Will 1 sec 1 Million Super Pi ever be achieved?

[Frank_Rizzo]

The CPUs I have had over the years have hit 1M SuperPi times of 40s, 30s, 20s and this week just under 8s.

This got me thinking. Will there ever be a 1 second time for 1M SuperPi?

Based on pure clock speed it looks very much unlikely. The 40/30/20 figures were obtained from 2Ghz/3Ghz/3.8Ghz clock speeds. This sub 8s figure is from a 2500K running at 4.8Ghz

It is now getting so much harder to double clock speed and thus halve SuperPi time. In order to get 1s from 8s the time has to be halved, halved, and halved again. This would mean that the 4.8Ghz speed would have to be doubled, doubled, and doubled again = 38.4Ghz

As SuperPi is single thread, and the clock speed is the biggest determining factor then it does seem that 1 second will never be achieved. At best 5 or possibly 4+ seconds will be world record holders forever.

This is the same as the world record for 100 metres. It will get to a point where humans will no longer physically run any faster, and thus the world record may never be beat (unless the times are started to be recored in .00001 seconds)

 

[Frank_Rizzo]

Uh for SuperPi clock speed is everything. Single thread.

If clock speeds physically hit a limit of 5/6/7Ghz then I can not see how SuperPi could ever break 4/5 seconds.

 

[Frank_Rizzo]

Cache size important too and this has helped figures over the years.

IMO 1 second will never be achieved with current home use technology. 10GHz+ surely can't be achieved and no new chipset, increase in RAM bandwidth and decrease in latencies will be sufficient enough to achieve 1 second.

Yes there are other ways of calculating pi figues. And SuperPi is now probably as irrelevant as achieving the highest score on Donkey Kong, whilst everyone else is amassing points in some MMO.

But just for the record. I'm sure there will be a world record set for SuperPi of around 4+ seconds and it will never be beat. That's the challenge ...

 

[Frank_Rizzo]

So it may be possible to get 38Ghz and thus 1 second in about ... 15-20 years?

I'm still not convinced. The speed barrier seems to be the biggest hold back at the moment. In the 80's and 90's it was easy to double processor speeds from 16Mhz to 33Mz .. 100Mhz, 200Mhz, 400Mhz. It sort of followed Moore's Law.

But in the past 5 or so years clock speed has been harder to increase and thus multi cores, bigger caches are used for performance gains. Thus with current architecture I don't think there will ever be stock speeds of 10Ghz, let alone the 30Ghz+ which would need to be required for 1M SuperPi.

It's like when the first jet engines were trialled. The sound barrier was a restriction. Of course, mach X speeds were achieved later once restrictions were overcome but it was decades later that it was available for commercial use.

1 second may be possible in 20 years but it would be a totally different architecture to todays technology. Maybe silicon has had it's days.

 

[Frank_Rizzo]

I was going to offer odds of 100/1 that it would not be achieved within 10 years.

I think there would be too many takers

 

[Frank_Rizzo]

And skimming threads without reading all of the information is such a simple taks that I am astonished some people do it.

SuperPi is single thread. The biggest influence on the performance of the app is the clock speed. The higher the clock speed the better the results. Adding 4, 8, 12 cores, increasing the cache by a Meg or so, increasing memory bandwidth will hardly make any impact on the results.

Look at the results on the hwbot site. There is a very strong correlation between performance and clock speed.

 

[Frank_Rizzo]

What's the point in continuing to conquer Everest?
What's the point in testing the new Focus over 0-62?
What's the point in continuing the Paris to Dakar Rally?
What's the point in benching rigs with SuperPi?

It's a challenge to be taken on.

My OP point is that I don't think 1 second will ever be achievable. This is much in the same way that no human with current body (chipset) and mindset (processor) will ever break the 3 minute mile, as there has to be a point where current physicality (technology) can no longer achieve a record.

I think that a 4 second 1M SuperPi will be the fastest ever and not ever beaten. If 4s is achieved it will be by silicon on something which would run at around 8-9Ghz, possibly in 2012/13 and that is it.

Technology always improves but it would need a sea-change in structure to be able to break the 1s 1M SuperPi barrier. A pointless exercise maybe. But for users which require pure grunt performance, rather than multiple processes handling all things at once, it is important.