Discussion Future ARM Cortex + Neoverse µArchs Discussion

[SarahKerrigan]

I had thought so myself previously but when I brought it up someone told me that it was still 2 wide to my surprise

It's 3-wide. ARM's own slides make this clear.

 

[soresu]

It's 3-wide. ARM's own slides make this clear.

View attachment 94108

I think that slide was probably the reason I originally said 3 wide ages ago.

I can't remember who replied to me or what exactly they said about it.

 

[soresu]

Ah, @SarahKerrigan it was this post:

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=posts/40164273/

 

[soresu]

Damn why doesn't linking a post just quote it....

Well, that's correct A75 is 3-wide at decode/front end.
But still identical width in scalar back end as A73: 2xALUs +1xBranch, 2xLSU

A73

https://images.anandtech.com/doci/11441/arm-a75_a55-cpu_diagram-a73.png

A75

https://images.anandtech.com/doci/11441/arm-a75_a55-cpu_diagram-a75.png

 

[SarahKerrigan]

Damn why doesn't linking a post just quote it....

Richie Rich had a weird fixation on ALU count that was not entirely un-annoying.

A75 is a wider, deeper, meaner, scarier core than A73 at pretty much every level.

 

[FlameTail]

Going beyond 8-wide is difficult, but not impossible.

 

[Nothingness]

View attachment 94135
View attachment 94136
Going beyond 8-wide is difficult, but not impossible.

That was already mentioned http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...rchs-discussion.2611782/page-16#post-41159378

 

[FlameTail]

Revisiting some old reddit threads about ARM...

https://www.reddit.com/r/hardware/comments/axyzju

https://www.reddit.com/r/hardware/comments/ayc2vp

Funny reading all the snarky comments directed at ARM

"Oh it can't sustain that performance like x86 CPUs!"

"ARM is for smartphones and low-power devices. It will never be suitable for PCs"

"It can't reach 5 GHz like x86 CPUs!"

"ARM only matches x86 in synthetics. In real life usage, it's not even close"

All this was before the Apple M1.

M1 came along and utterly shocked these naysayers, disproving many of their unfounded arguments.

And it will happen again with Snapdragon X, Nvidia ARM SoC, etc...

The Revolution is coming.

 

[Shivansps]

My bigest doubt about ARM is how power efficient it actually is when you go for high performance only. Small cores dont really count here, as even the A510 are so ridicusly slow that make Intel small cores look like an Zen 4.

 

[FlameTail]

As you can see from the precise power measurement here, Cortex X4 in 8G3 is drawing less than 7W at peak performance.

8G3 has same ST performance as Meteor Lake does. (Geekbench 6)

How much power does Redwood Cove in Meteor Lake consume in ST?

I bet it's atleast 2x that of Cortex X4.

 

[FlameTail]

That is the stock ARM cores.

We know the custom ARM cores from Apple and Qualcomm are far more formidable.

 

[SarahKerrigan]

Revisiting some old reddit threads about ARM...

https://www.reddit.com/r/hardware/comments/axyzju

https://www.reddit.com/r/hardware/comments/ayc2vp

Funny reading all the snarky comments directed at ARM

"Oh it can't sustain that performance like x86 CPUs!"

"ARM is for smartphones and low-power devices. It will never be suitable for PCs"

"It can't reach 5 GHz like x86 CPUs!"

"ARM only matches x86 in synthetics. In real life usage, it's not even close"

All this was before the Apple M1.

M1 came along and utterly shocked these naysayers, disproving many of their unfounded arguments.

And it will happen again with Snapdragon X, Nvidia ARM SoC, etc...

The Revolution is coming.

Nobody serious was saying that stuff. M1's performance characteristics only surprised people that weren't paying attention.

 

[FlameTail]

Leak with bad news just dropped:

https://x.com/Tech_Reve/status/1759950978898657679?s=20

Oh no💀

 

[soresu]

Leak with bad news just dropped:
View attachment 94196

https://x.com/Tech_Reve/status/1759950978898657679?s=20

Oh no💀

From China is a salient point considering Chinese SoC designers do not have access to cutting edge fab nodes anymore and we know that even Samsung is significantly behind TSMC for the moment.

