Intel's Superchilled Test Rig

Barence writes "Last week, PC Pro issued a challenge to see whose PC could render a 3D graphics benchmark in the shortest time. The competition was won by an entrant with a rather unfair advantage: Intel. The processor giant's superchilled rig is overclocked to nearly 5GHz. As PC Pro explains: 'The rig itself uses phase-change cooling: in other words it's attached to a chuffing great freezer, which I believe is the big box on the right of the photo. That yellow meter with the readout is showing the temperature of its output: yes, that's minus 40 degrees Celcius.'"

Comments

In a galaxy far, far away

[PmanAce]

Judging from that photo, we are still in the infancy of computing. The Millenium Falcon looks like that everywhere!

Re:In a galaxy far, far away

[Anonymous Coward]

I believe the politically correct term now is "presidentially-engineered".

Re:In a galaxy far, far away

[Entropy98]

We call it Jimmy-Rigged
 
--
  Nibble Codec Pack

Re:In a galaxy far, far away

[Curtman]

-40 is superchilled? That's what we call "morning" from December through February here.

-40C

[confused one]

No problem. I've got an environmental test chamber which I can use to bring the entire system to -40C. Had it at -75C last week; but, I don't think it could reach that temp with the heat load of a PC. Wonder what numbers I'd get.

Re:-40C

[WrongSizeGlass]

-40C is also -40F ... so -40 it is

Re:-40C

[Anonymous Coward]

then one might make the mistake of -40k

Re:-40C

[petermgreen]

How cool can it get the CPU though? Afaict the CPU usually runs quite a bit hotter than the surrounding air.

Also I wonder how the rest of the PC would survive those temperatures.

Re:-40C

[by (1706743)]

Negative Kelvin is actually possible...

Re:-40C

[McGiraf]

-40k ?

Re:-40C

[arose]

But 'k' isn't it.

Re:-40C

[blueg3]

Sort of.

Re:-40C

[Sulphur]

If dry ice temperatures are low enough, then you can use dry ice in acetone for cooling.

Re:-40C

[BlitzTech]

Acetone will dissolve most of the plastics in computers, and the dry ice/acetone combo only keeps the acetone at CO2's sublimation point of -78C. Not to mention the need for liquid pump system to carry the heated acetone away adds more heat relative to the ventilation fan you'd need in an ultracold freezer.

Re:-40C

[ewanm89]

that's change in energy/temperature, not stable temperature it's running at. It's like I have a 5 litres of water and I take out 2 litres and say the change is -2 litres of water. That's not how much I have though.

Re:-40C

[Sulphur]

We are only trying to reach -40. and as long as solid dry ice is present the temp will be maintained below that.
My cpu is in a ceramic case.
Make a frustrated pyramid with the bottom a heat sink to press against the cpu, and the top closed enough to stop spills but not gas tight. Add styrofoam on the sides to stop frost.
No pump is needed IMO.
If having solvent near plastic is not acceptable, then apply cooling to recirculating air. The motor and bearings of the fan can be external to the airflow.
An experiment carried out without attached computer, and then with a heat source to see how it behaves would be a good idea.

Re:-40C

[nickersonm]

No, it's actually negative temperature. It can occur in systems that have a limit to their possible entropy (number of configurations), where high temperature (simplified: disorder) leads to an equal population of states.

An example would be excited vs. unexcited states of atoms in a lasing medium - a finite number of states with a finite number of configurations means a finite maximum entropy. If one were to heat it towards infinite temperature, the states would become equally populated (maximum possible disorder, or number of configurations). During lasing operation, though, most of the states would be in the higher energy configuration, thus the overall system would be "hotter" (more energetic) than infinite temperature, but actually have less entropy. This is negative temperature; the scale wraps around because of the way temperature is defined with relation to the energy and entropy of a system.

