New Arm cores deliver up to 40% better performance
phonescoop.com
phonescoop.com
"Arm Unveils 2023 Mobile CPU Core Designs: Cortex-X4, A720, and A520 - the Armv9.2 Family"
https://www.anandtech.com/print/18871/arm-unveils-armv92-mob...
( posted in https://news.ycombinator.com/item?id=36109734 )
I keep thinking if I should dig up and repost all the bookmarked link from people who keep blasting Apple is and always will be 5 years ahead of everybody. And these are prominent member in their programming / tech circles.
I hope Anandtech does another detail article on their GPU as well. While I am not a fan of Mali but still want a detailed analysis on Immortalis.
Another question worth exploring would be how good are the other IPs from Mediatek. From Modem, I/O, DSP, NPU as compared to Snapdragon. As Mediatek is now competing directly with Qualcomm in flagship SoC.
the article is about the performance of one of the reference designs.
This differs from what happened during the last decade when Cortex-A53 and Cortex-A55 have been the only little cores implementing respectively Armv8.0-A and Armv8.2-A, while new big or medium-size cores were being introduced each year.
This is good, because, as I have seen just yesterday at a link provided by someone here on HN, a set of measurements made on many smartphones has shown very convincingly that unlike Cortex-X2, which was indeed better than Cortex-X1, both Cortex-A710 and Cortex-A510 have regressed a lot in power efficiency in comparison with the older Cortex-A78 and Cortex-A55, when made with the same manufacturing process, despite the contrary claims made in the presentations of the Arm company.
Replacing Cortex-A510 looks like an admission that it sucked. Hopefully Cortex-A520 and Cortex-A720 are the first decent implementations of low-power Arm cores with SVE2, the first that are truly better than the old cores with Armv8.2-A.
EDIT: Going through all the presentation slides at AnandTech shows that they have bumped the implemented architecture to Armv9.2-A.
This would have required a new little core anyway.
Nevertheless, they have reverted some of the microarchitectural changes originally made between Cortex-A55 and Cortex-A510, so it is likely that the introduction of a new little core has been driven both by the desire of implementing the additional features of Armv9.2-A and by the need of achieving acceptable power efficiency.
Maybe rumors of TSMC's N3 troubles have been exaggerated. Still, it seems unlikely that Spring 2024 will see X4 cores. It seems like there's a decent likelihood that Qualcomm will push back their Snapdragon 8 Gen 3 by 6 months to get the X4 cores (or create a Gen 3 with smaller enhancements). If TSMC's N3 is ready for Fall 2023 and N3E is a year behind that, we aren't going to see X4 cores until Fall 2024.
It sounds like ARM is targeting a TSMC process that's farther out than normal - or at least seems farther out based on the information TSMC has given to the public.
Apple to occupy 90% of TSMC 3nm capacity in 2023
https://news.ycombinator.com/item?id=35947339 (219 comments)
2. Snapdragon 8 Gen 3, so expect the usual 2024 flagship phones, early 2024 Smartphone release.
My understanding was that Amazon/Qualcomm/Mediatek/Google use ARM's cores but that Apple had their own core design and just uses the ARM architecture under special license.
AirTags have an nRF52xxx SoC by Nordic. The only Apple silicon in them is the U1 UWB chip (which is also present in iPhone, HomePod, Watch, ...) which runs on a 16nm Cortex M4.
Recent AirPods run on H1 / H2 SoCs though, which are custom Apple chips. Wouldn't be surprised if they are built around the E-cores from the iPhone SoC.
I read it as: The performance for some tasks is 40% better, but for other tasks might be 90% worse.
So is that a good thing or a bad thing? Why didn't the author give me the average performance instead?
And the -90% is very unexpected from the context, so that's a possible, but not a proper reading
You don’t want to say 40% faster and have the tech crowds harping on you because a single benchmark performs at 25% faster.
Especially when it comes to silicon lottery, thermal designs etc causing variances
“Up to” is a get out of jail free card on liability and pretty much every product will use that unless they’re specifically talking about a singular case.
15% more IPC looks very impressive.
