Massive speed increase attained with compiler optimisations for Opteron

[Urban Ranger]

For its unreleased product, the Portland Group is making performance improvements claims averaging 34% over PGI Workstation 4.1 on several industry standard benchmarks. Now that is a huge hike in performance from just compiler optimizations. The SPEC2000 floating point benchmark shows one Opteron result improving by almost 375%.

link



The new AMD64 compiler makes it easy to port code over to the i86-64 CPU's:

"Porting our software, which represents well over a half million lines of code, from 32-bit X86 processor-based systems to 64-bit AMD Opteron processor-based systems has been very straightforward. We have also been extremely impressed with the performance of both legacy 32-bit applications and 64-bit applications re-compiled to take advantage of the new features of the AMD Opteron processor."

Looks like the table is turned on Intel again.

[Provost Harrison]

*awaits Asher for the fireworks to commence*

Looks cool, I am keeping my eye on processors at the moment as I am looking at buying a system in the near future...but in terms of desktops, it looks like Intel have the lead at the moment...

[Spiffor]

"Opteron".
Does it means it uses optic fiber ?

[Az]

great. now all I need is money

[Wraith]

Well, I am quite looking forward to the A64, but I must point out that the compiler in question really isn't much of a speed demon to start with. Even it's new optimizations are (IIRC) leaving it behind the Intel compiler.

AMD does need to get into this area more, though, so this is a promising start. If they can get some platform specific compiler optimizations going they should get a good long-term benefit out of it.

--"Does it means it uses optic fiber ?"

As much as Intel's "coppermine" chips used copper. Which is to say, no.

Wraith
Shift to the left! Shift to the right! Pop up! Push down! Byte! Byte! Byte!

[Asher]

Re: Massive speed increase attained with compiler optimisations for Opteron

Originally posted by Urban Ranger


The new AMD64 compiler makes it easy to port code over to the i86-64 CPU's:

Looks like the table is turned on Intel again.

How have the tables turned?

And why didn't you post another Opteron article on TheInquirer's front page

Opteron uptake slower than AMD hoped
Stocks building, CPUs gathering dust bunnies
Thursday 03 July 2003, 10:45

[Spiffor]

Aaaaaah
We were waiting for you GLonkie

[Urban Ranger]

Re: Re: Massive speed increase attained with compiler optimisations for Opteron

Originally posted by Asher
How have the tables turned?

Because Opterons will unseat P4 Xeons.

Originally posted by Asher
And why didn't you post another Opteron article on TheInquirer's front page

Opteron uptake slower than AMD hoped
Stocks building, CPUs gathering dust bunnies
Thursday 03 July 2003, 10:45

You mean this one?

Opteron tested against Intel chips

Performance excellent, testers conclude

The programs were all optimised for the Intel processors but not for the Opteron, because he didn't have one to use.

Ho ho ho

[Traianvs]

AMD! AMD! AMD for life!

intel must die!

[Urban Ranger]

Originally posted by Wraith
Well, I am quite looking forward to the A64, but I must point out that the compiler in question really isn't much of a speed demon to start with. Even it's new optimizations are (IIRC) leaving it behind the Intel compiler.

I think the speed gains will be enough to attract people to Opteron servers and workstations over P4/Xeon ones, when further optimisations will yield even better speed increases.

[Asher]

Re: Re: Re: Massive speed increase attained with compiler optimisations for Opteron

Also, the compiler speed improvements they talk about are against the Portland Group's old compiler, not the Intel C++ compiler that AMD uses to submit results.

Originally posted by Urban Ranger
Because Opterons will unseat P4 Xeons.

Where did it say that?

You mean this one?

Opteron tested against Intel chips

Ho ho ho

Again, this is a bit misleading.

According to the author, the AMD system had 1GB of DDR333, the P4 1.7GHz had 256MB of PC800 RDRAM, and the 2.53GHz had 512MB of DDR266. And they have the 533MHz system bus and no HyperThreading.

The author also says the reason why the Opteron is fast in this benchmark is the code fits in the Opteron's 1MB L2 cache but not in the P4's 512KB, coupled with the slow memory and outdated system bus on the P4 it doesn't look so well.

It also doesn't mention how these machines are clustered, for all we know the P4s were connected with 10Mbps LAN and Opterons 1000Mbps...

Not to mention just how plain fishy the numbers are, a 1.7GHz P4 beat a 2.53GHz P4...

[Wraith]

--"Not to mention just how plain fishy the numbers are, a 1.7GHz P4 beat a 2.53GHz P4..."

Actually, this isn't fishy at all. Over on the Ace's Hardware forum some people have been in contact with the guys who ran the benchmark to get further info, and this fits with what their code is. It's got highly unpredictable branching, which means that the P4 takes a rather large hit compared to everything else since it's got such a long pipeline.

Wraith
If you're not part of the solution, you're part of the precipitate

[Asher]

So how does that explain how a 1.7GHz P4 beats a 2.53GHz P4?

[Wraith]

Because the P3 doesn't take nearly the same hit from a branch mispredict, as you know very well. The small L1 cache isn't helping, either.

You should also note that they said their code doesn't use more than 32 MB of memory, so the size of RAM is pretty much irrelevant for these benchmarks, although the type won't be.

The Opterons are running in 32 bit mode on those, btw. They didn't have any 64 bit OSes installed.

Please, Asher, you understand this stuff. Take off your Wintel blinders for a bit and think about the code and the architectures. This is a very specific type of scientific code. Yes, there are some things the P3s will do better than the P4s. There are some areas the P4s will kick butt in, and there are some where they're going to be completely spanked by the Opterons. That's just how it works when you have different design goals.

Wraith
A=B and B=C therefore A=C {except where void or prohibited by law}

[Asher]

Wraith...it's a 1.7GHz P4 beating a 2.53GHz P4.