The prime-crunching on dedicated hardware FAQ

[markr]

Originally Posted by jasong

... he optimized the code for INTEL CPUS.

No, not true. There have been significant optimisations for AMD CPUs as well as Intel. Probably more work has been put into code for Intel CPUs, but the installed & potential user-base are both majority Intel, so there's more benefit to the project from the programming time.

An example: for ages only Intel CPUs had SSE2, which gave a massive advantage at LL testing. GW would have been mad not to put in the time to use the SSE2 instructions - the first version of prime95/mprime to use them was three times faster than the previous version, according to whatsnew.txt. Other changes in the same version benefited Athlons by a mere 20%. Was that part of the conspiracy too?

On the other hand, AMD CPUs run rings around similarly-clocked Intel CPUs at TF up to 2^62. [This may have changed a bit recently, but was certainly true for years.] This was largely because the software was written to take advantage of particular features in the hardware.

So you can't say AMD has been totally neglected. I run prime95/mprime on AMD and Intel hardware, and as Batalov says, it doesn't suck on AMD.

Originally Posted by jasong

... GW is either lying(yes, I said it) or mistaken when he says making a graphics cards implementation isn't worthwhile.

It's not just GW saying that. Are they all in on the conspiracy or mistaken? [Wrong answer gets a nice big o rly.] You either think they are, or change your conclusion.

It's not like someone wanting to write million-digit prime testing code for a graphics card would have to start from scratch. All the relevant parts of the prime95 code are available to read & learn from, even if it's not open source, and it's not the only FFT-type code around. Why do you think no-one has done it yet?

Disclaimers
1... I have been a happy user of AMD CPUs for years. And Intel ones. I've seen strengths & weaknesses in each.
2... My default position on conspiracy theories is disbelief - they just never seem to have any evidence.

[ET_]

I noticed that standard GMP code runs faster on AMDs than on Intels. That may mean that general implementation on Intels sucks. Kudos to George who analyzed and solved the problem. I read a document written by George and dealing with assembly optimizations on Intel/AMD platforms: he spotted and manually corrected a couple errors present on the Intel machine code manual.

Luigi

[cheesehead]

Originally Posted by jasong

My apologies

You owe some apologies to George for what you say later.

The slowdown on AMD hardware with his code is so pronounced

What slowdown, exactly? I'm not aware that any has been reported. If you have personally experienced a "pronounced" slowdown on AMD, please report it instead of slandering George. If you haven't, then by what evidence do you write as though this slowdown were an established fact?

that one might conclude that there's a George Woltman conspiracy going on.

Like many other conspiracy theories, this seems to lack evidence, and points instead to a unfounded bias on your part.

Suffice it to say, building new code specifically for AMD hardware probably isn't something anyone would want to undertake, so it's doubtful someone will show up with good AMD code.

As others have pointed out, George _has_ included code specifically to improve AMD performance. Unless you show us evidence to back up your claims, you owe George a big apology.

10-20 graphics cards could run circles around everyone in GIMPS, including the teams.

Show us figures and specific details to support that assertion. According to all I know, it's simply false.

I don't have the skills to build a graphics card implementation, and the person I got my information from doesn't want to come forward, but GW is either lying(yes, I said it) or mistaken when he says making a graphics cards implementation isn't worthwhile.

So, you disallow the possibility that what is worthwhile for one person may not be worthwhile for another person? On what basis?

He's either greedy for the prize or is sick of working on Prime95, in my opinion.

That opinion is arrogant, slanderous and based on ignorance.

This is the worst posting I've ever seen from you, Jason.

Do you understand the principle of providing data to support a claim? If so, then apologize for not showing us evidence to back up your claims.

[davieddy]

Originally Posted by cheesehead

This is the worst posting I've ever seen from you, Jason.

That's saying something

The FAQ was created with him and his likes in mind.
Has he read them? (rhetorical question).

[jasonp]

Jason, there's no room in the FAQ for trolling or magical theories. Any more such posts in this thread will be deleted.

[jasong]

Originally Posted by jasonp

Jason, there's no room in the FAQ for trolling or magical theories. Any more such posts in this thread will be deleted.

All it would take to prove you people wrong would be a number-testing program that runs on a graphics card that tests numbers faster than GW's code correct? Since graphics cards ARE faster and better at the task prime-finding programs excel at(FFTs), simply getting the program to run on a graphics would take one most, if not all, the way to proving the whole double-precision theory wrong.