"Arm is claiming that, based on its pre-silicon simulation numbers, the Cortex-X4 will deliver a 15% IPC uplift at iso-frequency and iso-bandwidth versus the Cortex-X3 used in last year's flagship Android SoCs"
(link above)
https://www.anandtech.com/print/18871/arm-unveils-armv92-mob...
Yes, but that's because Qualcomm and Samsung are about as incompetent at CPU design as early-2010s AMD was; their products have been 3-4 years behind the leading edge for the past decade.
This might change with Qualcomm's acquisition of Nuvia, but they're still going to take a few years to be comparable with what you can get in a 400-dollar iPhone/iPad and be unavailable in the sub-1000 dollar Android market (just like the Snapdragon 888- whose performance is on par with the 2nd gen iPhone SE- is today).
Sadly this will probably mean some apps will be able to steal more CPU time for doing telemetry and serving ads because the user won't notice a difference in response time and performance.
The Anandtech breakdown (https://www.anandtech.com/print/18871/arm-unveils-armv92-mob...) goes into detail on this. While all the cores are claimed to perform better at the same power draw, they draw less power if they have to perform at the same level as the previous generation.
Smartphones will use a combination of big (X-4), medium (A720) and small (A520). Since the small and medium cores are more performant and use less power, that opens up opportunities for battery savings. There could be loads where the larger core remains asleep while the smaller ones finish the job, since they’re more capable. Or the same cores do the job, but use less power.
So if you buy a flagship in a couple of years, the CPU will consume less power. The battery tech will have progressed as well. However, the screens might be higher res or higher refresh rate (120Hz instead of 60Hz). Or it might use a 5G radio that needs more power. So the net effect is that battery lasts about a day.
phone that couldn't last full day, lasted over 3 with about maybe 40% battery (idle and stationary but connected to wifi)
My Android is 8 years old and it was flagship at the time and I also don't remember anything being slow on it.
I have also low end Android phone - which is 3-4 years old, it was cheap like $50-$100 range and on that phone nothing really works and even when it was new it was quite slow anyway. But I keep it as a backup in case my main dies and I need some phone to put sim card in to be able to call people.
Ascalon is a RISC-V microarchitecture expected to be released in 2024, with similar performance to projected AMD Zen5 (also 2024), but lower power consumption.
If ARM is still trying to catch up to Apple, they'll be leapfrogged by RISC-V, as early as that.
ARM has very successful architectures and cores already in the market. The current iteration of the Snapdragon 8cx does a pretty good job in general purpose computing scenarios [0].
There are no RISC-V cores in the market to compete with it.
Vendors can and will promise breakthroughs every now and then, that is just how marketing works. It happened before with things like Graviton, and of course Tenstorrent is doing it now.
[0] https://www.notebookcheck.net/SD-8cx-Gen-3-vs-M1-vs-M2_14600...
Depends on what "it" is.
For the whole Cortex-M line and A53/A55, SiFive has had designs with better power-performance-area for years now.
For the higher performance A70 line, SiFive has better designs competing up to A76.
There's several other vendors, with their own competitive RISC-V cores.
>This is pure speculation.
Not when it's based in announcements and talks by the relevant companies.
>It happened before with things like Graviton, and of course Tenstorrent is doing it now
Wait, why do you think Tenstorrent will not deliver? Do you actually have anything?
I have no reason to believe that strong teams that have successfully delivered competitive microarchitectures in a range of companies won't do it with RISC-V now that they're with Tenstorrent, MIPS, Rivos, Ventana, Qualcomm or any others.
Ventana's Veyron and Tenstorrent Ascalon are imminent. The other companies are less public about what they're doing.
Any one of these efforts could be delayed or underdeliver. All of them? That seems highly unlikely.
RISC-V is inevitable.
Are RISC-V “better designed” than x86 processors, even when the latter are overwhelmingly more popular and faster than the former?
> Not when it's based in announcements and talks by the relevant companies.
Again, this is exactly what marketing is. I cannot begin to count how many companies formed by wildly successful leaders of the industry, have failed to deliver.
> Any one of these efforts could be delayed or underdeliver. All of them? That seems highly unlikely.
Core design and implementation takes years to complete.
