Alibaba Reports Their XT910 RISC-V Core To Be Faster Than An Arm Cortex-A73

[microcode]

Wonder when we see the first RISC-V based Smartphones from china, to avoid the trade bans

I there is a lot of investment going into Android (with JVM and V8) on RISC-V right now. This is just the thing to get RISC-V momentum in consumer products.

I am basically not going to buy consumer electronics from PRC, but the Android, JVM, and V8 ports know no borders.

[lowflyer]

Really? Check this wikipedia page:
https://en.wikipedia.org/wiki/Crypto_AG

I know that old story. But you will not be intimidated or even abducted by the Swiss government if you speak out against Crypto AG or the Swiss government.

Or, are you earnestly saying that "because someone did something somewhere, we cannot trust anyone and therefore nothing really matters - So, go on and sell your soul to China"

I know that covert socialists really *like* China and are unable to tolerate anything negative about China - despite the million-high death toll.

[lowflyer]

true. same for american companies. if you care about data security, dont trust american and chinese companies.

Nope. There is much more reason to trust American companies than Chinese. And don't make the mistake to trust European countries more than America.

[kuco]

Nope. There is much more reason to trust American companies than Chinese. And don't make the mistake to trust European countries more than America.

You should not trust ANY company. There are like 100 reasons to not trust an american company for every reason you can find to trust them. About the same for european companies. Yeah, the ratio is a little worse for chinese companies, but who really cares, when your privacy is screwed anyway.

[tuxd3v]

Even if the next generation beats Cortex-A73, remember that is a 4 year old design. Recent Arm cores are more than twice as fast, and improving yearly, so it would be difficult to catch up.

yes the cortex-a73 is a somewhat old, or not so recent, at least is not the top of the line from ARM now..

This actual generation already beat the cortex -A73 in absolute perf, ...what we don't know is perf/watt.
But I assume the next generation will solve the problems/latency outside of the core, and maybe improve the perf/watt figures..

Ofcourse in some years, ARM will also have a better design.

[PerformanceExpert]

yes the cortex-a73 is a somewhat old, or not so recent, at least is not the top of the line from ARM now..

This actual generation already beat the cortex -A73 in absolute perf, ...what we don't know is perf/watt.
But I assume the next generation will solve the problems/latency outside of the core, and maybe improve the perf/watt figures..

Ofcourse in some years, ARM will also have a better design.

Right now they have a 1.6GHz version, so it's not beating Cortex-A73 in absolute performance. That slide shows power consumption per core but you'd want to compare total power consumption while running SPEC.

[tuxd3v]

Right now they have a 1.6GHz version, so it's not beating Cortex-A73 in absolute performance. That slide shows power consumption per core but you'd want to compare total power consumption while running SPEC.

thanks for the link..
In previous paper they states 2-2.5Ghz, what this link shows is a transition from fpga to 28nm asic @ 1.6Ghz..with a CoreMark of 7.1CoreMark/Mhz..
I got the impression from the previous paper that their design can scale from 2 - 2.5Ghz, but in that timeline scale they are still to introduce the 12nm FinFET node process..by 2020-09..
Tough we don't know if there are already internal 12nm samples, if it don't, then they have a bit more work to do to achieve the 2.5Ghz in the next iteration..
And for what it seems,
This end product will still come with the 10% slower Performance in SPECInt( in my opinion is due to latency outside the core.. ), but will be then faster than Cortex-a73 @2.5Ghz..

Yes you right, what we want is not only idle power consumption, or power consumption per core only, but consumption per CPU, with all SMP/Interconects/cores working out on [email protected]..
But then, we would need to find the exact same situation for Cortex-a73 to compare..which would be difficult, as ARM protects itself, and you don't find usually CoreMark results for ARM designs..

They state 0.3mW/core/Mhz,
I don't really know with what benchmark( probably with CoreMark..makes sense.. ), which would translate to [email protected] per core( if power was indeed linear, which it isn't in any cpu anyway.. ), but even then, power consumption with SpecInt per CPU, would be a nice addition

Even tough we only have the Wattage per core, it seems to me to be pretty good values 3 [email protected] ignoring the power consumption outside the 4cores, maybe..
ARM is also playing into 0.75W/Core( but I believe those figures were with it running SpecInt ).

