IBM Contributing A2O Processor Core To OpenPOWER Community

[phoronix]

Phoronix: IBM Contributing A2O Processor Core To OpenPOWER Community

IBM announced today at the OpenPOWER Summit 2020 that they are contributing the A2O POWER processor core and Open Cognitive Environment to the OpenPOWER community...

http://www.phoronix.com/scan.php?page=news_item&px=IBM-OpenPOWER-A2O

[schmidtbag]

Pretty cool - assuming this has compatibility with existing ppc64el packages, I imagine this would be the most mature open-source architecture. RISC-V still seems to have a long way to go, because unfortunately most of those who take part in it seem to be doing their own downstream changes. IBM has an incentive to keep this appealing to a wide audience.
I'm sure IBM realizes that their POWER architecture just doesn't have the edge that it used to. AMD can compete with cores. ARM can compete with efficiency and cost. Intel commands more influence in the server space. GPU-based servers need more PCIe lanes than what IBM offers. In a day and age where people are paranoid about their security, I think IBM offering an open-source platform is an easy way to attract customers.

[stormcrow]

OMG.

That is big and significant news. Nice to hear that some things are moving in the right direction.

IBM finally waking up to the fact that they better let others manufacture the CPUs, if they want their status to recover from the Wintel fiasco of '80s.

Recover? Recover what? A PC business they were never committed to to begin with? IBM never much cared about WinTel in the executive suites and even when the PC industry they helped start began to leave them behind they never really "suffered". IBM's bread and butter is services and support contracts. They do that whether you run IBM mainframes with PC terminal software, or if you're using Solaris or Linux.

IBM suffering? Please. They're a Fortune 50 company and they aren't in any shape or form hurting financially. Releasing a narrowly tailored ISA that probably contributes little to their bottom line is more about Public Relations than anything else. Sure anyone can use the ISA now, but who are the hardware engineers going to call for support contracts? IBM.

[jntesteves]

It didn't take too long for IBM to realize they should not be trying to compete in the CPU market that was mostly commoditized by ARM a decade ago. Since they know the CPU business is complemental to their services business, they will want to keep the trend of commoditizing their complement. And this is good for the end users.

[tuxd3v]

"On 7nm production IBM estimates the A20 processor core can operate at 3.0GHz at 0.25 Watts or 4.2GHz at 0.85 Watts per core. But this core was designed originally for 45nm where it is consuming 1.49 Watts at 2.30GHz."

"4.2GHz at 0.85 Watts per core"... this values are better than ARM ones, we know that ARM ones are around 0.95W at 2Ghz( I don't remember now the process node for those values..probably 28, or 12 nm, I checked them when looking for Alibaba RISCV cpu comparison.. but I don't really remember now )

I would love to see a comparison of this processor at 28 nm( to compare it with other ARM designs )..
At 28 nm is should consume maybe a bit more than ARM heavy cores( but its a maybe..it could even be less, those values are very good ).

At 7 nm this processor shines, and eclipse completely ARM, in performance/latency for sure, since it can operate at 4.2Ghz..

What we need now is somebody with lots of money, to produce it on silicon, and if they can replicate, the RTL estimations, well, POWER seems to be here to stay!

[edwaleni]

The one session at OpenPOWER Summit I am interested in is:

"How easy is it to run on ppc64le?"

This presentation is about how easy it is to migrate workloads on OpenPOWER systems, mainly looking at the ppc64le architecture. Firstly we will look at some examples of how difficult it is to migrate existing open source software to ppc64le. In the second part, some optimizations will be looked at, and how these optimizations can lead to better performance and the effort to this is not as difficult as expected.

Seems much of any software ported to ppc64le isn't using any of the optimizations that come in the CPU. I am curious how they can be exploited.

[tuxd3v]

Seems much of any software ported to ppc64le isn't using any of the optimizations that come in the CPU. I am curious how they can be exploited.

That is a problem indeed, porting is a burden, but it would be more difficult to port and take advantage of the full instruction set, usually the ports only take in consideration the basic instruction set( it turn the port more generic to run in more hardware, but the downside is that you don't take advantage of the good instructions on newer cpus )..

I am waiting for the session for this project:
Tuesday, September 15 • 10:35am - 11:05am
Open Hardware GNU/Linux PPC64 Laptop Potential - Roberto Innocenti, Power Progress Community (not profit association)

Does any one knows were to see this presentation?
Thanks in Advance.

[PerformanceExpert]

At 7 nm this processor shines, and eclipse completely ARM, in performance/latency for sure, since it can operate at 4.2Ghz..

Frequency != performance. This is a simple 2-wide design with very limited resources, so likely similar to Cortex-A55 in terms of IPC. A high-end 6-wide core like Cortex-A78 would beat it running at half the frequency...

Remember, performance = IPC * frequency. A high frequency isn't useful when paired with low IPC.

[ahab]

The one session at OpenPOWER Summit I am interested in is:

Seems much of any software ported to ppc64le isn't using any of the optimizations that come in the CPU. I am curious how they can be exploited.

GCC / CLANG after 2016 both take advantage of the Power optimizations. Some code is hard coded for x86 or ARM vector instructions... in that case it might not be so easy. The hardware optimizations that are hardcoded are generally the biggest problem with porting to any platform.

Other issues might be x86 code is set up to use Intel's MKL rather than OpenBLAS, which are generally compatible, but will still require some re-work in a make file.

[tuxd3v]

Frequency != performance. This is a simple 2-wide design with very limited resources, so likely similar to Cortex-A55 in terms of IPC. A high-end 6-wide core like Cortex-A78 would beat it running at half the frequency...

Remember, performance = IPC * frequency. A high frequency isn't useful when paired with low IPC.

I don't know about cortex a78,
But cortex A75 is a 4 way fetch, like the Power A20, and the Power A20 also has a 4 way issuer, Power A20 is also SMT 2, so it should be a lot more powerfull than the A55.. and the A55 can't do 4Ghz, not even 3Ghz..