[TechReport]ARM unveils Cortex-A72 CPU, Mali-T880 graphics, and more

[NTMBK]

I wonder how 20nm vs 32nm would go. 32nm probably doesn't have a hard time competing against 20nm.

Which is why Intel's 32nm Atom phones sold absolute gangbusters, sure.

 

[witeken]

Which is why Intel's 32nm Atom phones sold absolute gangbusters, sure.

If you have a crappy platform, crappy BOM, crappy pricing strategy, crappy and slow roadmap and crappy architecture, even a 0.1fm process node still wouldn't matter. (So never mind a node that's simply on par in terms of power.)

 

[Arachnotronic]

If you have a crappy platform, crappy BOM, crappy pricing strategy, crappy and slow roadmap and crappy architecture, even a 0.1fm process node still wouldn't matter. (So never mind a node that's simply on par in terms of power.)

I seriously doubt any foundry 20nm is "on par" with Intel 32nm in terms of electrical characteristics. I'm sure they're ahead.

 

[imported_ats]

Do you not remember how bad initial revisions of the A15 were? I wouldn't be surprised at all if their 1.9x claim is true.

Even if the initial revisions were bad, they aren't going to do 1.9x with a single node and not even a real single node. Even assuming the best numbers available from TSMC for 20nm to 16FF+, they need to clear another 30-40%. I just don't see it. And I certainly don't see it in any practically relevant case. ie, its a lot easier to increase thermally constrained performance in multi-threaded cases, but the reality of the software is that multi-tread performance isn't really at all relevant to the actual end user.

 

[oobydoobydoo]

Just a bit.

I see your point, but I feel like performance claims and roadmaps are on entirely different levels. It's a bigger deal to say "we're gonna beat our competitors by x amount" than it is to say "we'll have a new product in this timeframe." Really though, I don't think roadmaps -- unaccompanied by other statements -- are hype at all.

Intel did do the same thing, back when they disclosed Silvermont's architecture. It was probably, at least in part, a response to ARM's claim that Intel was about to get beaten with Silvermont, which was not even close to being the case. Given this, I'm not really surprised that people are wary of ARM's claims, but I personally find them reasonable.

I do recall the disappointment about A15, however I would argue that when we talk about these roadmaps doubt should be implied, and constantly calling every claim hypeperbole or hype (without any evidence) detracts from the discussion.
When Intel publishes a roadmap I don't have 100% trust nor 0%, it all must be taken with a grain of salt. I would say the same thing about calling intel's claims hyperbole without evidence. There is no need to say that this might be exaggerated, we all know that.

None of this is criticism directed at you specifically at all, just FYI.

 

[Nothingness]

Intel did do the same thing, back when they disclosed Silvermont's architecture. It was probably, at least in part, a response to ARM's claim that Intel was about to get beaten with Silvermont, which was not even close to being the case. Given this, I'm not really surprised that people are wary of ARM's claims, but I personally find them reasonable.

It's that Intel started to make comparisons against ARM chips back in 2008 when they launched Atom. I laughed a lot when I looked at their slides back then :biggrin: If anyone can find them...

EDIT: I find ARM way of presenting their data misleading to say the least... I think the data is correct, it's just very oddly presented.

 

[III-V]

Even if the initial revisions were bad, they aren't going to do 1.9x with a single node and not even a real single node. Even assuming the best numbers available from TSMC for 20nm to 16FF+, they need to clear another 30-40%. I just don't see it. And I certainly don't see it in any practically relevant case. ie, its a lot easier to increase thermally constrained performance in multi-threaded cases, but the reality of the software is that multi-tread performance isn't really at all relevant to the actual end user.

I'm sure they were talking about overall, multithreaded performance, and yes, its relevance is dubious, though not entirely useless. FinFETs are just a really big deal, though, and I imagine they're quite critical to achieving the "sustained performance" they mention.

Here's an earlier slide comparing their A15 to the A57:

They claim about 50% node-agnostic improvement for the A57. This is with an important caveat -- they're probably factoring in that the A57 is 64-bit capable, and has the benefits and penalties that come along with it. Wiith the 20nm process, it bumps to ~1.9x. That's a ~26% contribution from the process. I imagine 16/14 will be a bit larger of a gain.

The A57, according to Anand, was not a very large change from the A15. It's still 3-wide/3-issue, but it adds 64-bit support and I'd imagine small-but-effective "fixes" to the issues that A15 had.

Given this, there is certainly more room left on the table. Decode and execution width could be increased -- potentially to better efficiency, potentially not. I imagine there is some pressure, particularly thanks to Apple-jealousy, to move to a wider design. ARM's not really good at giving up their finer details of their architecture, so it's difficult to see where they have room to grow, but I imagine that they could achieve it if they're less concerned about giving into pressure and bloating core area in exchange for better competitiveness with Apple.