I have a lot of free time, and one of the things I intend to spend a lot of time on in the next few months is learning how to program in Linux. Would this forum like to challenge me to write a number-testing program for graphics cards?

[ET_]

Originally Posted by jasong

All it would take to prove you people wrong would be a number-testing program that runs on a graphics card that tests numbers faster than GW's code correct? Since graphics cards ARE faster and better at the task prime-finding programs excel at(FFTs), simply getting the program to run on a graphics would take one most, if not all, the way to proving the whole double-precision theory wrong.

I have a lot of free time, and one of the things I intend to spend a lot of time on in the next few months is learning how to program in Linux. Would this forum like to challenge me to write a number-testing program for graphics cards?

Jason, we would LOVE to be proven wrong, because that would mean that graphic cards are THE way for prime proving activities. All we know at this time is that, with actual hardware and software, this is not feasible.

If you think you can disprove our (mis)conception, feel free to do it. We will thank you should you succeed.

Luigi

[Mini-Geek]

Originally Posted by jasong

All it would take to prove you people wrong would be a number-testing program that runs on a graphics card that tests numbers faster than GW's code correct? Since graphics cards ARE faster and better at the task prime-finding programs excel at(FFTs), simply getting the program to run on a graphics would take one most, if not all, the way to proving the whole double-precision theory wrong.

I have a lot of free time, and one of the things I intend to spend a lot of time on in the next few months is learning how to program in Linux. Would this forum like to challenge me to write a number-testing program for graphics cards?

Go ahead and try. I hereby whereas fortowith and whatnot challenge you. I hope you can write a program for graphics cards that is faster than Prime95 (give me more of an excuse to buy good graphics cards for gamin- I mean prime hunting ), but I seriously doubt it because of all the reasons posted in the first post of this thread.

[xilman]

Originally Posted by jasong

I have a lot of free time, and one of the things I intend to spend a lot of time on in the next few months is learning how to program in Linux. Would this forum like to challenge me to write a number-testing program for graphics cards?

I can't speak for the forum as a whole, but I would like to challenge you to produce a program which runs on a graphic card of your choice and which can LL-test a prime exponent in the range 30M to 40M at least as fast as one-tenth of the speed of George's code running on a single-core 2GHz Intel cpu. The benchmark test is to be conducted by a mutually agreed third party or third parties.

So that it's not an open-ended challenge, let's stipulate that you have to demonstrate such a program before the end of August 2009.

If you succeed in your challenge, we'll have a whip-round and pay you the price of the graphics card on which your program runs. If you fail, you agree to stop making these sorts of claims. OK?


Paul

P.S. Any others reading this forum prepared to put their money where their mouth is?

[Mini-Geek]

Originally Posted by xilman

... which can LL-test a prime exponent in the range 30M to 40M at least as fast as one-tenth of the speed of George's code running on a single-core 2GHz Intel cpu.

...

P.S. Any others reading this forum prepared to put their money where their mouth is?

I don't like to make bets, but if he can make it faster than Prime95 (which I'll define as that I can get more work done on my dual Athlon running his than CPU, especially if it's LLR not just Prime95, but I'd expect this would be general multiplication code and an LLR-like implementation would be fast in the coming if this was made), I'd be happy to donate, say, $10-$20 (US) to the genius programmer. I know it's nothing much, but I don't exactly get much money. If it's close to those specifications but not quite by the deadline, and optimization makes it closer/at/past those within another year or two I'll probably donate.

[ET_]

Originally Posted by xilman

I can't speak for the forum as a whole, but I would like to challenge you to produce a program which runs on a graphic card of your choice and which can LL-test a prime exponent in the range 30M to 40M at least as fast as one-tenth of the speed of George's code running on a single-core 2GHz Intel cpu. The benchmark test is to be conducted by a mutually agreed third party or third parties.

So that it's not an open-ended challenge, let's stipulate that you have to demonstrate such a program before the end of August 2009.

If you succeed in your challenge, we'll have a whip-round and pay you the price of the graphics card on which your program runs. If you fail, you agree to stop making these sorts of claims. OK?


Paul

P.S. Any others reading this forum prepared to put their money where their mouth is?

I am prepared to pay up to half the price of the (commercial) graphic card he uses under this challenge.

Luigi