News ARM Cortex X3/Cortex A715 announcement

[soresu]

ARM just announced their 2022 slate of CPU + GPU silicon IP yesterday, including the codenamed Makalu-ELP Cortex X3, Makalu Cortex A715 and a power optimised rehash of A510 (codenamed Klein).

It appears that a true successor for A510 codenamed Hayes is scheduled for 2023 to coincide with Hunter/Hunter-ELP big/bigger cores, and 2024 big cores codenamed Chaberton, all it seems in keeping with the mountain naming scheme.

There is also the obligatory new GPU IP Mali G715, and a new flagship GPU branding with Immortalis G715 - this flagship also has ray tracing on top of the Mali feature set.

Cortex X3 info link.

Cortex A715 info link.

Cortex A510 rehash info link.

Mali/Immortalis G715 info link.

 

[DisEnchantment]

Wow, I did not know ARM removed all the 32 bit stuffs on X2 and X3, good job

According to Chris Abernathy, the core’s chief architect. The process of eliminating 32-bit and optimizing for the 64-bit ISA exclusively has been a 2-step process. With the Cortex-X2, the underlying circuitry used for handling 32-bit architectural-related elements was removed, saving on transistors and simplifying some structures.

 

[DrMrLordX]

No real update to the A510 until mid/late 2023? Well okay, I guess it won't hurt too bad but. Hmm. The performance delta between The biggest/big cores and little cores continues to grow.

 

[soresu]

No real update to the A510 until mid/late 2023? Well okay, I guess it won't hurt too bad but. Hmm. The performance delta between The biggest/big cores and little cores continues to grow.

Better than the ridiculously long time between little core updates A53 -> A55 -> A510 at least.

2 years is easily the shortest time in a long while for that - hopefully it's more impressive a change than A710 -> A715 and we get 15+% IPC and even better perf/watt increases.

 

[DrMrLordX]

Better than the ridiculously long time between little core updates A53 -> A55 -> A510 at least.

2 years is easily the shortest time in a long while for that - hopefully it's more impressive a change than A710 -> A715 and we get 15+% IPC and even better perf/watt increases.

Okay, you have a point. A510 was pretty disappointing overall, and it took a long time to get there.

 

[soresu]

Found a proper 2024 roadmap graphic on PC World's website and altered it a bit for moar legibility....



Presumably Titan and Krake are the 2023 and 2024 GPU codenames respectively and Tower is a codename for ARM's next gen interconnect.

The linking of Hayes with 2 gens of big cores implies it should be matching them for the same v9-A ISA increment/feature set and likely breaking from the TCS21 and 22 gen CPU IP ISA.

 

[Panino Manino]

A510 with "up to 5% power reduction"... my god, how decadent Qualcomm is.
Their big core still can't catch up with Apple, and they came up with these SOCs using three different cores that it seems it's too much work for them to be able to properly update at the same time. The A715 is only 5% "performance uplift".

 

[soresu]

The A715 is only 5% "performance uplift".

Yeah overall this generation appear to be a wash.

The 6 wide upgrade in X2 -> X3 being the only notable feature of the whole lot CPU wise, and even that doesn't bring particularly notable performance benefits.

I'm still not sure who designed what here - for the moment I'm just going to hopefully assume that X3 and A715 are still Austin's work, because this is Sophia-Antipolis' worst output ever if it is theirs 😭

Here's hoping that Hunter and Hayes next year are more impressive.

So far the move to v9-A hasn't exactly been accompanied with great IPC improvements, though I suppose that the +30% IPC and 2x SIMD of v8-A's X1 was a hard act to follow.

 

[soresu]

Found a higher quality picture of the 2024 roadmap and cropped the fluff:

 

[Thala]

So far the move to v9-A hasn't exactly been accompanied with great IPC improvements, though I suppose that the +30% IPC and 2x SIMD of v8-A's X1 was a hard act to follow.

Excuse my ignorance, but they achieved a 29% IPC increase from X1->X3 - are you complaining about the missing 1%? Do you know anyone else, who has achieved close to 30% IPC increase over the last 2 years? Certainly neither AMD, Intel or Apple.

 

[soresu]

Excuse my ignorance, but they achieved a 29% IPC increase from X1->X3 - are you complaining about the missing 1%?

That's the part that confuses me.

The wikichip article has a slide with only 11% IPC gain for X2 -> X3:

But the widely distributed PR slide has more than double that:

I'm sure that you can understand my confusion at the mixed signals they are giving off here 🤔

 

[Kryohi]

Excuse my ignorance, but they achieved a 29% IPC increase from X1->X3 - are you complaining about the missing 1%? Do you know anyone else, who has achieved close to 30% IPC increase over the last 2 years? Certainly neither AMD, Intel or Apple.

Nope, the ipc from X1 to X3 only increases by 22%.
In my opinion it's not that bad though, considering the X1 already had an ipc close to Zen 3.

 

[Lodix]

I think people are not considering very important points.

ARM updates their architecture yearly. I don't think over 10% IPC improvements every year can be considered bad.

This predictions of performance are at ISO PROCESSES.

When AMD or Intel publish their performance targets are combined with the new manufacturing in mind.

This improvements of performance come with an increase of efficiency. They can give better performance at ISO POWER and ISO PROCESS.

And 20% less power consumption at same performance in the same node in 1 year for mobile first IP is a huge improvement.

