Making an ordinary Athlon run as fast as a P4, per clock

[Dresdenboy]

Originally posted by E_tron
the main reason i was so upset was the fact that the CPU die was larger than a t-bred B. This could affect cooling(it adds up when you use Peltier effect coolers).

I am pleased with the CPU now, because i converted it to a full barton . now if i can do this to an applebred .

I'm not sure about this, but it is likely, that the disabled cache won't consume any power. It just depends on the way, how cache is disabled.

A cache reactivation on Applebred has been done by some hardware reviewers on the web (maybe x-bit), so it is possible.

[PrimeCruncher]

Originally posted by PageFault
Newsflash:

Checking logic before posting would have shown you that, once normalized for mean speed and cpu efficiency,

P4 18134 machines
AMD 8490
P3 1210
Cel 1440

First off, thanks for the support S00113. Second off, I too would like to know where those numbers come from.

Third, considering you're so into "logic" try this: According to your numbers P3s still make up over 1200 PCs. That's a LOT of computers and if we could optomize any further for them I would say go for it. P3s haven't been widely available in YEARS and yet there are still a large number of them on PrimeNet. If we're still going to have 1200 Athlons on PrimeNet five years from now I say any optimizations are worth the effort.

[PageFault]

Well, it shows something that you need support from an assshole ... but if you want to know, those numbers are derived from benchmarking the client on various systems I have used over the years. The machines were normalized by MHz to a 3000 MHz P4C and then ranked according to iteration times, i.e., a hypothetical 3000 MHz PIII did on average 44% of the work. Then, cpu speed needs to be considered, the PIII started at 500 MHz and ended at 1000 MHz, so the average speed of 750 MHz is the best (only) assumption. Comparing to the P4, which has an average assumed speed of 2300 MHz (production began at 1400 MHz and is up to 3200 MHz), the net effectiveness of a PIII is about 15% - hardly surprising as a result.

Originally posted by PrimeCruncher
First off, thanks for the support S00113. Second off, I too would like to know where those numbers come from.

There are indeed 1200 P4 equivalents of this hardware type. Look closer: that is about 5% of the total P4's. In five years, Athlon is going to be like a P75 in today's terms. The point which is difficult to make here is future effectiveness - in five years, the Prescott6 / whatever they want to name it will likely have SSE5 and run at 10 GHz. Developers tend to think of this future potential, not the "how do I get more out of the 486's still kicking around" mentality exhibited by some in this thread.

Third, considering you're so into "logic" try this: According to your numbers P3s still make up over 1200 PCs. That's a LOT of computers and if we could optomize any further for them I would say go for it. P3s haven't been widely available in YEARS and yet there are still a large number of them on PrimeNet. If we're still going to have 1200 Athlons on PrimeNet five years from now I say any optimizations are worth the effort.

[PageFault]

Big deal, you managed to pass one team in the past 90 days ... something you would have done anyways regardless of client version. You didn't show how many % attributable to the optimizations either ... phear your mighty PIII's ... not.

BTW if you want to start mud slinging we can go to the soapbox ... so far you just showed that a user with access to antiquated hardware can be a jackass, rather than contributing to the discussion.

Originally posted by S00113

I thought so about PIII too, but guess what! The latest version promised a performace boost for PIII, so I just installed it on my main production cluster. See the result yourself at the ARS "Team overtake report".

[PrimeCruncher]

Originally posted by PageFault
Well, it shows something that you need support from an assshole

I didn't say I liked his attitude but he does make some good points.

... but if you want to know, those numbers are derived from benchmarking the client on various systems I have used over the years. The machines were normalized by MHz to a 3000 MHz P4C and then ranked according to iteration times, i.e., a hypothetical 3000 MHz PIII did on average 44% of the work. Then, cpu speed needs to be considered, the PIII started at 500 MHz and ended at 1000 MHz, so the average speed of 750 MHz is the best (only) assumption. Comparing to the P4, which has an average assumed speed of 2300 MHz (production began at 1400 MHz and is up to 3200 MHz), the net effectiveness of a PIII is about 15% - hardly surprising as a result.

I don't think that's the best or only assumption for figuring the average CPU speed for a given type. Volume shipment records are probably available somewhere.

There are indeed 1200 P4 equivalents of this hardware type. Look closer: that is about 5% of the total P4's. In five years, Athlon is going to be like a P75 in today's terms. The point which is difficult to make here is future effectiveness - in five years, the Prescott6 / whatever they want to name it will likely have SSE5 and run at 10 GHz. Developers tend to think of this future potential, not the "how do I get more out of the 486's still kicking around" mentality exhibited by some in this thread.

Okay, first off, read the thread title. We're talking about making an Athlon work as fast as a P4 per clock. Yes, from a MHz perspective Athlons are and always will be slower. In comparison to a P4 you say there are 8490 Athlons. That's 46% of the P4s. So if we could make Athlons work faster that number would get a lot bigger.