The supposition is that a handful of relatively new companies will take the processor market by storm in a year or two, with less than a decade of work behind them, and a tenth of the budget, compared to companies like AMD or Intel.
That is speculation if I ever seen any.
I mentioned it: power-performance-area.
In the case of SiFive, they seem to like comparing each of their designs with a specific one from ARM. They typically manage to perform better, use less area and consume less power, at the same time, relative to the ARM design they picked.
>Core design and implementation takes years to complete.
These years already happened, these are several years underway. What follows is the products that we are about to see.
>and a tenth of the budget
There's several billions already invested. Source: RISC-V Foundation's main talks at RISC-V Summit last December.
In 2021, the theme was acceleration. In 2022, it was momentum, because it is already there. RISC-V is huge.
RISC-V is inevitable.
>The supposition is that a handful of relatively new companies will take the processor market by storm in a year or two, with less than a decade of work behind them, and a tenth of the budget, compared to companies like AMD or Intel.
Exciting times to live in.
I expect AMD and Intel to survive, and find a way to thrive. They can e.g. provide RISC-V CPUs with excellent x86 acceleration, making them an easy choice in markets that depend on x86 software still.
I do not have faith ARM will be able to adapt, due to their management's recent behavior. They are dinosaurs. Won't know what hit them.
It’s like saying that the better architecture is the one with available CPUs manufactured on the smaller node and with the higher CoreMark score.
> Won’t know what hit them
This is still a extremely naive take.
Qualcomm and Apple are not going to ditch ARM any time soon, and there is no market where RISC-V currently has a meaningful share. Hell, there are more x86 SBC out there than RISC-V by a double digit factor.
What I’m saying is that RISC-V is a hypothetical contender until manufacturers adopt it.
Of course, I will eat my hat if we get competitive RISC-V cores across the market in less than five years, but I’m confident that will not be the case.
Just two notes:
- Qualcomm is a top-tier RISC-V member and has an ugly lawsuit going on with ARM about Nuvia's IP. They did multiple presentations in RISC-V Summit where they were very clear about their intent to move to RISC-V.
- Apple has been hiring RISC-V experts for a while (now years). Keywords included performance and optimization.
So do not count on either to remain faithful to ARM.
>What I’m saying is that RISC-V is a hypothetical contender until manufacturers adopt it.
As per RISC-V's foundation presentation in RISC-V Summit December 2022, RISC-V had at that point already shipped 10 billion cores.
That's some serious manufacturing, and we're talking about what already happened. It's safe to assume much more of that is ongoing.
It’s any company’s duty to explore alternatives. Intel used to make ARM CPUs, Mediatek makes MIPS chips, etc.
It does not mean that it is possible to completely reverse course in less than a decade.
Qualcomm is definitely not going to dramatically reverse its core business direction to go full RISC-V in less than a decade. It makes no sense, especially after dumping $1.5B in Nuvia.
ARM’s case against Qualcomm is a licensing one. So you suggest that Qualcomm may be ditching ARM because of an argument about licensing ARM, which would make no sense whatsoever. It would be akin to Intel ditching AMD64 and adopting some exotic architecture, because of a disagreement about cross-licensing.
Apple just switched to ARM, of which they have a perpetual license. They are also getting into display production and modem design. Their aim is to control the production line to the last bit, they do not need to be compatible with anything else, so why would they ditch ARM any time soon? It would make no financial sense.
> As per RISC-V's foundation presentation in RISC-V Summit December 2022, RISC-V had at that point already shipped 10 billion cores.
There were 7 billion ARM cores shipped in the last quarter of 2020 alone, many of them inside actual consumer products, not SBCs or niche platforms. Virtually all portable, IoT, and in-car devices run on ARM. ~14% of laptops run ARM and the share is growing, same goes for datacenters. And most importantly, software support for RISC-V is still in its infancy, with no large scale consumer products currently available.
This is to say that RISC-V has a very long road ahead, and it is not going to topple ARM dominance any time soon, simply because the economics of it make no sense.
Explain the RISC-V cores that already moved.
It's inexpensive. What a feat.
Can you point at your source for that? I am not interested in speculation.