The projected performance at ISO PROCESS is over 20% combining IPC + frequency at similar power. For a 1 year refresh architecture.

This combined with the improvement in manufacturing that different vendors will choose will mean a very healthy improvement in just 1 year. Notably in efficiency for mobile SoCs next year. Samsung has just announced their production of 3nm with big power and performance improvements.

So if they meet their claims I am happy with what has been delivered for next year smartphones.

 

[soresu]

Nope, the ipc from X1 to X3 only increases by 22%.
In my opinion it's not that bad though, considering the X1 already had an ipc close to Zen 3.

Thanks for that, which article did you find that slide in - was it the wikichips or something else?

ARM have been infuriatingly scattershot with this PR, it seems all over the place.

 

[coercitiv]

The wikichip article has a slide with only 11% IPC gain for X2 -> X3:

But the widely distributed PR slide has more than double that:

Trying to understand this as well and my interpretation is as follows:

• IPC uplift is 11% with ISO numbers for the entire memory subsystem (L2, L3, memory latency and bw)
• performance uplift is 25% when adding frequency increase for the core and also potentially better memory subsystem

The marketing slide is still extremely vague since I have to do all the work to pull the additional right numbers for the comaprison to make sense (expected X2 clocks for real silicon, difference in memory performance if any... etc)

 

[soresu]

ARM updates their architecture yearly. I don't think over 10% IPC improvements every year can be considered bad.

I'm really measuring against my frustrations at ARM steadily falling further and further behind Apple coupled with the extremely slow release of new cores on ARM SBCs😅

I love to root for the underdog, but there is a point where I just run out of woo and lapse into pessimism 😆

 

[Thala]

It is X1->X2 (16%) here
and X2->X3 (11%) here

for overall 29% based on SpecINT

 

[DrMrLordX]

A510 with "up to 5% power reduction"... my god, how decadent Qualcomm is.
Their big core still can't catch up with Apple, and they came up with these SOCs using three different cores that it seems it's too much work for them to be able to properly update at the same time. The A715 is only 5% "performance uplift".

It's an ARM reference design, it's not exclusively Qualcomm's baby. Plus I expect Qualcomm to have their own Nuvia-enriched IP soon.

That being said, ARM Ltd. is being very conservative. They are clearly not trying to catch up to Apple.

 

[Thala]

I'm really measuring against my frustrations at ARM steadily falling further and further behind Apple coupled with the extremely slow release of new cores on ARM SBCs

That being said, ARM Ltd. is being very conservative. They are clearly not trying to catch up to Apple.

Not sure, what ARM is trying relative to Apple, as their design goals are different - however they are de-facto catching up to Apple, as ARMs IPC uplift has been greater than Apples over the last 2 years. Apple did not have anywhere close to 30% IPC uplift over the last 2 years.

I wonder what numbers you have, which would suggest otherwise?

ps. I am also frustrated about ARM SBCs - but as i mentioned in another thread, the SoC manufacturers for SBCs are partially still on TSMC 28nm.

 

[DrMrLordX]

Apple did not have anywhere close to 30% IPC uplift over the last 2 years.

Depends on which core you're talking about. "small" cores? Definitely not, Apple is still improving those steadily. "Big" cores though? A78-A710-A715 isn't really a big deal, and nobody's using multiple X cores per chip (yet) so I'm not sure how to do a comparison there.

 

[Thala]

Depends on which core you're talking about. "small" cores? Definitely not, Apple is still improving those steadily. "Big" cores though? A78-A710-A715 isn't really a big deal, and nobody's using multiple X cores per chip (yet) so I'm not sure how to do a comparison there.

We compare the fastest cores if you are interested in "catch-up'. What the device manufacturers integrate is up to them - DSU-110 allows 8 X-cores. Also 8CX Gen 3 integrates 4 X cores.

 

[Thala]

That's the part that confuses me.

But the widely distributed PR slide has more than double that:

I'm sure that you can understand my confusion at the mixed signals they are giving off here 🤔

That is consistent, when you read the fine-print. The 22% includes a 10% frequency uplift. Its +11% IPC and +10% frequency - they essentially compare devices clocked 10% higher.

 

[Lodix]

IPC numbers at ISO PROCESS and memory for the X3.

 

[DrMrLordX]

We

We?

compare the fastest cores if you are interested in "catch-up'.

It's valid to examine the Medium and Small cores as well. The bulk of reference ARM device performance will still be determined by the Medium cores (A78/A710/A715). You could and should definitely compare those to Thunder->Firestorm->Avalanche.

What the device manufacturers integrate is up to them - DSU-110 allows 8 X-cores.

And yet . . .

Also 8CX Gen 3 integrates 4 X cores.

. . . the only manufacturer that seems interested in using more than one X-core per SoC is probably moving away from reference designs.

 

[Thala]

It's valid to examine the Medium and Small cores as well. The bulk of reference ARM device performance will still be determined by the Medium cores (A78/A710/A715). You could and should definitely compare those to Thunder->Firestorm->Avalanche.

Whatever you compare, the IPC gains of ARM are larger over the last 2 years, than Apples - in fact Apples IPC gains are single digit.
If the device manufacturers are doing reasonable decision with respect to choice of cores is another question. It could very well be, that for Phones going with 6 cores instead of 8 but using faster ones might be the better option.