Re:-40C

[confused one]

And therein lies the problem. I did the post kind of tongue-in-cheek. I do thermal tests of instrumentation boards all the time. The system would be at -40 and the board might survive; but, some capacitors on the board might stop functioning. The processor is an interesting problem: Since I'm not directly cooling the cpu I'd have to use an air cooler which brings it's own problems. Without re-plumbing the chamber to supply a cold head, the cpu temp would probably be 30 or 40 degrees above the chamber temp. Even if I tried to crank the chamber to below -75C and put the air cooler in the path of the cold air, immediately after the evaporator, I'd still not see -40 at the cpu; and, at those temperatures you'd start to have problems with the other components on the boards.

Correction

[Alsee]

That yellow meter with the readout is showing the temperature of its output: yes, that's minus 40 degrees celcius.

Correction, it's minus 40 degrees fahrenheit.

-

Re:Correction

[Anonymous Coward]

Man, everyone else who replied to this comment is a fucking idiot. Stop trying to prove someone else wrong and just get the obvious joke.

Re:Correction

[Tridus]

This post deserves bonus points for getting so many people to reply while totally missing the joke.

Re:Correction

[icebike]

Mod parent Swoosh!

Re:Correction

[nacturation]

The post is practically trolling for stupid people.

Re:Correction

[m50d]

It's only obvious if you care enough about archaic measurement systems to know

AMD

[Tubal-Cain]

Any AMD stats for comparison?

Re:AMD

[TheEyes]

Sorry; AMD can't afford a freezer.

Re:AMD

[Threni]

According to one of the comments (about AMD winning $1B+ from Intel) they can afford a fairly substantial freezer...

Re:AMD

[TheEyes]

I'm aware; I've actually invested in some 2012 call options on AMD stock. Even as-is they should be worth $10 a share. If Bobcat can make them competitive in the ultraportable market (Android on ARM is going to eat Intel's lunch in the netbook-level arena; x86's crufty instruction set can't compete at that low level), and/or Bulldozer makes them competitive in the mid- to high-end desktop market, that should go up to $13-15, easy. It is a hell of a gamble, though; they're still almost a full processor node behind Intel, and that's hard to compete with.

Re:AMD

[garyisabusyguy]

Here is what AMD was doing last year with liquid helium, which would put the temp at about 5 degrees Kelvin (about -450 degrees Fahrenheit) and running at 7 giga-hertz

Here is an AMD news blurb
http://eon.businesswire.com/portal/site/eon/permalink/?ndmViewId=news_view&newsId=20091105006606&newsLang=en

And a nifty video
http://www.youtube.com/watch?v=z6Hf6d404QY&f=22

Re:AMD

[RightSaidFred99]

Riiight. Just like the x86's "crufty" instruction set couldn't compete in the scientific computing/HPC domain. Oh, wait... I think you'll find Moorestown already proves you wrong, and Medfield will prove you completely wrong. ARM is (and will be for a long time) dominant in the low end embedded market in terms of power/performance, but it's being threatened from above as the Atom chips get more and more integrated.

Re:AMD

[evilbessie]

IIRC some 'clever' people once used up 150l of liquid Helium to run an AMD Phenom II at something ridiculous like 6.5GHz.

Re:AMD

[Rockoon]

I do remember the days when they said x86 couldn't compete with the RISC chips (alpha/mips/etc..)

The reality is that (A) all processors offer instruction forms that are rarely used, and (B) complex instruction sets increase code density as long as the common ones are short. (C) extensive amounts of registers arent all that useful in most circumstances because most long term data has to be written back to memory anyways.

The read/modify/write instructions are often the best instructions to use these days on x86 because of code density. Instruction scheduling isnt nearly as important as it once was, so RISCifying your x86 assembler code is often a mistake.

Both CISC and RISC have their pro's and con's.

Re:AMD

[Tim C]

It's a minor nit-pick I know, but with a background in Physics I can't help myself - it's not "degrees Kelvin", it's just "Kelvin",.

in Soviet Russia

[FuckingNickName]

environment cools Intel.