As far as Intel goes, they recently achieved a ~2x increase in performance moving from Saltwell to Silvermont. They're moving to 14nm, which seems a bit disappointing for high performance, but for low power, it's doing the trick. Max frequency increases by ~12.5%, which is rather tame, but it should be able to have much better sustain.

I've been told from "those in the know" that Airmont is more than a simple manufacturing process improvement, and is like a "tick-plus" in this regard. There are quite a number of moves for Intel to make here, including the addition of a uOP cache, increasing decode and execution width (pretty much a given), adding an L3 cache, adding FIVR, bumping up L1 instruction cache to 32KB, more aggressive branch prediction, and increasing the size of various buffers. If Intel were to give it the full "tock treatment," I'd be surprised if gains were less than 30% per clock, ~46% overall, but we probably won't find out what they're doing until April.

Tying this all back to ARM, they're likely going to be catching up quite a lot, thanks to the substantial delays on Intel's part. If Intel can get Broxton out by the time ARM has its A72 out, they should be able to stand their ground and perhaps gain some small market share, but they keep missing great opportunities.

It's that Intel started to make comparisons against ARM chips back in 2008 when they launched Atom. I laughed a lot when I looked at their slides back then If anyone can find them...

Did they? I've never seen those.

EDIT: Like this one?

Gah, that reminds me -- I'm glad OMAP's gone. TI makes some great stuff... but OMAP was not one of those things.

 

[witeken]

I seriously doubt any foundry 20nm is "on par" with Intel 32nm in terms of electrical characteristics. I'm sure they're ahead.

Both are second generation HKMGs. And because of 20nm's density, it might even have a tailwind from leakage.

 

[Nothingness]

EDIT: Like this one?

That's it thanks! ARM11 was already old back then. Cortex-A8 was just around the corner, and Cortex-A9 had been announced. Everyone with a little bit of micro-arch knowledge knew Intel had missed their target.

Gah, that reminds me -- I'm glad OMAP's gone. TI makes some great stuff... but OMAP was not one of those things.

They were indeed pretty bad. Very late to market, poor memory controller, etc. OMAP4 slightly improved that, and OMAP5 could have been nice. Anyway I still want a BeagleBoard X15 and TI DSP are lovely monsters :biggrin:

 

[III-V]

That's it thanks! ARM11 was already old back then. Cortex-A8 was just around the corner, and Cortex-A9 had been announced. Everyone with a little bit of micro-arch knowledge knew Intel had missed their target.

They were indeed pretty bad. Very late to market, poor memory controller, etc. OMAP4 slightly improved that, and OMAP5 could have been nice. Anyway I still want a BeagleBoard X15 and TI DSP are lovely monsters :biggrin:

I had an OMAP3 on my Droid 2... not a happy time. I think the biggest issue was the 512MB RAM, though. I'm now using a Snapdragon 800... night and day. I feel like 14/16nm might be a very noticeable bump as well.

 

[imported_ats]

I'm sure they were talking about overall, multithreaded performance, and yes, its relevance is dubious, though not entirely useless. FinFETs are just a really big deal, though, and I imagine they're quite critical to achieving the "sustained performance" they mention.

Yeah, but lets just say I'm rather dubios of marketing slides with performance without any detail on what is actually being compared.

Here's an earlier slide comparing their A15 to the A57:

They claim about 50% node-agnostic improvement for the A57. This is with an important caveat -- they're probably factoring in that the A57 is 64-bit capable, and has the benefits and penalties that come along with it. Wiith the 20nm process, it bumps to ~1.9x. That's a ~26% contribution from the process. I imagine 16/14 will be a bit larger of a gain.

Yes, but did we actually even get a 50% improvement with A57? AFAIK, we didn't looking at the available data. Its almost like back in the day when all ARM would publish was DMIPS numbers. Don't get me wrong, ARM makes nice designs, its just that their marketing of them is smarmy at best. Pretty much everyone else in the industry when quoting performance numbers will give you what's being tested and under what conditions.

 

[Nothingness]

Don't get me wrong, ARM makes nice designs, its just that their marketing of them is smarmy at best. Pretty much everyone else in the industry when quoting performance numbers will give you what's being tested and under what conditions.

I certainly agree with the first part, but disagree with the second part. Try to reproduce any of Intel comparative benchmarketing result, but don't try too much, you'd waste your time All marketing stuff stinks.

 

[ShintaiDK]

I think we can agree that all marketing slides are misleading/deceptive. But I dont think I have seen anything as misleading as ARM. It is in a league of its own.