One interesting fact about the link you posted... the Asic [email protected] is labelled "C-SKY", which was a company acquired by Alibaba

[oiaohm]

Even if the next generation beats Cortex-A73, remember that is a 4 year old design. Recent Arm cores are more than twice as fast, and improving yearly, so it would be difficult to catch up.

Its not as much as one would think. For work loads like serving web requests that are can be massively multi threaded lots of arm server chips are still A72 not A73. So being faster than a A73 is a heck of a lot faster than A72 common mass multi-thread arm servers.

https://fuse.wikichip.org/news/2339/arm-unveils-cortex-a77-emphasizes-single-thread-performance/2/

A73 is also a 3 instruction decode.

https://www.nextplatform.com/2020/08/21/alibaba-on-the-bleeding-edge-of-risc-v-with-xt910/

Yes A73 and new alibaba risc-v are both 3 instruction decode.

The newer arm cores gain performing by coming wider from 3 instruction to 6 instruction in fly. So double of performance is just double the design to double the through put.

Big important note here Risc-v chip does not have the USA able to mess with means to make it. This chip being A73 level is good enough to replace your web server request arm servers. Trade restrictions does at times mean using good enough instead of the best.

[PerformanceExpert]

Its not as much as one would think. For work loads like serving web requests that are can be massively multi threaded lots of arm server chips are still A72 not A73. So being faster than a A73 is a heck of a lot faster than A72 common mass multi-thread arm servers.

https://fuse.wikichip.org/news/2339/arm-unveils-cortex-a77-emphasizes-single-thread-performance/2/

A73 is also a 3 instruction decode.

https://www.nextplatform.com/2020/08/21/alibaba-on-the-bleeding-edge-of-risc-v-with-xt910/

Yes A73 and new alibaba risc-v are both 3 instruction decode.

The newer arm cores gain performing by coming wider from 3 instruction to 6 instruction in fly. So double of performance is just double the design to double the through put.

Big important note here Risc-v chip does not have the USA able to mess with means to make it. This chip being A73 level is good enough to replace your web server request arm servers. Trade restrictions does at times mean using good enough instead of the best.

Cortex-A72 is actually bigger and wider than A73 (which is only 2-way), and runs at up to 3GHz due to a longer pipeline. Comparing with A72 would be more realistic since it is a server core while A73 is a mobile-only core, but it means the gap is much larger than 10%.

And you don't get twice the performance from doubling the width - improving performance is a lot of hard work on micro architecture. Getting anywhere near the performance of modern server cores would require a huge design team and many years of hard work. So I don't see a RISC-V version of say EPYC or Graviton 2 happening in the next 4-5 years.

The current trade restrictions are purely political - the US is even blocking the sale of Dutch EUV machines to UMC so China can't get access to EUV. Things will improve as soon as Trump is gone.

[oiaohm]

Cortex-A72 is actually bigger and wider than A73 (which is only 2-way), and runs at up to 3GHz due to a longer pipeline. Comparing with A72 would be more realistic since it is a server core while A73 is a mobile-only core, but it means the gap is much larger than 10%.

https://www.androidauthority.com/arm-cortex-a73-gary-explains-694455/

The A73 is the faster core. Sorry A73 is not a mobile only core it has been used in different server core designs. So winning against A72 is pointless to get customers using custom server parts based on A73.

And you don't get twice the performance from doubling the width - improving performance is a lot of hard work on micro architecture. Getting anywhere near the performance of modern server cores would require a huge design team and many years of hard work. So I don't see a RISC-V version of say EPYC or Graviton 2 happening in the next 4-5 years.

But you cannot get twice the performance without increasing the width either. Risc-v does have quite quite few huge design team working on things.

The current trade restrictions are purely political - the US is even blocking the sale of Dutch EUV machines to UMC so China can't get access to EUV. Things will improve as soon as Trump is gone.

https://www.gizmochina.com/2020/07/13/china-new-lithography-5g-chipmaking-discovery/

That blocking EUV access has backfired. Turns out China has developed their own way that does not need the Dutch EUV machines any more heck it don't even infringe on the EUV patents. Remember people thought it would take years to develop alternative EUV tech instead it was only months.

This should serous-ally be a warning that these trade restrictions on technology don't really stand a chance when China can really put a million+ people against a problem and make something to replace it.

If the USA really blocks access to arm and x86 from China the man power and resources china can bring to bare on the problem means Risc-v getting from A72-A73 to EPYC/Graviton 2 being measured in months not years. That is if the USA is foolish enough of course to cut of china access.