Five years from now if we had a Prescott 6 with SSE5 then we'll probably have an Opteron 4 with SSE4. Therefore any optimizations for the current Opteron are irrelevent by then. You're implying that, because of that, we shouldn't bother. We should, since even after that product is discontinued it will still account for a large number of computers, as the Athlon does. Comparing production of Athlons vs. P4s is irrelevent in this context because we're talking about optimizations that will change those numbers. Therefore the only relevent numbers are the amounts of certain types of machines on PrimeNet.

Now this entire thread seems to be pretty much a moot point since we don't know if George has read it or plans to do anything about it. I think before we get too much further we should get his input and find out.

[S00113]

Originally posted by PageFault
Well, it shows something that you need support from an assshole ...

Merry Christmas to you too. :-)

[S00113]

Originally posted by PageFault
Big deal, you managed to pass one team in the past 90 days ... something you would have done anyways regardless of client version.

I know that. I changed the client three days ago, and you can see the result under 7 Day Output.

You didn't show how many % attributable to the optimizations either ...

I have no idea. I might just be a coincidence that my production increased several years/day at the same time that more than a hundred PIIIs (I don't even know exactly how many) started using the new client.

About your numbers, I can just quote what you think about the graph with the number of clients running on each type of CPU:

BTW the linked plot is irrelevant - it doesn't tell me anything about installed capacity, which is the number required to infer production by cpu type.

I don't think your numbers are any better at showing anything.

BTW if you want to start mud slinging we can go to the soapbox ... so far you just showed that a user with access to antiquated hardware can be a jackass, rather than contributing to the discussion.

Oh, well. I don't think you read much more than the italic text before you wrote this. Sorry about that. At least some people were able to get my points.

By the way.. Most of my production come from P4s. I have two Athlons running mprime. If it was for my personal gain, I would insist on P4 optimizations too. But based on the numbers I can see (number of CPUs, development, etc), I think it would be much smarter to spend the time on optimizing the Athlon code to make it as fast as the P4 by clock.

[PageFault]

I didn't bother reading beyond the personal attacks and instead posted inappropriately. If you can be civil, I can as well and offer an apology for the remarks - stuff like this has no place, even in the soapbox.

Originally posted by S00113
Oh, well. I don't think you read much more than the italic text before you wrote this. Sorry about that. At least some people were able to get my points.

What I'm writing about is distant from personal gains. I too get most of my results from P4's. Rather, the thread opener asks a question and I attempt to reply analytically. So, what needs consideration:

Is the outcome guaranteed?

If yes, how fast?

If a timeframe can be estimated, does the cost justify the benefit?

This is how designers have to think. What I wrote stems from that ... time will tell as to what happens. In a year, it will be seen what the most productive optimizations were.

By the way.. Most of my production come from P4s. I have two Athlons running mprime. If it was for my personal gain, I would insist on P4 optimizations too. But based on the numbers I can see (number of CPUs, development, etc), I think it would be much smarter to spend the time on optimizing the Athlon code to make it as fast as the P4 by clock.

[PrimeCruncher]

Originally posted by PageFault
Is the outcome guaranteed?

If yes, how fast?

If a timeframe can be estimated, does the cost justify the benefit?

Answer to question one: yes, most Athlons will be upgraded and thus work will be completed faster.

Answer to question two: probably within a week or two most computers will be running the new client as people will be looking for any performance gain they can get.

Answer to question three: with a timeframe of six months or so, considering that after the Opteron optimizations are done there will be no other optimizations, the cost of six months does justify the benefit of tens of thousands of Athlons completing work in a shorter period of time.

[moo]

im thinking as i read this we have 2 types p4 we call it gold ok then althon iorn right now standing for p4 its about 23 carrot gold to get pure it would take a lot but we hagve hte althon iorn its mostly still slag so would it be more productive to optimize the gold to pure removeing smal impurites one at a time or the iorn removeing a large ammount of impurites the frist pass then mabey a second pass later on.

refine the althon mabey 10 to 25 precent boost easier and quicker then trying to squeeze life out of p4

[SalemTheCat100]

It's been pretty obvious that PRIME95 development has all but ignored AMD.

The Opteron has been out for 16 months, the Athlon 64 for over a year.
An Opteron was purchased for GIMPS with donated funds by users hoping to see an improved Prime95 program for the AMD64 platform.
Linux is 64 bit NOW.

So where is the optimized release for the AMD64 processors?

NO WHERE , that's where.

If intel had released the Opteron, prime95 would already be running on it six months ago.

And as can be seen by comments in this thread, they consider any PRIME95 system/user who has an Athlon as having an obsolete system. Only the mighty P4 is a non-obsolete system.

My PROTEST that started over six months ago continues.
I have turned off ten of my Athlon systems that were running Prime95.

Maybe more of the other 28000 Athlon Prime95 users can followed suit in protest. If enough Athlon Prime95 users went dark I'd expect to see a different response instead of the usual NOT_WORTH_THE_EFFORT (AMD is obsolete) response. Maybe some actual work would be done in optimizations for the AMD platforms.

SALEM