 

[III-V]

I think we can agree that all marketing slides are misleading/deceptive. But I dont think I have seen anything as misleading as ARM. It is in a league of its own.

Nvidia's pretty bad. Intel and AMD have certainly had their moments. Kind of hard to pick just one.

 

[SOFTengCOMPelec]

I think we can agree that all marketing slides are misleading/deceptive. But I dont think I have seen anything as misleading as ARM. It is in a league of its own.

You did not read THIS post then ? (joke)

But I agree, I'm finding the slides VERY annoying. I want to have valid points of reference, (such as benchmarks comparing it to currently released cpus), so I can put the figures into perspective.

 

[Exophase]

I've lost hope in ARM's marketing. I'm sad to agree that they really have appeared to have sunk beneath even nVidia, let alone Intel. The only saving grace is that they're so blatantly ridiculous that I doubt anyone really takes them seriously, unlike nVidia's which have more subtle and convincing problems.

This is part of a bigger trend where ARM has been less and less willing to reveal real data about anything instead of BS. With each new CPU core they released since Cortex-A8 the TRM contained less and less information on performance, for example instruction timing and hazards. A9 still had timing information but nothing on reordering, while A5, A7, and A15 (or anything since) have pretty much nothing at all. Uarch details were relegated to less formal presentation slides that were vague and confusing. Cortex-A57 and A53 got almost no mention of uarch details at all (although the A53 articles here helped a little). And the slides talking about performance comparisons said less and less about what they were testing. At this point I'm not even expecting them to reveal anything about A72 outside of these single magic numbers.

Meanwhile Intel and AMD give optimization guides with real information, although the Bulldozer family one is pretty bad.

This is especially annoying for the in-order cores that ARM still heavily depends on, where you really need to know this to write good code. Oh, we can pretend that compilers are always as good or better at assembly than the best programmers no matter how long that continues to not be the case, or maybe that all compiler writers somehow have better access to processor information than the general public in the first place.

I look at something like Raspberry Pi especially, now there's this Cortex-A7 version which is a great improvement but still low end. Yet a huge platform that could use all the fine tuned optimization it can get. Would be nice if someone was inclined to fine tune code the processor, but they can't really unless they or someone else wants to reverse engineer all of it themselves.

I'd really like to know what ARM is so afraid of. Do they really think knowing this information will give their competitors any real advantage?

That's it thanks! ARM11 was already old back then. Cortex-A8 was just around the corner, and Cortex-A9 had been announced. Everyone with a little bit of micro-arch knowledge knew Intel had missed their target.

Back around that time Intel was also very big on comparing Javascript benchmark scores, when the browsers had much inferior ARM JITs or were lacking them entirely.

And when Z series started making it into anything even resembling a mobile device they weren't clocked at anywhere near 1.6GHz. I remember that one hybrid Symbian/Windows Moorestown phone that halved the rated clock speed of the part and still drained the battery in under an hour in Windows. But Intel was talking about Atom hitting it big in MIDs (think like Nokia N800) long before Moorestown even.

They were indeed pretty bad. Very late to market, poor memory controller, etc. OMAP4 slightly improved that, and OMAP5 could have been nice. Anyway I still want a BeagleBoard X15 and TI DSP are lovely monsters :biggrin:

For better or worse I'm still getting an OMAP5-based Pyra when it comes out, and only a few years later than OMAP5 devices were supposed to It's really sad that for a device like this TI is still one of the only realistic options.

 

[Exophase]

But every once in a while an ARM engineer cuts through the marketing and comes to the rescue.. no sooner than I posted that found this on RWT:

http://www.realworldtech.com/forum/?threadid=147766&curpostid=147801

So there, an actual IPC comparison with Cortex-A57 (albeit a big range), that can put to rest the speculation

And 10-50% better IPC while using less power, under the same process and using the same macros is very impressive.

 

[Fjodor2001]

I wonder how long until Intel decides to exit this market. They've been trying for so many years now, and the mobile division lost a staggering $4.21 BILLION in 2014 alone. Kind of sad though, since Qualcomm really needs some competition.

 

[Nothingness]

But every once in a while an ARM engineer cuts through the marketing and comes to the rescue.. no sooner than I posted that found this on RWT:

http://www.realworldtech.com/forum/?threadid=147766&curpostid=147801

So there, an actual IPC comparison with Cortex-A57 (albeit a big range), that can put to rest the speculation

And 10-50% better IPC while using less power, under the same process and using the same macros is very impressive.

I hope that post will clear a few things up. That marketing slide sounded so deceiptive.

I'd really like to know what ARM is so afraid of. Do they really think knowing this information will give their competitors any real advantage?

