News New ARM roadmap up to 2022 IP cores

[soresu]

Seems ARM just dropped a new roadmap I would assume at the ARM TechCon conference that started on tuesday 6th.

The recent Neoverse roadmap update (from V1 announcement) gave us Perseus = Matterhorn/A79 at 40% IPC improvement from Ares/Enyo/A76.

From this I can infer that since A77 is 1.2x A76 and A78 is 1.07x A77 that Matterhorn's 40% over A76/Ares is about 1.09x A78.

Given the 30%/1.3x over A78 that Makalu represents on this new client roadmap I can infer that Makalu is about 1.19x Matterhorn.

If these are just the 'big' PPA balanced Axx cores then these are still a decent improvement - and obviously at least one of these gens will have an Xn core too paired with it.

Big and not unexpected news too is that Makalu will be 64 bit only, which sounds a lot like Makalu is v9-A finally ditching v7-A/ARM32 hardware level compatibility as I thought some time ago - though I don't doubt that ARM32 software emulation will still be possible.

Link here to the ARM blog entry.

 

[teejee]

up to 2022 cores? That is a lot of cores

 

[ElFenix]

up to 2022 cores? That is a lot of cores

ah, the ol' anandtech switcharoo!

 

[Lodix]

I know the blog mentions that the new cores are compared to A78. But looking at the graph it is clearly using X1 as the reference point and not A78. The graph is "Peak CPU Performance" per ARM that normally means peak single CPU Performance at max clock without throttling.

The jump from A77 is too large to be referencing A78. It only brings ~7% peak performance improvements at the same 3GHz. The graph mentions X1 next to A78 and I am sure it is using the X1 at 3GHz as the mark.

Also ARM previously mentioned Matterhorn is a "huge" core, probably the evolution of the X1.

 

[soresu]

I know the blog mentions that the new cores are compared to A78. But looking at the graph it is clearly using X1 as the reference point and not A78. The graph is "Peak CPU Performance" per ARM that normally means peak single CPU Performance at max clock without throttling.

The jump from A77 is too large to be referencing A78. It only brings ~7% peak performance improvements at the same 3GHz. The graph mentions X1 next to A78 and I am sure it is using the X1 at 3GHz as the mark.

Also ARM previously mentioned Matterhorn is a "huge" core, probably the evolution of the X1.

If so that is a markedly less impressive perf jump for 2 generations compared to the 1.56x from A76 to X1, or A73 to A76 for that matter.

X1 alone brought a 1.3x IPC increase over A77, yet now they are announcing ahead of time that all you get for 2 generations is 30%?

Sounds like more of the same 15% YOY IPC improvement promise that came months after A76 with the Deimos/Hercules roadmap.

 

[soresu]

Odd that the Neoverse track seems to be continuing with the Greek gods/heroes naming scheme while the Cortex Axx track has changed to mountains starting with Matterhorn.

Something to keep in mind though, the roadmap refers to Matterhorn and Makalu as generations rather than cores, the cores themselves may have different codenames, or perhaps the mountains only represent the Xn cores.

Going by this logic the name Andromeda for A79 would be perfect to match up with Perseus/N2 on the Neoverse roadmap.

 

[IntelUser2000]

I know the blog mentions that the new cores are compared to A78. But looking at the graph it is clearly using X1 as the reference point and not A78. The graph is "Peak CPU Performance" per ARM that normally means peak single CPU Performance at max clock without throttling.

Well, I'm not sure if its that clear. The starting point isn't at zero, a method used to amplify the differences.

 

[Lodix]

Well, I'm not sure if its that clear. The starting point isn't at zero, a method used to amplify the differences.

Yep. But you can clearly see that the jump over A77 to next one is similar to the improvement from A75 to A76 and A76 to A77, which both brought over 20% increase in IPC and overall single core performance. And the title of the graph is "Peak Performance" and the dot is placed over "Cortex X1/A78", of those the one with higher peak performance is obviously X1.

