Slashdot Mirror

http://www.spec.org/cpu2006/Docs/readme1st.html#Q15

There are several different ways to measure computer performance. One way is to measure how fast the computer completes a single task; this is a speed measure. Another way is to measure how many tasks a computer can accomplish in a certain amount of time; this is called a throughput, capacity or rate measure.
The SPEC speed metrics (e.g., SPECint2006) are used for comparing the ability of a computer to complete single tasks.
The SPEC rate metrics (e.g., SPECint_rate2006) measure the throughput or rate of a machine carrying out a number of tasks.

For the rate metrics, multiple copies of the benchmarks are run simultaneously. Typically, the number of copies is the same as the number of CPUs on the machine, but this is not a requirement. For example, it would be perfectly acceptable to run 63 copies of the benchmarks on a 64-CPU machine (thereby leaving one CPU free to handle system overhead).

SpecInt2006 and SpecFP2006 are both single core benchmarks, it is only SpecRate which is multicore. In which case this shows a 3Ghz Nehalem has more performance per core than a 5Ghz PPC. That's not too unexpected actually, Nehalem is out of order and POWER6 isn't.

Looking at int rate for 4 cores, which really is a multicore benchmark,

http://www.onscale.de/specbrowser/2006-ir-004.html

it seems like Nehalem is still ahead, but not by as much. Same with int rate for 8 CPUs

http://www.onscale.de/specbrowser/2006-ir-008.html

Now for FP rate the picture is different

http://www.onscale.de/specbrowser/2006-fr-004.html

Power6 does indeed come out ahead. And for the rate scores with more than 16 cores, pretty much all the scores are for Risc because no one bothers to make x86 machines with that many cores. Mind you Larrabee might change that. Larrabee is in order too, has loads of cores and a wide vector FP unit. If all the rhetoric about being vector complete is true, it seems like it would score very well on Spec FP rate for lots of cores, something x86 is still weak at.

http://www.spec.org/cpu2006/Docs/readme1st.html#Q15

There are several different ways to measure computer performance. One way is to measure how fast the computer completes a single task; this is a speed measure. Another way is to measure how many tasks a computer can accomplish in a certain amount of time; this is called a throughput, capacity or rate measure.
The SPEC speed metrics (e.g., SPECint2006) are used for comparing the ability of a computer to complete single tasks.
The SPEC rate metrics (e.g., SPECint_rate2006) measure the throughput or rate of a machine carrying out a number of tasks.

For the rate metrics, multiple copies of the benchmarks are run simultaneously. Typically, the number of copies is the same as the number of CPUs on the machine, but this is not a requirement. For example, it would be perfectly acceptable to run 63 copies of the benchmarks on a 64-CPU machine (thereby leaving one CPU free to handle system overhead).

SpecInt2006 and SpecFP2006 are both single core benchmarks, it is only SpecRate which is multicore. In which case this shows a 3Ghz Nehalem has more performance per core than a 5Ghz PPC. That's not too unexpected actually, Nehalem is out of order and POWER6 isn't.

Looking at int rate for 4 cores, which really is a multicore benchmark,

http://www.onscale.de/specbrowser/2006-ir-004.html

it seems like Nehalem is still ahead, but not by as much. Same with int rate for 8 CPUs

http://www.onscale.de/specbrowser/2006-ir-008.html

Now for FP rate the picture is different

http://www.onscale.de/specbrowser/2006-fr-004.html

Power6 does indeed come out ahead. And for the rate scores with more than 16 cores, pretty much all the scores are for Risc because no one bothers to make x86 machines with that many cores. Mind you Larrabee might change that. Larrabee is in order too, has loads of cores and a wide vector FP unit. If all the rhetoric about being vector complete is true, it seems like it would score very well on Spec FP rate for lots of cores, something x86 is still weak at.

http://www.spec.org/cpu2006/Docs/readme1st.html#Q15

There are several different ways to measure computer performance. One way is to measure how fast the computer completes a single task; this is a speed measure. Another way is to measure how many tasks a computer can accomplish in a certain amount of time; this is called a throughput, capacity or rate measure.
The SPEC speed metrics (e.g., SPECint2006) are used for comparing the ability of a computer to complete single tasks.
The SPEC rate metrics (e.g., SPECint_rate2006) measure the throughput or rate of a machine carrying out a number of tasks.

For the rate metrics, multiple copies of the benchmarks are run simultaneously. Typically, the number of copies is the same as the number of CPUs on the machine, but this is not a requirement. For example, it would be perfectly acceptable to run 63 copies of the benchmarks on a 64-CPU machine (thereby leaving one CPU free to handle system overhead).

SpecInt2006 and SpecFP2006 are both single core benchmarks, it is only SpecRate which is multicore. In which case this shows a 3Ghz Nehalem has more performance per core than a 5Ghz PPC. That's not too unexpected actually, Nehalem is out of order and POWER6 isn't.