Re:in Soviet Russia

[Jeremy Erwin]

Shame that it's no longer winter.

Re:under 60 seconds?

[ak_hepcat]

Dammit, /., i'm logged in. why did you post me as AC?

Sigh.

Whatever.

Re:IS THAT IN CELCIUS OR AMERICAN?

[Anonymous Coward]

40c

The c stands for COMMUNIST!

40f

The f stands for Fuck Yeah!

Big advantage?

[leachlife4]

Phase change cooling is not really that extreme of a cooling system for benchmarking... go to Quakecon you will see quite a few people with it.
LN2 (or even better liquid He) on the other hand could be considered an unfair advantage.

Re:Big advantage?

[Anonymous Coward]

no a really big advantage would be putting the entire rig in a satalite sat in the earths shadow out in the void with a massive superconductor heat sink spreading the heat across 2 or 3 kilometers for really low temps and just beaming the results back via a micro wave trasmiter powerful enough to melt the icecaps... at least that how the borg would do it....

Re:Big advantage?

[Sycraft-fu]

I think they just meant unfair in that it was Intel doing it. First that means they can put whatever funds they like to the challenge. A normal user has limits but Intel could spend an effectively unlimited amount if they so chose. Second they can pick the processor. Maybe this is just a random EE of the shelf, or maybe he tested a bunch and found the best. Intel does make them, after all. Finally Intel's engineers probalby know about the limits and how to reach them better than anyone else.

Re:Big advantage?

[mat128]

If you read TFA (but this is /.), it says he used a retail processor. He was also limited to a single-socket solution, which means no multi-sockets server boards.

Re:Big advantage?

[blueg3]

Liquid helium's not better. Your primary problem at that point is not how low of a temperature you can achieve, but maintaining the temperature while the object you're cooling is producing such large amounts of heat. Liquid nitrogen has a better heat capacity than liquid helium (and is enormously cheaper), so it's going to work better.

Re:Big advantage?

[15Bit]

I'm not sure i'd class Liq He as an advantage. Having worked with it i'd call it a pain in the arse, and totally unsuitable for computer cooling. Low heat capacity and insanely low temp mean it has to be transported and stored in large, very well insulated containers, so it lacks the easy mobility of liq N2. Transferring liq He from one container to another also requires some skill if you don't want to evaporate the whole lot during transfer.

Oh, did i mention it gives you splendid burns too.

Re:Big advantage?

[eeCyaJ]

That's no moon. It's a Beowulf Cluster.

I'm buying real-estate on Pluto

[Tablizer]

it will be the happening place for gamers.

Re:I'm buying real-estate on Pluto

[Arbition]

Assuming it has air... Otherwise, why not in an interstellar void, or even better, intergalactic. Bring on the massiveo heatsinks please, we have a lot of radiating to do.

Re:I'm buying real-estate on Pluto

[GrumblyStuff]

Available energy would also be a concern. Diddly squat for solar and I don't really see there being either local fuel or oxidizers.

Re:I'm buying real-estate on Pluto

[tangent]

Sure, but the ping time is a real bitch.

Re:under 60 seconds?

[ak_hepcat]

user@system:~/src/smallpt$ time ./smallpt 5000
Rendering (5000 spp) 100.00%
real 38m55.591s
user 301m41.140s
sys 0m5.950s

38m isn't too shabby, i guess.

Psh, only 29sec here

[Fry-kun]

on a 2x6core server at work ;)

[xxxx@xxxx smallpt]$ time ./smallpt 100
Rendering (100 spp) 100.00%
real 0m29.127s
user 5m41.044s
sys 0m0.093s

P.S. and compiling didn't take me hours, either, since I'm on Linux

Re:Psh, only 29sec here

[somenickname]

He must have forgotten to turn on optimization (or the MS compiler is *that* bad at optimizing). Even my lowly 2.5Ghz Core2 Duo T9300 renders this in 2m10s on Linux. Core for core, that makes my 2.5 year old laptop chip nearly twice as fast as his i7 980X.

