Slashdot Mirror
IBM Ships Fastest CPU on Earth
HockeyPuck writes "The 5-billion-instructions-per second Power6 processor from IBM would beat such rivals as the 3.73 gigahertz Pentium Extreme and the 2.4 gigahertz UltraSparc T2 from Sun. 'It's hard to make the average person understand just how fast this is,' said IBM Chief Technology Officer Bernard Meyerson, offering an example meant to explain his company's baby that still leaves the listener awed with the speediness of the two laggards. 'Hold your index finger out in front of your face,' Meyerson said in a telephone interview from IBM headquarters in New York. 'In less time than it would take a beam of light to travel from your knuckle to your fingertip, the new IBM chip would complete one task and start looking for the next, he said.'"
Nice chip. Now what OS and applications run on it?
I'd guess anything that runs on the Power archicture. Here's a list of the various OSs that have been supported on various iterations of the Power architecture at one time or another.
This guy's the limit!
Fadi Azhari must not be doing his homework...the POWER6 also uses a "different technical trick, called multithreading".
I has Altivec, but IBM calls it VMX
And no, it is not socket compatible to my understanding with the G5 (PowerPC 970).
It DOES run Linux!
...>ahem and other OS's too... like Windows NT...
(mod me down if you must - but I just HAD to...)
Seven Days with Ubuntu Unity
That would only be true if the CPU is able to retire a sustained average of one instruction per clock cycle. SFGate's article makes a raw comparison between chips with different number of cores, threads and other factors, considering only GHz...
Using the 'rule of thumb' (well, forefinger at least) - that light travels one foot in one nanosecond, and assuming he means the Metacarpophalangeal joint and not either of the Interphalangeal joints when he says "knuckle" - and for ease of calculation, assume that my forefinger is very roughly 2.4 inches from Metacarpophalangeal to tip. Then he's saying that the chip completes one "task" and starts another within 2.4/12 nanoseconds - 1/5th of a nanosecond. Five tasks per nanosecond is five billion tasks per second - or 5GHz. The chip is claimed to be a 5GHz chip - so a "task" means whatever the processor can do in one clock cycle. That's an odd definition of the word "task" - it might have been better to say "operation".
Well the big ones are.
Linux
AIX
and i5/OS
Applications?
DB2, Oracle, SAP, and goodness knows how many super advanced and mega expensive packages for specific industries that the average person never knows about.
In other words it isn't wasted on Office, Vista, and other low end applications.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Whether a faster clock speed or more cores will work in for any given application, is, of course, application dependent.
... different strokes for different folks as they say.
If the work you wish to do can be parallelized -- that is, broken into smaller pieces and then either reassembled when all the pieces are complete or, even, better, no assembly required -- and, more importanlty, your application is written to take advantage of parallelization then you will most certainly benefit from a CPU that can handle simultaneous threads.
OTOH, if your tasks can't be parallelized -- one task depends much on the other, than you should focus more on clock speed and less on simultaneous threads.
The bottom line is that the best CPU for you, as always, depends on what you're doing and how you're doing it. And there's usually more than one way to skin a cat, so
My blog
It's horses for courses. If you're serving web pages and running database queries from a well tuned database, the Sun Niagara chips are fast and very well suited. They serve the pages a little slower, but can serve many more at a time.
If, on the other hand what you're doing is not easily threaded then IBM probably have the upper hand. Say you're doing some mathematical analysis, where you have to do everything in sequence. IBM's faster processor can complete each stage quicker, moving on to the next part and delivering the result faster than a chip with more threads but slower speed.
Power6 is a big change from Power4 & Power5 series. The key factor is: it gains clock and SMT at the expense of OoOE. In-order execution means its performance is deeply dependent on perfectly tuned compilers.
Other than the lack of out-of-order, on paper it looks pretty strong. Dual core, lots of bandwidth, up to 7 IPC (5 in one thread, 2 in the other), big GHz, voltage & frequency slewing, and yes it has AltiVec.
p.s. No, it would not be good for Macs. POWER chips are all made for big iron.
Fortunately, Java is a very thread-friendly language. I'm sure I could think of a way to use those 256 processors. ;-)
That's no SNL quote, that's actually Muhammed Ali.