Or perhaps it's trying not to offer info some patent troll could try to use to attack ARM.

 

[xthetenth]

I wonder how long until Intel decides to exit this market. They've been trying for so many years now, and the mobile division lost a staggering $4.21 BILLION in 2014 alone. Kind of sad though, since Qualcomm really needs some competition.

I think they're going to wait until the platform costs they've steadily been dropping to the price point they've been getting so many design wins at, and then not leave the market because their investment has been rewarding them with more market share for less money spent.

 

[Exophase]

Or perhaps it's trying not to offer info some patent troll could try to use to attack ARM.

Ugh, I hate how good of a point that is. ARM already fought off patent trolls who thought they had a right to every processor that fetched more than one instruction at a time because they had a patent that described a stack machine with that "feature." Those trolls had already extorted license fees from several other CPU manufacturers at that point, and it was actually one of ARM's customers who got sued (good thing ARM defended them, although I suspect that was part of their contract). ARM won that case in court, but it still cost them over a 1 million GBP (peanuts compared to what they would have had to pay up in licensing or damages, but still not something to just shrug off) and I'm sure a lot of anxiety and frustration.

These scam artists and the technically inept patent office that lets people patent anything no matter how obvious, over-encompassing or already prevalent are ruining everything.

I guess that means maybe ARM does have real manuals for some people willing to sign an NDA that also forbids them from suing.

 

[Fjodor2001]

I think they're going to wait until the platform costs they've steadily been dropping to the price point they've been getting so many design wins at, and then not leave the market because their investment has been rewarding them with more market share for less money spent.

Not sure what you mean by that. Are you implying Intel will drastically drop the price on its mobile SKUs to gain market share? Because so far we're not saying many Intel design wins in this segment.

And even Intel does that, I'm not sure they'll succeed. ARM rules this segment, and has better products. In addition both Qualcomm, Apple and Samsung have their own ARM core derivatives, so they are firmly entrenched in the "ARM world" in mobile.

 

[Arachnotronic]

I've lost hope in ARM's marketing. I'm sad to agree that they really have appeared to have sunk beneath even nVidia, let alone Intel. The only saving grace is that they're so blatantly ridiculous that I doubt anyone really takes them seriously, unlike nVidia's which have more subtle and convincing problems.

This is part of a bigger trend where ARM has been less and less willing to reveal real data about anything instead of BS. With each new CPU core they released since Cortex-A8 the TRM contained less and less information on performance, for example instruction timing and hazards. A9 still had timing information but nothing on reordering, while A5, A7, and A15 (or anything since) have pretty much nothing at all. Uarch details were relegated to less formal presentation slides that were vague and confusing. Cortex-A57 and A53 got almost no mention of uarch details at all (although the A53 articles here helped a little). And the slides talking about performance comparisons said less and less about what they were testing. At this point I'm not even expecting them to reveal anything about A72 outside of these single magic numbers.

Meanwhile Intel and AMD give optimization guides with real information, although the Bulldozer family one is pretty bad.

This is especially annoying for the in-order cores that ARM still heavily depends on, where you really need to know this to write good code. Oh, we can pretend that compilers are always as good or better at assembly than the best programmers no matter how long that continues to not be the case, or maybe that all compiler writers somehow have better access to processor information than the general public in the first place.

I look at something like Raspberry Pi especially, now there's this Cortex-A7 version which is a great improvement but still low end. Yet a huge platform that could use all the fine tuned optimization it can get. Would be nice if someone was inclined to fine tune code the processor, but they can't really unless they or someone else wants to reverse engineer all of it themselves.

I'd really like to know what ARM is so afraid of. Do they really think knowing this information will give their competitors any real advantage?



Back around that time Intel was also very big on comparing Javascript benchmark scores, when the browsers had much inferior ARM JITs or were lacking them entirely.

And when Z series started making it into anything even resembling a mobile device they weren't clocked at anywhere near 1.6GHz. I remember that one hybrid Symbian/Windows Moorestown phone that halved the rated clock speed of the part and still drained the battery in under an hour in Windows. But Intel was talking about Atom hitting it big in MIDs (think like Nokia N800) long before Moorestown even.



For better or worse I'm still getting an OMAP5-based Pyra when it comes out, and only a few years later than OMAP5 devices were supposed to It's really sad that for a device like this TI is still one of the only realistic options.

Informative post. Thanks.

 

[Arachnotronic]

I guess that means maybe ARM does have real manuals for some people willing to sign an NDA that also forbids them from suing.

They do.

 

[witeken]

let alone Intel.

Intel always puts its testing methodology on the slide, so not sure why you're complaining.