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

[ShintaiDK]

The next ARM core unveiled. Mostly just PR slides tho. But it does give a little glimse of what to come.

Will be interesting to see how the actual real world performance plans out. Scott covered quite a few points in terms of what to expect.

http://techreport.com/news/27767/arm-unveils-cortex-a72-cpu-mali-t880-graphics-and-more

 

[witeken]

Doesn't excite me, but that's not wonder when PR slides come from a company that makes claims against products from competitors that haven't even been released into the market.

I'll see when it comes to market. Doubt there'll be much "premium mobile" delivered in '16.

 

[NTMBK]

Looking forward to this. Should be good for servers, too.

 

[teejee]

Exciting, the performance difference between smartphone SOC's and desktop PC's seems to be even smaller next year.
84% increase from A57 is impressive, that takes Intel about 5 years to achieve.

 

[Shehriazad]

I wonder if ARM will at some point be capable enough to challenge big CPUs...that would be interesting to see.

 

[witeken]

Exciting, the performance difference between smartphone SOC's and desktop PC's seems to be even smaller next year.

Why? An EOY '16 CPU, which we only have marketing slide from, that will likely have a hard time competing against even '14 Cyclone or its successors, whatever Nvidia releases then and Qualcomm's next architecture? I wouldn't be surprised if it has to face Cannonlake-Y.

 

[Sweepr]

The Cortex A57 + Mali T760 combination is already fast and they are promising ~80% more CPU/GPU performance within smartphone power budget as soon as next year? Really impressive if they deliver this.

 

[monstercameron]

is that interms of ipc or just process node for the performance uplift of the a72 core?
if it is ipc then that is getting pretty close to intel ipc-wise.

 

[Arachnotronic]

Great stuff! Can't wait to see it in the real world.

 

[teejee]

Why? An EOY '16 CPU, which we only have marketing slide from, that will likely have a hard time competing against even '14 Cyclone or its successors, whatever Nvidia releases then and Qualcomm's next architecture? I wouldn't be surprised if it has to face Cannonlake-Y.

Cortex A57 is doing well, good performance and power efficient, so 84% better performance is a big improvement.
Nvidia does not seem to do SOC's aimed at smartphones at all anymore, only used in tablets.
ARM's core IP's are very important since they will be used in many different SOC's. E g cheap A72's from Mediatek will be common in 2017.

 

[Arachnotronic]

Cortex A57 is doing well, good performance and power efficient, so 84% better performance is a big improvement.
Nvidia does not seem to do SOC's aimed at smartphones at all anymore, only used in tablets.
ARM's core IP's are very important since they will be used in many different SOC's. E g cheap A72's from Mediatek will be common in 2017.

Not sure how cheap those SoCs from MediaTek will be since they'll be quite large and high performing, but yes, this should be a great core for the ARM licensees that don't have their own in-house high perf cores. :thumbsup:

 

[Sweepr]

is that interms of ipc or just process for the performance uplift of the a72 core?
if it is ipc then that is getting pretty close to intel ipc-wise.

Unfortunately we don't know the specifics. Comparing the latest 28nm A15 SoCs to 20nm A57 SoCs (Exynos 5433 and Snapdragon 810) I'd say they were a bit too optimistic with the 90% better CPU performance claim (not sure what they mean by sustained performance though). That's not to say Cortex A72 won't be an impressive core but let's be cautious.

 

[ShintaiDK]

is that interms of ipc or just process for the performance uplift of the a72 core?
if it is ipc then that is getting pretty close to intel ipc-wise.

Its not IPC, its everything, plus a large dosis of "optimism".

Dont expect real world numbers to be anything near, as Scott also explains. Its all hidden in "substained performance", "same workloads" etc.

 

[Fjodor2001]

84% CPU performance increase! That's actually massive. How long can they continue improving at this rate, while we're seeing ~8% performance increase on Intel desktop CPUs?

 

[witeken]

Not sure how cheap those SoCs from MediaTek will be since they'll be quite large and high performing, but yes, this should be a great core for the ARM licensees that don't have their own in-house high perf cores. :thumbsup:

A cheap SoC on an expensive process node = oxymoron.