A quick google search turns up a couple references, such as this one: http://www.nzlistener.co.nz/issue/3483/columnists/8092/great_greater_greatest.html;jsessionid=F86BAF04332CA229F91CA1A92B340560
http://en.wikipedia.org/wiki/POWER6 Each core has two integer units, two binary floating-point units, and a decimal floating-point unit, and is capable of two way SMT. The binary floating-point unit incorporates âoemany microarchitectures, logic, circuit, latch and integration techniques to achieve [a] 6-cycle, 13-FO4 pipeline,â according to a company paper.[6] Unlike the servers from IBM's competitors, the POWER6 has hardware support for decimal arithmetic and will include the first decimal floating-point unit integrated in silicon. More than 50 new floating point instructions handle the decimal math and conversions between binary and decimal.[7] This is a feature being added to the processors powering IBM's System z.[8] So it has a 5Ghz clock rate but can actually manage a bit more than 5 Bips peak. But A notable difference from POWER5 is that IBM moved from an out-of-order design to an in-order design, a drastic change which should require software recompilation for top performance. However, the processor still achieves significant performance improvements even with unmodified software, according to the lead engineer on the POWER6 project.[2] Hmmph. I'd bet it's got a really long pipeline to reach that clock speed. The POWER6 has approximately 790 million transistors and 341 mm large fabricated on an 65 nm process. It was released on the 8th June 2007, at speeds of 3.5 GHz, 4.2 GHz and 4.7 GHz[2], but the company has noted prototypes have reached 6 GHz.[3] POWER6 reached first silicon in the middle of 2005[4]. Wow it's huge, almost twice the size of a Core 2 Duo.
I think IBM is doing taking the NetBurst approach - a long pipeline to get to high frequencies. Plus it's a server chip only used in their servers so they can design for a much higher TDP than Intel or AMD and rely on water cooling.
I think this guy is spot on
http://aceshardware.freeforums.org/praising-the-power-6-design-t426.html Later this year Intel will release the 65 nm bulk CMOS Tukwila and
it will likely easily outperform the 65 nm SOI CMOS Power6 on the
benchmarks of most interest to buyers of business critical servers
despite running at less than half its clock frequency and having
less than half its socket level bandwidth. IBM might have created
a better product and closer competitor to Tukwila better if Power6
had been a quad design based on a Power5 core worked over to
improve performance/power but then its wouldn't have the mega-
giga for headlines in the WSJ and given IBM Micro a measure of
bragging rights to help justify its continued existence.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
Um... they are NOT all processing elements
The T2000 (for example) has one die clocked at 1.2 or 1.4 Ghz. On that die are 8 processor cores. Each of these has 4 CMT threads (sort of what Intel used to call hyper-threading). 32 "virtual" cpus, 2U form factor. $6,995 base
The M-series, lets take the M5000 is built by Fujitsu. That has 8 processing elements, each clocked at 2.1 Ghz, with each dual-core. 16 "virtual" cpus. $47,000 base. 10U form factor.
The 5220 has 8 cores, 8 CMT per core, for 64 virtual cpus, 1.4 Ghz. $10,995 base. 2U,
The issue here is that all of this threading doesn't help "straight line" programs. The Power6 does much better (at 5 Ghz)! For "straight line" code, the best machine here (of SUNs offerings) is the M5000 -- machines (more or less) selected randomly from SUNs catalog. Only 16 virtual CPUs, but boy does it cook! (and, its 5 times the size, and 5 times the price).
Just another "Cubible(sic) Joe" 2 17 3061
I do believe that's a (rather poorly executed) reference to Admiral Grace Hopper and her "nanoseconds".
Javascript + Nintendo DSi = DSiCade
That would ~300 feet per instruction on the 3.25MHz ZX81. About one instruction per light-Statue-of-Liberty to use Journalist-friendly units.
I'm not sure whether the comment was intended to be funny or not.
But I think you are correct. Java has synchronisation primitives built into the language and the standard libraries have lots of concurrent save types (e.g. a Vector can be used like a thread safe ArrayList). Compared to other languages where the concurrency is a bolt on (sorry pthreads), this really does make it easy to thread an app and get the synchronisation correct. Sun's tutorial goes into more depth:
http://java.sun.com/docs/books/tutorial/essential/concurrency/locksync.html
-- Mike
Hmmmm, 5 billion instructions per second divided by 186,000 miles per second * 5280 feet per mile * 12 inches per foot = 2.36 inches per instruction at the speed of light.
He should have said "width of your hand" instead. Mod -1 Poorly Chosen Analogy Due To Innumeracy
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
I have found that "Inside the Machine" by Jon Stokes is a great book for explaining the inner workings of processors to non-hardware engineers. Not too heavy, but not all fluff either.
Support the mob or mysteriously disappear.
Financial institutions are required by law to perform financial calculations on a computer as they would on paper, so a chip that can do these calculations natively have a built-in market that is willing to pay the extra for the features. This is a special-purpose processor. There have been and will continue to be purpose-built calculating machines, so it's really not fair to say that IBM is simply trying to dazzle us to justify their existence. As long as no one else makes these machines, their existence is justified.
It's only overclocking if you're running the chip at a clock speed faster than it is rated/designed for. If the chip is designed to be water-cooled, I wouldn't be callinng that overclocking.
I replaced the control PCB in a washing machine about a year ago that used a 68020 CPU. No, that's not a typo. No, I don't need glasses. Yes, really a 68020, like in Mac II.