Paul A. Clayton (paaronclayton@gmail.com) on 1/13/11 wrote:
---------------------------
>Wilco (Wilco.Dijkstra@ntlworld.com) on 1/12/11 wrote:
>---------------------------
>>Paul A. Clayton (paaronclayton@gmail.com) on 1/12/11 wrote:
>>---------------------------

>I believe Linus is comparing the highest performance
>ARM currently available against even a typical laptop x86--
>not with the lowest performance non-embedded x86 available.

Well according to your CoreMark numbers a mobile phone ARM already beats a Celeron, ie. low-end laptops. We already know ARM is competitive with Atom today - the whole thread is about ARM becoming competitive with low/mid-end x86 in the next few years.

You might find this A15 info interesting:

http://www.arm.com/files/pdf/AT-Exploring_the_Design_of_the_Cortex-A15.pdf

>>Being 8% slower is not only a tiny gap, it is Atom which is struggling to be competitive
>>as it needs a 60% faster clock.
>
>Once again, frequency (or even power use) is not an
>issue when comparing performance alone.

We're not comparing just performance. If performance was the only issue then AMD would be long gone and we'd all be using Alpha or Power. I'm typing this on a 1.8GHz Athlon64. Beating average laptops/desktops is all that is required.

>While the gap should be narrowed--it is much
>easier to increase performance at a lower starting
>performance (doubling issue width is much easier and has
>much greater benefit going from 1 wide than from 4 wide;
>doubling frequency is easier and has much greater benefit
>starting from a 5-stage pipeline than from a 20-stage
>pipeline)--and there are diminishing returns for the end
>user as well,

Exactly, this is why ARM can quickly close the gap. But leapfrogging or even becoming competitive with high-end x86 is near impossible.

>such does not mean that there will be a
>single-core ARM breaking 10000 CoreMark (which is not
>exactly a real-world benchmark) in the next few years.

Of course ARM will break the 10000 CoreMark in the next few years - afterall, a 2GHz A9 does already ~5800. The A15 has 2.5x the frequency and ~50% better IPC, so 2881 * 2.5 * 1.5 = ~10800.

Consider this: a 1GHz T1 does ~19000 (OK it's old but that's just 2375 per core...), a single-threaded 3.2GHz i5 does ~13000, Xeons do ~20000 per core (multithreaded). While I agree CoreMark is not the best benchmark, the gap is not that wide. With a 2GHz A9, you'd need 4 times as many cores to beat the above. With the A15, it's only twice.

>>Tegra2 uses A9. The GCC score is a bit worse than the published result of 2881.
>>Also, if you use a multi-threaded Atom then it is also fair to use a dual core A9.
>
>I did not see that result--the Tegra2 result I did find
>used 2 threads on 2 cores (5866.39-->2933.20). The halved
>2 core result implies a speed gain of less than 11% (not
>the 25% projected by some for the A9 improvement over A8
>[though the original Tegra may have used an enhanced A8]).
>Your figure gives less than a 9% improvement.

I've no idea how you could possibly come to that conclusion. The 600MHz A8 figure is 1221, the 1GHz A9 does 2933. That means the A9 is 44% faster than the A8 on CoreMark/MHz, far more than the expected 25%, right?

Note Tegra1 uses ARM11, Tegra2 has a dual core Cortex-A9. A8 has never been used in Tegra.

>ARM is currently only targeting such low-power systems, but
>some seem to imply that ARM will take the world by storm
>in a very short period of time. I think Linus is reacting
>to this kind of claim.

ARM has already taken over the world by storm. They just stayed out of the x86 space until now.

>>In servers, performance per Watt, performance per m^3, or performance per dollar
>>is often far more important than maximum single threaded performance.
>
>But there is also some desire for a single binary to run
>across a range of systems.

In the server space people are more willing to compile for a new architecture. Companies like Google don't buy off the shelf x86 software. If they can reduce their datacentre bills significantly by using ARM, they will.

>Note also that Sun/Oracle's
>T-series is not taking over the datacenters of the world.
>Some server workloads demand significant single-thread
>performance, even if only for shortish bursts.

Sure, that's why ARM is improving single-thread performance significantly. They don't have to beat the high end, just be good enough to make the switch from x86 worthwhile.

Wilco