Discussion Future ARM Cortex + Neoverse µArchs Discussion

[FlameTail]

Do you guys reckon we will see ARMv10 sometime this decade?

And will ARMv10 be a major upgrade (like ARMv8) or a minor upgrade (like ARMv9) ?

 

[soresu]

All eyes are on Rockchip

RK3588 was first announced in 2019, it's certainly more than due a successor.

 

[Panino Manino]

Mediatek isn't the only culprit. Samsung is still using A78 cores for their midrange SoCs.

Exynos 1280 : 2×A78 + 6×A55
Exynos 1380 : 4×A78 + 4×A55
Exynos 1480 : 4×A78 + 4×A55

3 generations of SoCs using the same cores. I am beyond outraged.

I would be somewhat sated if they had put an X1 in the 1480. But they didn't. The poor ST performance of the A78 really hurts the experience in the midrange Galaxy phones, becuase Samsung's OneUI android skin is really heavy and needs strong ST performance to have a smooth experience.

More than enough.
The only problem I see is that ARM still don't have a new small core! How is this possible? Is ARM really unable to create a better small core that can compete with Apple has?

edit;

 

[soresu]

Do you guys reckon we will see ARMv10 sometime this decade?

And will ARMv10 be a major upgrade (like ARMv8) or a minor upgrade (like ARMv9) ?

It's not just v8 or v9, it's v8-A or v9-A.

The ISA for the M type is separate (v8-M has its own SIMD called Helium).

As for v10-A, it's entirely possible that it won't happen and v9-A will just stretch on in perpetuity.

At this point there is little low hanging fruit left, excepting perhaps ISA extensions that would make it easier to run non ARM apps on ARM CPUs, for which solutions already exist on v8-A with Apple Mx.

 

[Shivansps]

Mediatek isn't the only culprit. Samsung is still using A78 cores for their midrange SoCs.

Exynos 1280 : 2×A78 + 6×A55
Exynos 1380 : 4×A78 + 4×A55
Exynos 1480 : 4×A78 + 4×A55

3 generations of SoCs using the same cores. I am beyond outraged.

I would be somewhat sated if they had put an X1 in the 1480. But they didn't. The poor ST performance of the A78 really hurts the experience in the midrange Galaxy phones, becuase Samsung's OneUI android skin is really heavy and needs strong ST performance to have a smooth experience.

A78 is fine... heck... even A76 is decent enoght, see how many devices are using dual/quad intel small cores before ADL-N. They perform around that level.
The problem is A55... and all ARM small cores in general, they are criminally slow.

Btw, Exynos 1380 has a Mali G68 and the 1480 a RDNA2 one... i do not need to mention how big it is going from a 1st gen Valhall to asomething i can actually call a GPU.

 

[soresu]

The only problem I see is that AMD still don't have a new small core!

Atom is not exactly blowing anyones socks off for efficiency.

AMD's strategy of taking their existing performance core and optimising it for area and power efficiency to cater to hyperscaler (lots and lots of threads) customers in the cloud business seems to fit the bill for the time being.

Intel had to gimp their performance core's AVX512 execution just to prevent problems with code migration in their big/little config because they didn't bother to plan and implement AVX10 prior to Alder Lake.

AMD on the other hand just skipped the entire mess and ended up leaving egg on Intel's face the entire time Zen4 has been available to consumers with AVX512 support.

I'm more interested in what the future holds for AMD a la x64 only µArch, what that means for legacy 32 bit Windows software and this other new APX extension set that Intel proposed.

 

[eek2121]

Meaning what?

Advertised benchmarks don’t end up meeting reality due to clocks and/or thermal constraints.

All parts that actually make it into devices are much lower.

I am pretty sure there are articles on AT about it. It is absolutely ridiculous and it is one of the reasons I am not a fan of ARM, Qualcomm, etc. There is always something big not mentioned in their marketing materials.

Expect production versions of these chips to clock lower. My bet is 3.2-3.4ghz tops. Depending on the v/f curve they might go a bit higher. I will be shocked if any device hits advertised clocks unless it is a PC or something.

EDIT: I have to get off here now, but ARM has been resistant to even moving the performance needle much for many years. It took their partners cooking up faster designs to force them to actually focus more on performance. I suspect that is the only reason they are doing so now.

 

[soresu]

It has been a while since i used WOA, but the x86-64 emulator did not emulate AVX when I used it, hopefully Microsoft has a solution for that.

You want to avoid the emulator anyway since it eats into power efficiency

The latter is a given with emulators and power efficiency - that's why you need a much faster computer to emulate an older, slower one on a different hardware platform.

For the former WoA has its own x86-64 -> ARM emulator, I was talking about FEX-emu which is a separate open source effort employing at least one of the ARM64 devs from Dolphin Emu (Sonicadvance1).

 

[FlameTail]

It's not just v8 or v9, it's v8-A or v9-A.

The ISA for the M type is separate (v8-M has its own SIMD called Helium).

As for v10-A, it's entirely possible that it won't happen and v9-A will just stretch on in perpetuity.

At this point there is little low hanging fruit left, excepting perhaps ISA extensions that would make it easier to run non ARM apps on ARM CPUs, for which solutions already exist on v8-A with Apple Mx.

ARM could possibly switch their SIMD vector implementation to something like RISC-V's vector implementation.

