- RealityVoid 5y agoI keep hearing this on HN and I do not buy it. The initial cost is much much greater than the competition and even at lower power draw, I don't see it being competitive. And I don't think Apple would even want to be competitive in this segment.
- Zababa 5y agoI know that AWS already have their own ARM chips: https://aws.amazon.com/ec2/graviton/ https://aws.amazon.com/ec2/graviton/. I don't know about chips enough to understand how large is the gap between AWS and Apple when it comes to designing chips though.
- BackBlast 5y ago
- qaq 5y agoWhy no 8k support :( Apple always has to hold something back
- qaq 5y agoThe GPU obviously can support it god forbid people buy UP3218K instead of Pro Display XDR
- ReaLNero 5y agoI have an M1 Max. If anyone wants me to benchmark any workflows, I'd be glad to help!
- pvinis 5y agoMaybe you could try this one? https://github.com/ashfurrow/xcode-hardware-performance https://github.com/ashfurrow/xcode-hardware-performance
- KptMarchewa 5y agoDo you have two 4k displays? If yes, then does running them affects performance significantly? On i9 MBP this was causing extremely severe performance issues, to the point where processor was throttled to 30% performance.
- EricE 5y agoBiggest reason to get 64GB of RAM if you intend to do this - the Unified Memory is shared between the CPU and GPU. More pixels in total that the machine has to render, the more RAM will be needed for the GPUs. Don't skimp on RAM or buy into the hype that ARM halves the amount of required RAM - it's total BS.
- ReaLNero 5y agoI only have one 4k display, and I don't notice any difference in performance.
- codeisawesome 5y agoI run two 5Ks and my 2019 Intel feels quite sluggish. I still pay that price for the real estate.
- KptMarchewa 5y agoThat's what I would love to avoid.
- ReaLNero 5y ago
- brailsafe 5y agoYa, but has anyone tested it's capability of running the Instagram website or Reddit!?
- 00deadbeef 5y agoAny devs here switched from an old 15.4" MBP to a new 14" MBP? I'm curious how you find the screen size for doing work? My 15.4" can scale to "looks like 1920x1200" and the 14" to "1800x1169" so there's not a lot of difference. Does the smaller size make it feel cramped?
- samcolson4 5y agoI switched - regretting it and may return mine in favour of the 16". As great as the 14" is (amazing machine), the loss of screen real estate has been felt.
- q-base 5y agoCan you help me understand what that means? Is it not possible to run it at the "full" resolution? Will it be scaled to what equates to the 1800x1169 that you mention? That would equate to less "real estate" than what I have on my T460s FHD screen? Of course the screen would be light years ahead to look at. But just sounds a little bit strange with regards to actual pixels/space on the screen?
- masklinn 5y ago> Can you help me understand what that means? Is it not possible to run it at the "full" resolution? Will it be scaled to what equates to the 1800x1169 that you mention? Yes, Apple doesn't provide for running at the native resolution out of the box (utilities like SwitchResX make those options available, though I don't know if they work with the M1s yet), all the display modes they propose are scaled. Historically there was a default mode at 2x scaling with 2 above and 2 below, apparently on the M1s the default 2x scaling is the second-to-last and you only have one mode above. So on the 14" with a physical 3024 x 1964 display, the modes provided natively by macOS are: 1800 x 1169 (1.68x) 1512 x 982 (2x, default) 1352 x 878 (2.24x) 1147 x 745 (2.64x) 1024 x 665 (2.95x) Don't ask me why they selected these wonky-ass resolutions aside from the 2x one though. The 1024 one could make some sense but it seems like adjusted luck: the 16" bottoms out at 1168 instead. Could be a question of pixel density I guess but you'd have to run the computation.
- bullen 5y agoI'm having a hard time finding mips/flops for the M1 CPU, has anyone been able to measure it?
- vmception 5y agoHow is the M1 Max treating you all, heat wise?
- circa 5y agoWell I was on the fence before. Now I must upgrade. haha.
- lincpa 5y agoApple M1 just follow my "Warehouse/Workshop Model". Its mathematical prototype is the simple, classic, vivid, and widely used in social production practice, elementary school mathematics "water input/output of the pool", Therefore, it is scientific. The "von Neumann architecture" used by computers now has no mathematical model support. So it cannot prove its scientificity. My theory rebuilt the theoretical foundation of the IT industry, The IT industry is fundamentally associated with mathematics. Solved the most fundamental and core major problems in the IT industry. Therefore, my "warehouse/workshop model" will surely replace the "von Neumann architecture" and become the first architecture in the computer field, and it is the first architecture to achieve a unified software and hardware. Software and hardware are factories that manufacture data, so they have the same "warehouse/workshop model" and management methods as the manufacturing industry. In the IT field, only it and binary system fully comply with these 5 aesthetics: Simplicity, Unity, order, symmetry and definiteness. It has a wide range of applications, from SOC to supercomputer, from software to hardware, from stand-alone to network, from application layer to system layer, from single thread to distributed, from general programming to explainable AI, from manufacturing industry to IT industry, from energy to finance, from the missile's "Fire-and-Forget" technology to Boeing aircraft pulse production line technology, from myth to Transformers. Link: The Grand Unified Programming Theory: The Pure Function Pipeline Data Flow with Principle-based Warehouse/Workshop Model https://github.com/linpengcheng/PurefunctionPipelineDataflow https://github.com/linpengcheng/PurefunctionPipelineDataflow Why my "warehouse/workshop model" can achieve high performance and low power consumption (take Apple M1 chip, Intel AVX-512, Qualcomm as examples): https://github.com/linpengcheng/PurefunctionPipelineDataflow/blob/master/doc/why_wwmodel_fast_en.md https://github.com/linpengcheng/PurefunctionPipelineDataflow...
- mmcnl 5y agoI'm curious: how is Docker performance with x86 images? Can anyone elaborate?