Anyway, there is a good number of ESP32 microcontrollers using RISC-V cores, replacing the aging ARM, e.g. ESP32-C3, ESP32-P4.
Get off your bubble. SBCs are a tiny fraction of the CPU market.
But custom/semicustom is huge, and is where I expect most already shipped cores would be. Most of those cores would not be user visible.
Western Digital probably accounts for a lot of shipped cores, as they were early adopters and moved their HDDs controllers to RISC-V.
I’m pretty sure nothing is “inevitable”.
I mean there are many good reasons why RISC-V still hasn't busted out yet. It's just an ISA, that's it. The implementations are wildly different, which is not something devs what to deal with. Risc-v didn't even have a rev 1 debug spec until a few months ago i mean come on get it together (what about the embedded -march that has been in limbo for 5 years). They're catching up with arm, slowly, but arm has an entire ecosystem that people trust that risc-v don't have.
ARM doesn't need to catch up with Apple. Apple hasn't been buying ARM's designs for many years and there's no chance that will change. Even if Apple is faster, the whole Android world needs ARM's designs and most customers aren't choosing iPhones because Apple has faster processors. Apple's no threat to ARM in the data center since Apple isn't interested in the market.
ARM probably isn't that afraid of RISC-V at this point. Yes, there are some companies doing good things with RISC-V, but you don't have to look far to see how hard it is to create a great processor. Qualcomm has way more resources than RISC-V startups and they bought Nuvia (with engineers who made Apple Silicon) and they still aren't pushing their own cores. When you look at software benchmarks on ARM, there's often some lag (compared to x64) due to less mature compiler optimization and that's going to be even worse for RISC-V. While a lot of software is portable, it's not a drop-in replacement. A lot of Linux distros aren't even supporting ARM that well at the moment.
None of this is meant to belittle RISC-V. It's simply meant to note that there's a big hill to climb and it takes a long time to climb that hill. Even Apple who control their whole ecosystem took a very long time. Apple started shipping their own designs in 2010 and it was a decade before they made their way into MacBooks. Apple has the kind of money and ecosystem control that make such a move easier. That's not to say that RISC-V can't make a big impact, but rather to note that the threat from RISC-V is probably significantly farther out.
There are always threats, but the big threat right now is Qualcomm (and others) moving away from ARM's designs. If they're able to just pay instruction set licenses instead of the core licenses, it'll dry up ARM's revenue. As ARM's revenue dries up, they won't have the money to invest in new core designs. That will accelerate the decline of ARM's revenue as it becomes easier to beat their core designs.
It has also been available for a while (years).
Today, it is found in the cheap VisionFive 2, Star64 and Milk-V SBCs.
It is nice to see competition to ARM in the embedded market, but it is still far behind.
[0] https://www.elektroniknet.de/international/from-microcontrol...
Google is one of the X280 licensees, already using the resulting chips, where they paired the X280 with a specialized accelerator, using a very tight interface between accelerator and cpu, co-designed with SiFive.
P670 and P470 are newer, vector-enabled higher performance designs.
This all reads like moving goalposts to me.
Best I could find is that boards based on either A55 or U74 have a power draw to wall of ~5W under load, with negligible differences in performance.
I learned about it from press releases back when it launched. Here's the first one I found with 15 seconds of Google effort[0]. Relevant page is 5.
SiFive has since revised U74's design multiple times providing even higher performance on the same envelope[1].
>Best I could find is that boards based on either A55 or U74 have a power draw to wall of ~5W under load, with negligible differences in performance.
SBCs typically have hardware besides the SoC, and the CPUs themselves constitute a small part of the draw of the SoCs used, which in the first place could be fabbed on different nodes.
They do not make for very useful comparisons, e.g. my pi4 and VisionFive2 have virtually the same power draw, but the VF2 has a M.2 SSD mounted, whereas the RPi4 does not support USB-PD, so there should be further losses in voltage conversion. Too many variables hide CPU power efficiency.
0. https://c.a.segger.com/fileadmin/documents/Press_Articles/20...
1. https://linuxgizmos.com/sifive-speeds-up-risc-v-u74-cores-as...
Alright then.
40% performance gain is ZERO if you have to feed it 40% more electricity.