Edit: If you do the math to reach that claim of 2'5x performance you need that 1'3x Integer IPC increase of the X1. So my point is proven.

 

[name99]

I know the blog mentions that the new cores are compared to A78. But looking at the graph it is clearly using X1 as the reference point and not A78. The graph is "Peak CPU Performance" per ARM that normally means peak single CPU Performance at max clock without throttling.

The jump from A77 is too large to be referencing A78. It only brings ~7% peak performance improvements at the same 3GHz. The graph mentions X1 next to A78 and I am sure it is using the X1 at 3GHz as the mark.

Also ARM previously mentioned Matterhorn is a "huge" core, probably the evolution of the X1.

Remember "huge" is in the context of ARM.
Even Apple's cores are ~smaller than Willow Cove (this depends on EXACTLY what you want to call "core", if you want to include L3 and ring stops, etc, but they're each basically about the same size. And of course TGL is notably less dense than ICL.
I expect X1 (and these successors) to be ~smaller than Apple.

 

[name99]

If so that is a markedly less impressive perf jump for 2 generations compared to the 1.56x from A76 to X1, or A73 to A76 for that matter.

X1 alone brought a 1.3x IPC increase over A77, yet now they are announcing ahead of time that all you get for 2 generations is 30%?

Sounds like more of the same 15% YOY IPC improvement promise that came months after A76 with the Deimos/Hercules roadmap.

When the best your competitor can offer is 18% ever 4..5 years...
AMD gets (apparently) 19% after 18 months [Zen3 over Zen2], but will they sustain that in 2022?

 

[soresu]

When the best your competitor can offer is 18% ever 4..5 years...

Fiddle dee dee, I'm not interested in AMD/Intel vs ARM - I was talking about ARM and ARM alone.

I'm perfectly happy to have a Threadripper monster in one corner and a silent but powerful SBC beast in the other.

AMD gets (apparently) 19% after 18 months [Zen3 over Zen2], but will they sustain that in 2022?

Depends on your use case/workload really.

They addressed a big problem point for gamers with Zen3, but did not make any significant changes to SIMD execution.

I would expect a minimum 50% SIMD improvement with Zen4, possibly even 2x and with significant AVX512 instruction set support (though to what level given its fragmentary nature).

For me even 50% greater SIMD would be a win, but 2x as with Zen2 would be stellar - for content creators there are many workloads that will gobble up all the SIMD execution power they can get their hands on.

Couple this with an increase to 96-128 cores per socket and N5P and you have an awesome increase in content creation compute power with a commensurate efficiency per watt increase, just as with Zen2.

Remember "huge" is in the context of ARM.

Actually I'm pretty sure that the "huge" comment was in the specific context of ARM prior to the unexpected X1 5 wide "let loose the dogs of war" reveal.

That's not to say it's impossible that X2 will jump straight to 6 wide, but it sounds like a bit of a stretch, especially if most implementations of it are still going to be on the same 5nm node as X1, or pretty close.

They got 3 big Axx gens so far out of 4 wide too, I wouldn't expect them to jump to anything higher than 5 wide so soon, likewise I wouldn't be at all surprised to find that A79 is still 4 wide.

With them making such a big change to the codename scheme with Matterhorn I do wonder if they won't make a dramatic official naming scheme change to all the Cortex Axx cores going forward, though clearly Cortex X branding is still shiny new and needs no refinement or change.

 

[DrMrLordX]

Should we expect A78 and X-1 to be in consumer silicon by next year sometime?

 

[Lodix]

Should we expect A78 and X-1 to be in consumer silicon by next year sometime?

Before the end of the year there will be a couple of new SoC's announcements with the new IPs inside them. Sometimes they are in the consumer hands in the same year as they are announced ( the ARM IPs ) but if not, then Q1 next year.

 

[soresu]

Sometimes they are in the consumer hands in the same year as they are announced

Huawei will not be doing that again for the foreseeable future, and they are the only company short of Mediatek (MT8173 in Fire TV 2nd gen) that ever did such a thing to my knowledge.