I am not knowledgeable about the details, but I recall @Nothingness mentioning that RISC-V's vector implementation is better.

 

[soresu]

ARM could possibly switch their SIMD vector implementation to something like RISC-V's vector implementation.

Unlikely.

They have kept NEON through v7-A to v9-A and code development for SVE(2) is only just starting in earnest in the OSS world.

I don't see ARM going with a whole new vector instruction set in v10-A unless it is purely complementary to SVE.

 

[Nothingness]

ARM could possibly switch their SIMD vector implementation to something like RISC-V's vector implementation.

I am not knowledgeable about the details, but I recall @Nothingness mentioning that RISC-V's vector implementation is better.

If I ever wrote that, I must have been severely drunk. I don't know well enough R-V V extension to state such a thing.

 

[Nothingness]

They have kept NEON through v7-A to v9-A and code development for SVE(2) is only just starting in earnest in the OSS world.

The SVE situation at the moment doesn't look better than AVX a few years ago (or AVX-512 at the moment, thank you Intel, and kudos to AMD!). Too few devices support it (and many that do support it disable it in firmware).

 

[soresu]

Interesting points brought up by that Dolphin/FEX dev:



Same poster:

Oh interesting, X925 doesn't have the same zero latency moves and instruction fusion of A725.

 

[Panino Manino]

Atom is not exactly blowing anyones socks off for efficiency.

Wrote it wrong, not AMD but ARM.
They still don't have a new and better small core.

 

[FlameTail]

If I ever wrote that, I must have been severely drunk. I don't know well enough R-V V extension to state such a thing.

Sorry, I must have misremembered.

 

[soresu]

Wrote it wrong, not AMD but ARM.
They still don't have a new and better small core.

During v8-A they only updated it once: A53 -> A55, which took 4 years from 2013 to 2017.

Then another 4 years from 2017 to 2021 for A510 as the 1st v9-A smol core.

Comparatively they released A520 last year, so they have at least doubled their cadence for that segment, and we might expect A530 along with X930 and A730 next year, possibly even with a new ISA increment to v9.4-A or something.

Either way, by their own PR they seem to be branding A7xx as their 'efficiency' core now, I wouldn't expect them to make any big futures moves for A5xx.

 

[Shivansps]

It has been a while since i used WOA, but the x86-64 emulator did not emulate AVX when I used it, hopefully Microsoft has a solution for that.

You want to avoid the emulator anyway since it eats into power efficiency

Last time i tested it i think it was back in december or so. And no, no AVX.

Ill need to try again with a post 26100 Windows/Kernel because it has MAYOR changes for ARM64.

 

[FlameTail]

During v8-A they only updated it once: A53 -> A55, which took 4 years from 2013 to 2017.

Then another 4 years from 2017 to 2021 for A510 as the 1st v9-A smol core.

Comparatively they released A520 last year, so they have at least doubled their cadence for that segment, and we might expect A530 along with X930 and A730 next year, possibly even with a new ISA increment to v9.4-A or something.

Yes

Either way, by their own PR they seem to be branding A7xx as their 'efficiency' core now, I wouldn't expect them to make any big futures moves for A5xx.

I once said that ARM should make a Cortex A6xx, an OOO E-core design like Apple.

 

[FlameTail]

It's notable that, with Cortex X925- ARM has caught upto or even exceeded the width (IPC) of Apple's M4 P-core. For instance, X925 has more SIMD units (6 vs 4).

What ARM is seemingly lacking now, is the ability to clock their cores as high as Apple, or having Apple-like efficiency.

 

[SarahKerrigan]

Idk man. Mongoose cores were using more die area than Apple's P cores, while having similar performance to ARM's A7x cores and worse efficiency. In fact, Mongoose M6, the 2021 core (which was cancelled), was rumoured to add SMT.

SMT in phones!

Edit: corrected to M6

Huawei is shipping SMT-equipped phone SoC's today.

 

[SarahKerrigan]

ARM could possibly switch their SIMD vector implementation to something like RISC-V's vector implementation.

I am not knowledgeable about the details, but I recall @Nothingness mentioning that RISC-V's vector implementation is better.

I have looked a fair bit at RVV and did not come out particularly impressed with it. (But then, I'm a bit of a skeptic of variable-length vector extensions to begin with, and when I'm feeling contrary, I'm a skeptic of any client SIMD beyond 128b. I kind of feel like VMX already got this right the first time, in the ways that matter.)

 

[FlameTail]

Huawei is shipping SMT-equipped phone SoC's today.

Taishan CPU core.

 

[adroc_thurston]

Whew, given the specINT numbers, X5 is a total, absolute dud. Too bad!

Taishan CPU core.

It's a meme but hell yeah it exists.

 

[LightningZ71]

Yes, imagine: https://www.lantronix.com/products/snapdragon-8-gen-3-mobile-hardware-development-kit/

Now imagine the price

That price is ludicrous... An all-up phone using what is the same SOC and including all the cameras, sensors, battery, case, display etc. is roughly half the cost of just the board. I get economies of scale, but that's just milking the market...

 

[adroc_thurston]

That price is ludicrous... An all-up phone using what is the same SOC and including all the cameras, sensors, battery, case, display etc. is roughly half the cost of just the board. I get economies of scale, but that's just milking the market...

Devkits are always $$$, don't be silly