Witnessing paradigm shift in semiconductor technology

Today we all carry our BlackBerries from meeting to meeting. Soon we’ll be

toting far more powerful devices-ones that will be at home in the netbook’s current market space while also performing some of the tasks now assigned to laptops.

The emerging platforms-including tablets that can be docked to a keyboard and clamshell designs with two screens, such as the MSI prototypes demonstrated at the International Consumer Electronics Show-promise to revolutionize electronic support of collaborative work.

Several questions about these devices come to mind. Who will own the major sockets once tablets enter the mainstream? Whose CPU will be the go-to processor for powering such a device-will it be Intel’s Atom, or will ARM find its cores in use? Which companies and technologies should we be watching?

When I got the iPad, Acer’s Aspire One netbook had already been collecting dust in my desk drawer for more than a year. I had never made good use of the Acer netbook; the screen felt small and clumsy, with its Windows panels and other accessories, and even for casual Web browsing, I preferred my larger laptop.

The iPad manages screen real estate much more carefully. The ability to rotate the screen to find the right orientation for the task at hand is a great feature; combined with Apple’s famous interface to pan and zoom via touchscreen finger gestures, it allows me to browse almost as comfortably as I can when using my 24-inch desktop monitor.

But would the iPad have enough power to enable me to work with business applications, which nowadays are predominantly Web-based? To determine this, I ran some software-based system testing.

I first pointed the iPad to the well-known SunSpider JavaScript benchmark, a test designed to measure a browser’s JavaScript execution performance. The tablet executed the test in 10.2 seconds.

I then ran the same test on the Aspire One, and the result was 20.7 seconds. Disappointing, considering that the Atom N270 used n the Aspire One is still found in plenty of netbooks, and, according to CPUBenchmark.net, the modern Z and N series CPUs for netbooks and mobile Internet devices are pretty much comparable to the N270.

Next, I ran Google’s V8 test suite and got an even bigger difference.

Whereas SunSpider code measures the cumulative time to run the test, the Google V8 test suite computes the inverse of the geometric mean of multiple test components. Essentially, the SunSpider test measures “how slow,” while the Google V8 test measures “how fast.” For the V8 test, the more points, the better. The iPad scored 97.8 points and the Aspire One a mere 21.6.

The intimal tests on the Aspire One were done using the default Microsoft Internet Explorer browser. I subsequently installed Apple’s Safari on the Acer device, and the netbook sped up like a Ferrari: The SunSpider test took a mere 2.8 seconds, and Google V8 scored 446 points.

At the end of the day, the Aspire One is more than four times faster than the iPad, provided you change the default IE browser to Safari or something comparable; otherwise, the Acer netbook is two times slower than the Apple tablet.

Considering both benchmarking results and the browsing experience, the Atom-powered netbook with default IE8 has severe performance limitations when using rich Web applications, while the ARM-powered iPad is quite adequate with Safari. The Atom processor’s speed advantage seems to be lost as soon as you run certain Microsoft software on it (.net and IE).

Indeed, for Web applications the IE engine is substantially slower than virtually anything else on the market. Microsoft has promised to improve the speed in IE9. But the increase might come at the expense of other features, such as security.

I also ported Netlib’s Whetstone and Dhrystone to the iPhone software development kit and to Microsoft Visual Studio. For the Microsoft platform, I compiled them both into .net and native code. Dhrystone is a fixed-point test; Whetstone focuses on floating-point. In modern computing, serious number crunching ought to be done in optimized primitives libraries (vector operations, compression, image and voice coding, BLAS, cryptography, XML parsing, regular expression parsing and the like), so I used single precision for the Whetstone test. The results are shown in the table below.

An interesting benchmark published in IEEE Journal suggests the recent ARM-based Tegra 2 outperforms recent Atoms.

Benchmarking is good, but today two things matter more in terms of performance: the speed of primitives, and the speed of bytecode (potentially, JITcompiled) written in Java, C#, Python, Perl or Flash (increasingly used to drive business logic and tie together primitives in multimedia application and games).

The Apple SDK provides for native, noninterpreted code using a rapid application development (RAD) environment, whereas on most other platforms (Android, Windows, webOS, Google applications), RAD normally involves some sort of bytecode. The two noticeable exceptions are the QT library, currently owned by Nokia, and Apple’s SDK. This is why code compiled in Apple’s RAD tool outperforms code produced by Microsoft’s RAD tool.

Nonetheless, my feeling is that bytecode will still be the choice for the majority of business logic. It will be interesting to compare the performance of the ThumbEE/Jazelle-enabled ARM platform with a recent Atom for Java and .net (for Atom only). I believe that ARM received a major competitive advantage through its years of work on bytecode acceleration (Jazelle) and JIT execution (ThumbEE).

On the primitives side, Intel’s open-source OpenCV and threading building blocks are widely used. Intel has the outstanding Integrated Performance Primitives, Math Kernel Libraryand XML Parsing Accelerator, but the blocks are not part of any popular development platform; each must be purchased separately.