I can't believe that ARM's CEO would have been warbling so positively about it less than 6 months before the IP release if it was suffering such problems.

That would be far worse PR wise than simply casting X5 as a merely evolutionary core design.

 

[poke01]

Nvidia ARM SoC

This will be most interesting thing when it happens on PC cause you know the graphics will actually be great.

DLSS and NV good GPU drivers for Windows. They also have expertise with ARM consumer products ie Nintendo Switch/Switch 2. If Nvidia ever comes to PC with ARM SoC that will be Qualcomms main issue.

 

[soresu]

This will be most interesting thing when it happens on PC cause you know the graphics will actually be great.

Standards compliant more than great I think (though Zink and DXVK are more and more making standards compliance a non issue as long as you can supply a Vulkan driver).

It will depend on how much the quality of their Windows driver stack is tied to x86/64 specific optimisations.

I would be surprised if their Wintel stack just transfers painlessly and even more so for their ARMdroid stack for that matter given its Linux/POSIX base.

I certainly would not be surprised to see some teething problems in their initial WoA run.

 

[FlameTail]

If Nvidia ever comes to PC with ARM SoC that will be Qualcomms main issue.

For sure. Nvidia will be Qualcomm's greatest rival in WoA PCs. But even as a Qualcomm fan, I welcome Nvidia's entry into the market. It will incentivize Qualcomm to make more competitive products, and also:

1. Nvidia making ARM SoCs will give the WoA project a lot more legitimacy, in the eyes of the tech community.

2. It will incentivize game developers to make ARM ports of their games.

Qualcomm stands to benefit from both these things.

 

[igor_kavinski]

If developers target Vulkan API for their games, they should be able to port their games painlessly.

Speaking as a non-developer though.

 

[FlameTail]

Cortex X5 flaunting Dimensity 9400 hits 2.7K in GB6 ST.

 

[FlameTail]

https://x.com/Arm/status/1760336178971107516?s=20

ARM announces Neoverse V3 and Neoverse N3.

Arm Announces Neoverse V3 and N3 CPU Cores: Building Bigger and Moving Faster with CSS

www.anandtech.com

No deep dive architectural details or general performance uplifts are mentioned.

 

[SarahKerrigan]

https://x.com/Arm/status/1760336178971107516?s=20

ARM announces Neoverse V3 and Neoverse N3.

Arm Announces Neoverse V3 and N3 CPU Cores: Building Bigger and Moving Faster with CSS

www.anandtech.com

No deep dive architectural details or general performance uplifts are mentioned.

Performance increase is indeed mentioned. They provide graphs for both N3 and V3.

I'm assuming E3, which was mentioned only in passing, is A520.

 

[FlameTail]

So is V3 based on X5 or X4?

 

[Hitman928]

https://x.com/Arm/status/1760336178971107516?s=20

ARM announces Neoverse V3 and Neoverse N3.

Arm Announces Neoverse V3 and N3 CPU Cores: Building Bigger and Moving Faster with CSS

www.anandtech.com

No deep dive architectural details or general performance uplifts are mentioned.

Expected performance:

 

[soresu]

Expected performance:

If that represents general expectations for X5 and A730 then it looks like they have a dud on their hands outside of AI perf if they are in fact ground up designs.

Here's the codenames for the next cores though:



Might be the first time we had a codename for an E core since E1.

I can only assume that E3 is Hayes/A520 as we aren't expecting anything new this year on the low end.

 

[soresu]

So is V3 based on X5 or X4?

It's certainly a good question.

If it is X5 then they indeed have problems.

The expected perf would seem to gel more with X3 -> X4 improvement on this slide:



Taking 15% as an average between 10-20% of the lions share of the V3 results that is.

If ARM's CEO was talking about X5 having the best IPC improvement since X1 and meant this then it's a dirty rotten lie at best.