Re:Psh, only 29sec here

[illumin8]

If you RTFA you'll see they intentionally limited the Intel guy to 1 socket. What fun would it be if he ran it on an 8-way server in the lab that you couldn't even buy in real life?

4.94 GHz at 1.62V!!!

[fontkick]

Insane voltage... the 980 is rated up to 1.375V. I'm happy with a i7-860 @ 3.6 GHz running on 1.2V.

Intel's made upgrading much more fun considering you can get a 30-40% CPU speed increase in about 10 minutes of research and bios tweaking. Next fall there will be 8-core/16-threads on the desktop. I am loving Intel these days.

Getting colder...

[sznupi]

For extra effect, they should put Debian 6.0 "Squeeze" there; also recently frozen.

February in Ottawa

[rueger]

So let me understand this. They cool the thing down an average winter temperature in half of Canada, and it's a big deal?

Already I see thousands of gamers running cables out their bedroom windows and leaving their rigs in the snowbank outside.

Re:February in Ottawa

[Shados]

It doesn't get to -40C very frequently in most of the populated (well, more than small villages anyway) regions of canada :). -35 happens, but its far from an average, unless you live all the way north in the middle of nowhere or in the territories.

You are missing the SECRET AGENDA!

[HiggsBison]

"Nearly 5GHz". The whole point here that everyone seems to be missing is that they made something go more than 1000 times as fast as the original 4.77 MHz IBM PC.

Now if they could give it 640MB of memory and a 110MB floppy drive...

Re:You are missing the SECRET AGENDA!

[Korin43]

What's interesting is that 5 Ghz isn't all that impressive a number anymore. People have gotten i5's to 5 Ghz on air.

Re:You are missing the SECRET AGENDA!

[gandhi_2]

What's interesting is that someone has a job where they can just go grab the latest and greatest intel chip whenever they want to, for free.

Re:You are missing the SECRET AGENDA!

[Dogtanian]

"Nearly 5GHz". The whole point here that everyone seems to be missing is that they made something go more than 1000 times as fast as the original 4.77 MHz IBM PC.

Well, only if you're referring to the clock speed. In terms of raw processing speed (by any reasonable benchmark), they're likely *significantly* more than 1000 times faster, since clock speed refers simply to the number of cycles per second, and doesn't account for how much work can be done per cycle.

Re:You are missing the SECRET AGENDA!

[Bobnova]

5ghz is absolutely nothing in the benchmarking world. 980x cpus can generally reach 6ghz cooled by liquid nitrogen, some are close to 7ghz. Single Stage Phase Change cooling (what is pictured) is used by many to rough in an overclock and get familiar with a given chip, but few people use it for their final results as it doesn't get nearly cold enough. Dry ice gets closer, but nothing beats LN2 for 980x cpus.
This is really not very impressive.

Wow...

[Sack]

Imagine a Beowulf cluster of those!

Re:Editing?

[pushing-robot]

Yeah, I had to use this.

Re:I'ld rather have a recently-fabricated HP Alpha

[mat128]

RAMBUS is dead. I remember the old days where you could get a Pentium 4 that used RAMBUS. This shit was always overrated and super expansive. I knew people who had 128MB of RAMBUS (and you had to buy this shit in pair too) who wanted to upgrade to something descent for the times, like 1G. They ended up getting a whole new computer for the price they would have paid for their RAMBUS, and their new computer was much faster than their old one.
Also, Intel EPSD does server stuff. Check it out.

Re:under 60 seconds?

[cynyr]

so, is it multi-threaded? because it seems to use 1-4 cores max, from the benchmark numbers i see....

Re:under 60 seconds?

[SwedishPenguin]

time ./smallpt 100
Rendering (100 spp) 100.00%
real 0m54.164s
user 3m33.343s
sys 0m0.083s

Intel(R) Core(TM) i7 CPU 860 @ 2.80GHz (quad)
4GB RAM
2.6.35-rc6-rc
Arch Linux

With spotify, chromium, a few terminals, KDE and ark also running.

Screw yer Giga-hurts nonsense!

[sherpajohn]

I am more interested a a FPU (food processing unit) than a CPU - how long to render Natalie Portman in hot grits?

-40C is superchilled?!

[Yvan256]

Dudes, don't visit the north of Canada in the middle of winter!

Re:-40C is superchilled?!

[dbIII]

Well, it's 60C colder than what I've got in winter here.

Re:-40C is superchilled?!

[dbIII]

I'm not quite in the tropics so it's winter even if it never gets below zero C.

Re:under 60 seconds?

[Sir_Lewk]

[user@system] ~ $ time ./smallpt 5000
Rendering (5000 spp) 100.00%
real 16m28.799s
user 262m48.590s
sys 0m2.280s :)

Windows optimizations

[tomz16]

There is something seriously wrong with the optimizations in his windows binary...

Ran in 36 seconds on a 4 x 8224 SE AMD opteron IBM x-server running linux (8 total cores at 3.2GHz)

 

Re:Windows optimizations

[illumin8]

Why would you assume that? The engineer from Intel was limited to using a single socket system. I could argue that there is something seriously wrong with your Linux compiled binary since you have 4x as many sockets and ran less than twice as fast.

Re:Windows optimizations

[tomz16]

Why would you assume that? The engineer from Intel was limited to using a single socket system. I could argue that there is something seriously wrong with your Linux compiled binary since you have 4x as many sockets and ran less than twice as fast.

and I could argue that you are flat out wrong
- Intel has 6 cores on the latest architecture running at 5Ghz
- I had 8 cores that are now 1 generation old, running at 3.2GHz, with slower memory (and cores are a little slower clock-for-clock than the i7)

but most importantly :
- most posts from similar hardware show that the people using a linux binary are *several times* faster on identical hardware

It's pretty clear that the author SERIOUSLY messed up the compilation on windows (also betrayed by the fact that it took him several hours of futzing with visual studio in order to get it to work)

Re:Windows optimizations

[Orestesx]

I have replicated your result, 252 seconds in windows, ubuntu VM with access to 4 cores gives the following for Phenom II X4 965 @3.4: sean@sean-ubuntu-2:~/Desktop/smallpt$ time ./smallpt 100 Rendering (100 spp) 100.00% real 1m22.444s user 5m22.656s sys 0m1.688s And yes the images are the same. I agree that windows optimizations are messed up. AMD users fret not.

Re:Windows optimizations

[Rob Riggs]

It's pretty clear that the author SERIOUSLY messed up the compilation on windows (also betrayed by the fact that it took him several hours of futzing with visual studio in order to get it to work)

I'm not sure that it is all that clear. My initial though was that the PC Pro author didn't port the Unix code properly. But there are issues reported with the code on AIX, and I am running into performance issues on SPARC/Solaris with the source code AS IS compiled with both GCC (4.3.2) and Sun Studio. It may be that something about the smallpt source code is preventing the Microsoft compiler from fully optimizing the code. Or the problem may lie with the OpenMP implementations on these platforms. On Solaris, for example, the process never fully utilizes all of the CPU cores.

Real Story: Windows Benchmark is Slow

[Rob Riggs]

The posts from users running Linux on the forum are showing times that are 4-5x faster than those posting benchmarks from Windows. What's going on there?

Re:Real Story: Windows Benchmark is Slow

[B1oodAnge1]

Differences in compilers? I'm also curious.

This is like trying to hold a contest

[mysidia]

Of who can blow up a rack of PCs into the smallest size pieces.

And having the US military unexpectedly enter into the contest, using a tactical nuke to blow up the rack, for their entry in the contest

You know... for PR... to bolster recruitment rates.

.

Same difference... AMD could probably best Intel, if they spent more money on a rig of their own. I think it kind of defeats the point to have large corporations with massive resources the average high-end user could barely dream of seeing in person participate in contests like this

Re:This is like trying to hold a contest

[The boojum]

It's not as though they're competing for a prize here. It's just a friendly competition. You're free to ignore their result and focus on the rest if you like.

Personally, I find this kind of thing interesting as far as seeing what's possible, even if it's not exactly practical.

Re:under 60 seconds?

[Sir_Lewk]

It most certainly is multithreaded, I have it using 16 cores at once on one of the machines I have access to.

In fact, if you look at the code you might not see at first how exactly it's threaded. That's because it's using OpenMP, which it turns out is an absurdly easy and concise way of parallelizing code. Check out this for more info on using OpenMP with gcc, it is really slick stuff.

Weird

[Arancaytar]

I can hardly think at all at that temperature, let alone faster.

OS X Compile?

[appleguru]

I tried to compile and run this on OS X (SL, 10.6.4, gcc 4.2.1). I downloaded the .tar.gz from http://kevinbeason.com/smallpt/, and ran make (which runs g++ with compile flags of -O3 -fopenmp...). It compiled fine. Running it gives a Bus Error though.. any ideas?

Re:OS X Compile?

[JimR]

One of the comments on the smallpt page recommends adding the following line:

if (depth > 100) return obj.e;

after line 55 in the code to prevent a stack overflow.

It seems to do the trick for me.

Re:OS X Compile?

[JimR]

And FWIW compiling with:

g++-fsf-4.4 smallpt.cpp -o smallpt -O3 -fopenmp -ffast-math -m64 -march=native -ftree-parallelize-loops=8 -funroll-all-loops

gives a runtime for smallpt 100 of 2m31s for my MacBook ("Early 2008" model).

Poor code for a benchmark

[hvdh]

I had a 30 minute look at the source code. It's clearly optimized for shortness, not for speed.
There are some obvious performance no-gos, see lines 44-45, using a double variable as a loop counter.
Performance depends to a good extent on the erand48 implementation and whether OpenMP knows that erand48 is MT-safe.

Light speed limit

[mangu]

Clock speed has reached the ultimate physical limit, light speed.

If you take a measuring tape to a motherboard and do some math, you'll see that once we got past a few GHz there's no way a bit can go from one chip to the other within one clock cycle.

The result of that is that chips need local caches and pipelines, etc, until the complexity starts digging into the performance. And power consumption skyrockets.

Still not good enough...

[PFritz21]

And yet, this system still couldn't run StarCraft II without melting the graphics card...

Re:Beated

[Rockoon]

The pcpro smallpt distro is a 32-bit application and uses the 32-bit version of OpenMP.

Re:Beated

[Rockoon]

How big is that overclock?

PCPro's $1000 6-core i7 980 Extreme (which is stock 3.33ghz) took 73 seconds.

My $200 6-core Phenom II 1055T (overclocked to 3.34ghz) took 188 seconds.

Your $200 6-core Phenom II 1055T took 44 seconds.

Thats a big assed difference. I was using the same executable as PCPro (the Win32 version of smallpt)

Not cool

[supermariosd]

Cheating on a benchmark test given is NOT cool, no matter what the temperature of the machine.

Priority?

[istartedi]

Task Manager, Right-click process, set priority to realtime?

Just a guess.

Maybe there's some way to do it on the command-line too. Either that, or the Express version of the compiler doesn't optimize as well. Maybe they built a debug version.

very thoughtful

[stumblingblock]

very thoughtful of them to choose -40 degrees, so those favoring Celsius AND Fahrenheit will understand.

Re:I'ld rather have a recently-fabricated HP Alpha

[Khyber]

they made single-stick + terminator blank pairs. I ever only had one RAMBUS RAM stick.

What ...

[sigmoid_balance]

... Swedish motherfucker, do you speak it? It's Celsius(https://secure.wikimedia.org/wikipedia/en/wiki/Anders_Celsius), not Celcius.