Looking at int rate for 4 cores, which really is a multicore benchmark,

http://www.onscale.de/specbrowser/2006-ir-004.html

it seems like Nehalem is still ahead, but not by as much. Same with int rate for 8 CPUs

http://www.onscale.de/specbrowser/2006-ir-008.html

Now for FP rate the picture is different

http://www.onscale.de/specbrowser/2006-fr-004.html

Power6 does indeed come out ahead. And for the rate scores with more than 16 cores, pretty much all the scores are for Risc because no one bothers to make x86 machines with that many cores. Mind you Larrabee might change that. Larrabee is in order too, has loads of cores and a wide vector FP unit. If all the rhetoric about being vector complete is true, it seems like it would score very well on Spec FP rate for lots of cores, something x86 is still weak at.

Quad core Power6 systems from 2007 beating quad core i7 Extreme Edition systems from 2009 in floating point performance:

http://www.onscale.de/specbrowser/2006-fr-004.html/

Not just integer but also floating point; the linked chart shows older Power6 4.7Ghz speed chips schooling intel's nehelem chips in floating point performance:

http://www.onscale.de/specbrowser/2006-fr-004.html/

Here are the SpecInt scores

http://www.onscale.de/specbrowser/2006-i.html

The top score is an Intel Xeon X5570 at 2933Mhz
SPECint2006 = 36.3
SPECint_base2006 = 32.2

Look at SpecFP

http://www.onscale.de/specbrowser/2006-f.html

The same chip is on top there too

SPECfp2006 = 42.0
SPECfp_base2006 = 39.3

Here are the results you linked to for a 5Ghz PPC were

SPECfp2006 = 24.9
SPECfp_base2006 = 20.1

So even at SpecFP where Risc has traditionally been quite a bit ahead, x86 is now on top. On SpecInt it's been like that for ages, at least since Athlon64.

Here are the SpecInt scores

http://www.onscale.de/specbrowser/2006-i.html

The top score is an Intel Xeon X5570 at 2933Mhz
SPECint2006 = 36.3
SPECint_base2006 = 32.2

Look at SpecFP

http://www.onscale.de/specbrowser/2006-f.html

The same chip is on top there too

SPECfp2006 = 42.0
SPECfp_base2006 = 39.3

Here are the results you linked to for a 5Ghz PPC were

SPECfp2006 = 24.9
SPECfp_base2006 = 20.1

So even at SpecFP where Risc has traditionally been quite a bit ahead, x86 is now on top. On SpecInt it's been like that for ages, at least since Athlon64.

The problem of distributed consistency has kept researchers occupied for quite a while. For example, see project Scalaris. They are using a distributed hash table to distribute data among many nodes. This should be relatively easy, at least once you have a good hashing function on your hands. But a lot of research has been done on P2P networks during the last decade, so there is quite a lot of stuff to read and take ideas from.
The interesting part is that it can maintain consistency and support ACID properties. From the site it appears that they accomplish that by using a modified Paxos Algorithm which basically is a way to maintain consensus among many different peers in a non-Byzantine system (this means that there are no malevolent peers in the system -- peers can break down and cease working but not sabotage the system). Leslie Lamport of Microsoft Research has done a lot of work on this, anyone interested may take a look at his papers, very advanced stuff there.

IBM's Power6 destroys everything out there for FP (and has for several years now) but you need like 5 grand+ for an entry level system. Here are the specfp2006 rate results:

http://www.onscale.de/specbrowser/2006-fr-004.html

Your best bet is probably an i7 though.

I think if IA-64 ever achieved the kind of volume the x86 market has, it would end up being a fine processor with lots of room for improvement still. It never really stood a chance: it was marketed as a server processor and Microsoft offerer only a half-assed support for it (it's their best interest to keep computers a commodity and they will fight any attempt to differentiate in that space). In addition, by the time it could be a viable high-power desktop workstation for developers or data-crunchers (a space it shines in) there was no Fedora or Ubuntu for it.

Instead, AMD came out with a set of extensions to the crufty x86 and that is what we use today. We would be much better if we started from a clean sheet.

And much, much better, if binary compatibility to x86 wasn't such a big issue.

None of that is true. Microsoft ported NT based kernels to Itanium (and spent vast amounts of time doing so because there are some subtle issues). Still since it was made by Intel it was pretty much guaranteed to get Windows support.

An Opteron 246 had about the same SpecInt as an Itanium 2 even when both were running native code.

http://www.theinquirer.net/en/inquirer/news/2003/11/05/amd-benchmarks-opteron-against--itanium-xeons-piiis

An Itanium was much slower running x86 binaries. Even the second generation run x86 binaries slowly

http://www.builderau.com.au/news/soa/Itanium-loses-x86-hardware-support/0,339028227,339230300,00.htm

Microsoft Windows and major Linux versions include IA-32 EL. The emulation layer is considerably slower than a modern Xeon however: A 1.5GHz Itanium 2 processor runs emulated x86 instructions at about the same speed as a 1.5GHz Xeon processor, according to Intel.