Not sure why everyone is blindly accepting ARM's claims. Technology always moves on, certainly with new process nodes, so it isn't surprising that tech that's 2 years ahead of us seems exciting, but I wouldn't use this slide to separate a good CPU from a leading one, certainly not when we don't know what the competition will have.

Especially Nvidia is very good as this:

 

[witeken]

84% CPU performance increase! That's actually massive. How long can they continue improving at this rate, while we're seeing ~8% performance increase on Intel desktop CPUs?

Note how it says "sustained performance". If you scale Krait (or A57) down to that node, you'll also get some nice 2X gains or so. Also the testing methodology isn't mentioned.

 

[Fjodor2001]

Note how it says "sustained performance". If you scale Krait (or A57) down to that node, you'll also get some nice 2X gains or so. Also the testing methodology isn't mentioned.

The level of performance increase being displayed by ARM is still so substantial that even if you remove any such marketing fluff it will be truly impressive. So I think the question in my previous post is still valid. How long can they keep up this improvement rate, when Intel delivers 8% desktop CPU performance increase? Won't ARM hit a wall soon?

 

[ShintaiDK]

Yep, seems people already running with the ball for the hype.

Using the same metrics, we could also say that something like the 20W C2750 Atom is 3-4 times faster than a Haswell i5. Substained performance within a specific TDP. Nothing else.

Anyway, most of the benefits comes from higher clocks and the CCI 500. Dont expect anything revolutionary IPC wise.

 

[antihelten]

To me, the term "sustained performance", seems to indicate that at least some of the 84% increase comes simply from the fact that A72 will throttle less than A57, and not necessarily an increase in peak performance (which is the far more relevant metric with smartphones/tablets)

 

[Sweepr]

Anyway, most of the benefits comes from higher clocks and the CCI 500. Dont expect anything revolutionary IPC wise.

ARM claims 2.5GHz sustained frequency for mobile devices. That could easily account for a large portion of that 84% performance increase (depending on the sustained frequency of Cortex A57 @ 20nm). A57 @ 14nm should provide better comparisons.

 

[Exophase]

If you scale Krait (or A57) down to that node, you'll also get some nice 2X gains or so.

What, you think that'll allow you to more than double the average clock speed? Far from it.

 

[Haserath]

Arm should be topping out at the same frequencies Intel is at on the desktop, unless they can somehow achieve higher?

I'm curious if the performance increase will be from TSV memory and that power reduction along with finfet? It should at least have LpDDR4.

Intel still is able to deliver big speedups from laptops, though not quite as large because they aren't nearly as constrained. Intel has been bringing performance down to the power envelope of ARM while ARM scales up. A 50% power reduction would be similar to a 100% performance gain, though it most likely doesn't work out quite like that.

 

[Exophase]

Arm should be topping out at the same frequencies Intel is at on the desktop, unless they can somehow achieve higher?

They won't come anywhere close unless they want to reach similar levels of power consumption. You don't design a uarch that can scale to those frequency levels unless that market segment (desktops, high end laptops, servers) is a critical target. Because to do so means sacrificing efficiency at the lower end. This doesn't make sense for an ARM CPU whose most popular markets will remain phones and tablets by a wide margin.

For comparison see Atom which also isn't projected to exceed 3GHz any time soon, let alone 4GHz.

I'm curious if the performance increase will be from TSV memory and that power reduction along with finfet? It should at least have LpDDR4.

The memory controller/interface used doesn't have anything to do with the CPU core, and if ARM's estimations are remotely honest they'll be of the core(s) only (and maybe the interconnect)

 

[imported_ats]

Exciting, the performance difference between smartphone SOC's and desktop PC's seems to be even smaller next year.
84% increase from A57 is impressive, that takes Intel about 5 years to achieve.

And its probably like totally really completely a truthful 84%. No way would they ever be throwing around marketing BS.

 

[waffleironhead]

Doesn't excite me, but that's not wonder when PR slides come from a company that makes claims against products from competitors that haven't even been released into the market.

I'll see when it comes to market. Doubt there'll be much "premium mobile" delivered in '16.

Projected numbers from Arm against projected numbers from Intel. The only limit is the imagination of the marketers. :awe: