Discussion Future ARM Cortex + Neoverse µArchs Discussion

[FlameTail]

Microsoft is making it easier for developers to port/develop apps for Windows On ARM

https://twitter.com/x/status/1768314066315915482

 

[soresu]

Microsoft is making it easier for developers to port/develop apps for Windows On ARM

https://twitter.com/x/status/1768314066315915482

I wouldn't take it as revolutionary movement.

It has more of a feeling of "please bulk out our platform 🥺🥺" entreaties

 

[Shivansps]

The lack of SMT in client won't be significant enough for gaming. Vulkan/DX12 gaming leans more to SMT=Off having the highest FPS.

ARM can lean towards more energy efficiency with aiming for dead/gated units. Rather than completely filled units and thrashed regs/renames/caches/etc.

The real issue will be discrete GPUs. Since, both ARM's and Qualcomm's solutions are laptop leaning. So, it is going to hit that iGPU wall where x86-64 desktop can just slap latest phat dGPU in.

The only case for a successful ARM takeover for PC gaming is not adopting SMT but rather getting into the desktop market. Where CN gets it, Qualcomm/ARM will most likely never get it.
View attachment 95023

As well as SMT being replaced by BT [Bulk threading]. Ideal solution is >4 Cortex-X prime cores and >8 Cortex-A energy efficient dense option cores.
View attachment 95117
Where most of the back-end for games are being switched to bulk cores(Cortex A) while the front-end is handled by prime cores(Cortex X).

Getting to the "igp" wall is already good enoght to get a good portion of the market, the next Malis can probably rival RDNA2/3 IGP on the bigger configurations, just do a quick search to see how many systems are sold with less than that.

Its more of a matter of who is going to put the money to make a Windows driver and to provide Windows support as ARM does not seems to be interested.

 

[FlameTail]

https://twitter.com/x/status/1776759718893256998

Cortex X5, the new IPC king?

 

[Tup3x]

But can it clock high enough is the question. Well, not that important for mobile SoCs but to be competitive in WoA devices.

 

[gdansk]

If X5 manages power well

The biggest of ifs.
And what do you care if it's an IPC "king" if it clocks lower and uses more power than Everest to get a lower score?

 

[FlameTail]

.

 

[adroc_thurston]

The biggest of ifs.

it doesn't.

And what do you care if it's an IPC "king" if it clocks lower and uses more power than Everest to get a lower score?

yes fmax still bad. the crux of ARM issues really.

https://twitter.com/x/status/1776809351170359644

Cortex X5 in D9400.

15W phone SOC peak power

phone vendors need Jesus

 

[FlameTail]

Nah, it's about time we took these 15W SoCs out of phones and slapped them into laptops!

 

[adroc_thurston]

Nah, it's about time we took these 15W SoCs out of phones and slapped them into laptops!

QC tried (Snapdragon 7c, wasn't it?).
The I/O is too castrated to be useful outside of phones.

 

[gdansk]

QC tried (Snapdragon 7c, wasn't it?).
The I/O is too castrated to be useful outside of phones.

But MediaTek doesn't shy away from making a SoC even if it doesn't make sense.
A "fast Chromebook" SoC might even have a market.

 

[FlameTail]

Cortex X5 D9400 - 2700 GB6 @3.35 GHz doesn't sound good.

For comparison, Cortex X4 8G3 FG - 2300 @3.4 GHz.

 

[FlameTail]

QC tried (Snapdragon 7c, wasn't it?).
The I/O is too castrated to be useful outside of phones.

Snapdragon 7c was a travesty.

That thing had a gutted 32 bit memory bus.

Modern Snapdragon 8 series chips are a different class (120+ mm² die size).

 

[adroc_thurston]

But MediaTek doesn't shy away from making a SoC even if it doesn't make sense.

Oh absolutely, but they haven't tried doing Chromebook travesties much so far.

That thing had a gutted 32 bit memory bus.

It's a mainstream phone SoC and had it originally.

 

[FlameTail]

It's a mainstream phone SoC and had it originally.

*midrange phone SoC.

I was talking about flagship phone SoCs (Snapdragon 8 series, Dimensity 9000 series)

 

[adroc_thurston]

midrange phone SoC.

Mainstream.

I was talking about flagship phone SoCs (Snapdragon 8 series, Dimensity 9000 series)

Those don't have any I/O either.
Just like one or two USB roots and that's it.

 

[Tigerick]

Cortex X5 D9400 - 2700 GB6 @3.35 GHz doesn't sound good.

For comparison, Cortex X4 8G3 FG - 2300 @3.4 GHz.

X-Elite @ 4GHz scores about 2700 in GB6 1T, supposedly the same core used in upcoming 8G4.

If D9400 can achieve 2700 @ 3.35 GHz, I would say pretty impressive IPC in Cortex-X5, don't you think?

 

[Tigerick]

Mainstream.

Those don't have any I/O either.
Just like one or two USB roots and that's it.

How many more IO you need in tablet/mainstream NB, 2 USBs, SDIO, HDMI should be standard...

 

[FlameTail]

X-Elite @ 4GHz scores about 2700 in GB6 1T, supposedly the same core used in upcoming 8G4.

If D9400 can achieve 2700 @ 3.35 GHz, I would say pretty impressive IPC in Cortex-X5, don't you think?

No, the issue is that the IPC uplift is not good.

That sounds like only a ~15% IPC uplift for Cortex X5 (vs X4). Which sounds mediocre when we had this article a few months ago...

RESEARCH NOTE: Arm’s “Blackhawk” CPU Is An Audacious Plan To Have The Best Smartphone CPU Core This Year - Moor Insights & Strategy

For years now, there has been what I consider a healthy, competitive tension between Arm CPU instruction set licensees and Arm’s pe-packaged and pre-validated IP licensees. (I am sure some licensees would challenge me on “healthy” given Apple’s performance.) I think it made sense that Arm would...

moorinsightsstrategy.com

Conclusion: Either the 2700 GB6 ST leak for D9400 is fake, or ARM'S endeavour to create a "Custom ARM core killer" has failed.

 

[Tigerick]

No, the issue is that the IPC uplift is not good.

That sounds like only a ~15% IPC uplift for Cortex X5 (vs X4). Which sounds mediocre when we had this article a few months ago...

RESEARCH NOTE: Arm’s “Blackhawk” CPU Is An Audacious Plan To Have The Best Smartphone CPU Core This Year - Moor Insights & Strategy

For years now, there has been what I consider a healthy, competitive tension between Arm CPU instruction set licensees and Arm’s pe-packaged and pre-validated IP licensees. (I am sure some licensees would challenge me on “healthy” given Apple’s performance.) I think it made sense that Arm would...

moorinsightsstrategy.com

Conclusion: Either the 2700 GB6 ST leak for D9400 is fake, or

ARM'S endeavour to create a "Custom ARM core killer" has failed.

Arm CEO Rene Haas’s strategy to “eliminate the performance gap between Arm-designed processors and custom Arm implementations.” He did not say anything about "Custom ARM core killer" . He did say largest year-over-year IPC performance increase in 5 years.”

I already told you to take out calculator to compute, here let me create a table for you to show GB6 1T's performance per clock. By dividing score with clock speed, one can get PPC numbers as shown below: (PPC is not IPC but it is only metric to compare different platform)

• If ST score of D9400 is correct, Yeah PPC is the highest in the whole lineup. It is about 19% better than 8G3 and 5% better than A17 Pro.
• It's 19% performance uplift compared to Cortex-X4. I don't know it is largest IPC increase in 5 years but at least it is 5% better than A17 Pro.
• 5% PPC better than A17 Pro has proven that Cortex-X5 core can perform better than current Apple's A-series based on PPC.
• Yes, I would say Cortex-X5 has eliminated the performance gap between Arm-designed processors and custom Arm implementations based on PPC/IPC.
• As for upcoming A18 Pro and 8G4 with custom cores, we only know they are scoring around 3500 points in GB6's 1T but we don't know final clock speed.

	A14	A15	A16	M3	A17 Pro	A18 Pro		8 Gen 3 FG	X-Elite G1	X-Plus	8 Gen 4		D9400	D9300
GB6 1T	2090	2310	2566	3084	2908			2300	2900	2425			2700	2225
Clock Speed	2.99 GHz	3.23 GHz	3.46 GHz	4.06 GHz	3.78 GHz			3.4 GHz	4.2 GHz	3.4			3.35 GHz	3.25 GHz
Perf Per Clock	699	715	742	760	769			676	690	713			806	685
+%	+ 15%	+13%	+ 9%	+ 6%	+ 5%			+ 19%	+ 16%				Base	+ 18%
GB6 MT	4838	5627	6562	11564	7235			7501	14029					7236
Perf Per Core	806	938	1094	1445	1206			938	1169					904

Conclusion: First of all you don't even want to spend times to calculate IPC/PPC to get better idea. And secondly when I said the number is pretty impressive, you shown ARM article and try to BS about "Custom ARM core killer" when the CEO only said eliminate the performance gap between Arm-designed processors and custom Arm implementations. Clearly based on data in the table above, Cortex-X5 has accomplished the job. Geez...

 

[FlameTail]

Goood, Tigerick. Gooood.

*One caveat though: there are two GB6 ST scores circulating about the 8G4. 2800 and 3500. I am inclined to believe the former, as it's what most leakers (including Revegnus) have leaked. The 3500 is likely made up.

 

[poke01]

i'm sorry but

The 3500 is likely made up.

yep, gen 4 scoring higher than X elite is not possible especially when nearly the same core

 

[Tigerick]

Goood, Tigerick. Gooood.

*One caveat though: there are two GB6 ST scores circulating about the 8G4. 2800 and 3500. I am inclined to believe the former, as it's what most leakers (including Revegnus) have leaked. The 3500 is likely made up.

Yeah, 2800 makes more sense, and if the core is the same as X-Elite, we should be expecting 4GHz with 700 PPC. God, I can't believe Phoenix core is showing its ages before launching...

I have compiled all upcoming flagship SoC for mobile phone. And I think I know why A18 Pro scoring 3500 might be true due to low amount of CPU core counts. That's why Apple able to clock A18 Pro higher than 8G4 even though they are all made by N3E process. Let's hope SF manage to launch Exynos 2500, if not all flagship SoC are going to be made by TSMC...

	Node	CPU Config	Total CPU Cores	CPU	Max CPU Speed	GB6 1T	PPC	GB6 MT
A18 Pro	N3E	2 + 4	6	?	Highest	3500 ?	?
8 Gen 4	N3E	2 + 6	8	Phoenix	Second High	2800 ?	~700 ?
D9400	N3E	1 + 3 + 4	8	Cortex-X5	3.35 GHz	2700	806
Exynos 2500-A	SF3	1 + 3 + 4	8	Cortex-X5	3.2 GHz ?	?	?
Exynos 2500-B Laptop		1 + 5 + 4	10

 

[soresu]

Let's hope SF manage to launch Exynos 2500, if not all flagship SoC are going to be made by TSMC

No reason they shouldn't.

First gen SF3 has been operational for quite some time now - even with the brand new MBCFET (Nanosheet) device type they should be getting decent yields at this point.

 

[FlameTail]

they should be getting decent yields at this point.

Uhmm...