Apple iPad 2 As Fast As the Cray-2 Supercomputer

An anonymous reader writes "Presenting at the IEEE High Performance Extreme Computing conference, a researcher from the University of Tennessee presented evidence that the iPad 2 is as fast as the original Cray-2 supercomputer. Performance improvements were made to the iPad 2 LINPACK software by writing Python for generating and testing various Assembly routines. The researcher also found that the ARM Cortex-A9 easily beats the NVIDIA/AMD GPUs and latest Intel/AMD workstation CPUs in performance-per-Watt efficiency."

My wristwatch

[aglider]

Is more powerful than the Atanasoff machine [wikipedia.org]!

Researcher = Jack Dongarra

[gnasher719]

who has been publishing the Top 500 Supercomputer list for many, many years. I would bet that he ran Linpack himself on the Cray-2.

Re:OMFG

[PaulUTK]

I'm not sure you actually "read one more BS Apple story", this has nothing to do with how great Apple is. This was presented by Dr. Luszczek here in my research group at the Innovative Computing Laboratory to show the efficiency of ARM vs server class CPUs and GPUs. The only readily accessible ARM we could develop on at the time was the iPad2. As with most journalism, the main point of the presentation wasn't what the title of the story was.

Re:I knew the Cray-2

[ichthus]

27 or so years its senior. [wikipedia.org] Wow, pretty neat, huh? Also, my Galaxy S2 is waaaaay faster than my Atari 800.

Re:OMFG

[adonoman]

Except that at least one of the people interviewed had the current 4S, and was still blown away by the weight, look, and performance of the identical phone handed to him. These weren't people unfamiliar with iPhones.

Re:I knew the Cray-2

[gander666]

When I started my Master's thesis, I began learning to program the Cray-XMP. In Fortran still, with some C (pre-ANSI C for you whippersnappers). Then I got a job, and that opportunity fell by the wayside. I still am in awe with how those machines were optimized.

Of course today, I would just use Matlab, and if I needed more speed, I would compile it to C++ and run natively. But it has been a long time since I have done any serious number crunching.

For a good read, pick up "Turing's Cathedral", it is a good story of the birth of electronic digital computers, and an eye-opener.

remember the i486? whips the Cray-1

[swschrad]

you need to remember, however, that the software for these consumer devices is nowhere close to that on the Crays. no optimization is done any more... for you script kiddies, "optimization" means you manually with the assembly language, or automatically in the compiler, try several things and pick the one that uses the least memory/processor cycles/OSPF if multithreaded/whatever based on what you want to gain by optimizing code. all this "include.kitchensink" stuff just packs in extra code crap in case any of it is needed.

and Clippy or Bob never ran on a Cray, either.

Re:OMFG

[AcidPenguin9873]

Slide 18 from this slide deck [utk.edu] is where you compare energy efficiency across processors. I see two major flaws in your methodology:

1. You're using the TDP of each of the processors, instead of a measured power draw while running the benchmark. Are those other processors drawing their TDPs while running this benchmark? I doubt it. Usually the TDPs for any given processor are listed for some sort of power virus type test which is difficult if not impossible to hit running real code. It's possible that this benchmark hits the TDP of each of these processors, but I'd want proof of that, and generally I'd want measured power draws, not TDPs.
2. More importantly, dynamic power scales quadratically with Voltage (P=C*V^2*F) (Wikipedia reference [wikipedia.org]). If you run these processors at a slower clockspeed and lower voltage, their power draw drops by the V^2*F factor. The performance slows down because of the lower frequency, sure, but you get a squared factor by decreasing voltage, plus some power reduction due to lower frequency, while only having a linear slowdown factor due to the lower frequency. In other words, they can get into a much more efficient power band by not running at their highest voltage/highest frequency. They can run up at high voltage/high frequency because users want super-responsive computers and super-fast GPUs, but for doing long-running power efficiency comparisons, you'd never run them that way. You'd find the sweet spot on the V/F curve and run them there. Cortex-A9 is designed to live at a different point on the perf/power/V/F curve - it's effectively already down at a lower frequency/lower power/lower peak performance point, yet at its performance point it is very efficient. You'd need to sweep across a range of freq/voltages to find the sweet spot of each processor before you compare them like this.

Re:My desktop computer is way more powerful than t

[JohnSearle]

And in other news, the Asus Transformer Prime is 4x as fast as the Cray. Android (NVIDIA Tegra 3 T30 1300 MHz (4 cores) ) [primatelabs.com] vs Apple (Apple A5 (32nm) 1000 MHz (2 cores) ) [primatelabs.com]

I hate how everything must be compared against Apple iProducts. I don't recall every comparisons of yesteryear being brand specific. I don't care if the iPhoneX is 2x as fast as iPhoneX-1, or the iProductY is 2x as fast as the Cray. Give me damn benchmarks or clock speed of current day standards, and not a commercial.

Re:My wristwatch

[Tuidjy]

Yeah, and my bicycle smokes any Ferrari... in miles per calorie efficiency.

The iPad2 is the second coming of Christ, we got it already.

Re:My wristwatch

[atlasdropperofworlds]

According to wikipedia, the GPU in the XBox 360 puts out 240 GFLOPS [wikipedia.org]. The CPU is harder to nail down, but it seems to have a peak around 115 [xbox-scene.com] GFLOPS.

The iPad 3 has a CPU that, from what I hear, has a peak capacity of 1.5 GFLOPS. The SGX 543MP2 in the new iPad 3 has 4 cores and does 6.4 GFLOPS per core, per 200 MHz [wikipedia.org]. If we assume the 4 cores are clocks at 600 Mhz, that would mean the GPU output would be, in theory, 77.6 GFLOPS.

In short, whatever Carmack was thinking or testing, he sure wasn't hitting the peak performance of the Xbox - the console is still leagues ahead of the mobile CPUs and GPUs, and it's 7 years old.