At that point the fastest Xeon was much faster than 1.5Ghz

Opteron systems were much cheaper

http://www.tomshardware.com/forum/57718-28-opteron-kill-itanium

and they tended to win on real world benchmarks

http://www.infoworld.com/article/03/08/01/30FE64linux_3.html

Basically Itanium was a chance for a company with vast resources to start from scratch and it wasn't faster than x86. The Risc chips that NT supported actually had a better performance advantage, at one point up to 2x the SpecInt. And that wasn't enough to get people to bear the pain of switching over.

The fact is you can't judge computer architecture by aesthetic principles. x86 and x64 may look ugly but that is subjective. The thing that counts is performance and x86 has been beating competing architectures on SpecInt for ages.

Amd64 vs Ia64 was particularly dramatic. Intel had a huge financial advantage and at one point desktop Athlon 64s were the fastest processor in the world, beating far more expensive Ia64 server processors. It's the same now with Nehalem -

http://www.onscale.de/specbrowser/2006-i.html

it beats far more expensive non x86 chips, including ones from Intel.

Actually it wins on FP now, which is something that non x86 chips tended to do well at

http://www.onscale.de/specbrowser/2006-f.html

It's easy to say that it would be easy to start from a clean sheet, but Intel has tried that, poured money into it got the entire industry (including Microsoft) to announce transition plans from x86 to Ia64 and it still failed. Hell Ia64 isn't even that aesthetically pleasing, the more you look at it the more crufty it is.

I think if IA-64 ever achieved the kind of volume the x86 market has, it would end up being a fine processor with lots of room for improvement still. It never really stood a chance: it was marketed as a server processor and Microsoft offerer only a half-assed support for it (it's their best interest to keep computers a commodity and they will fight any attempt to differentiate in that space). In addition, by the time it could be a viable high-power desktop workstation for developers or data-crunchers (a space it shines in) there was no Fedora or Ubuntu for it.

Instead, AMD came out with a set of extensions to the crufty x86 and that is what we use today. We would be much better if we started from a clean sheet.

And much, much better, if binary compatibility to x86 wasn't such a big issue.

None of that is true. Microsoft ported NT based kernels to Itanium (and spent vast amounts of time doing so because there are some subtle issues). Still since it was made by Intel it was pretty much guaranteed to get Windows support.

An Opteron 246 had about the same SpecInt as an Itanium 2 even when both were running native code.

http://www.theinquirer.net/en/inquirer/news/2003/11/05/amd-benchmarks-opteron-against--itanium-xeons-piiis

An Itanium was much slower running x86 binaries. Even the second generation run x86 binaries slowly

http://www.builderau.com.au/news/soa/Itanium-loses-x86-hardware-support/0,339028227,339230300,00.htm

Microsoft Windows and major Linux versions include IA-32 EL. The emulation layer is considerably slower than a modern Xeon however: A 1.5GHz Itanium 2 processor runs emulated x86 instructions at about the same speed as a 1.5GHz Xeon processor, according to Intel.

At that point the fastest Xeon was much faster than 1.5Ghz

Opteron systems were much cheaper

http://www.tomshardware.com/forum/57718-28-opteron-kill-itanium

and they tended to win on real world benchmarks

http://www.infoworld.com/article/03/08/01/30FE64linux_3.html

Basically Itanium was a chance for a company with vast resources to start from scratch and it wasn't faster than x86. The Risc chips that NT supported actually had a better performance advantage, at one point up to 2x the SpecInt. And that wasn't enough to get people to bear the pain of switching over.

The fact is you can't judge computer architecture by aesthetic principles. x86 and x64 may look ugly but that is subjective. The thing that counts is performance and x86 has been beating competing architectures on SpecInt for ages.

Amd64 vs Ia64 was particularly dramatic. Intel had a huge financial advantage and at one point desktop Athlon 64s were the fastest processor in the world, beating far more expensive Ia64 server processors. It's the same now with Nehalem -

http://www.onscale.de/specbrowser/2006-i.html

it beats far more expensive non x86 chips, including ones from Intel.

Actually it wins on FP now, which is something that non x86 chips tended to do well at

http://www.onscale.de/specbrowser/2006-f.html

It's easy to say that it would be easy to start from a clean sheet, but Intel has tried that, poured money into it got the entire industry (including Microsoft) to announce transition plans from x86 to Ia64 and it still failed. Hell Ia64 isn't even that aesthetically pleasing, the more you look at it the more crufty it is.