Transmeta notebooks have been available for a week now, and although it's been a fascinating week for benchmark mavens, the casual punter hasn't been much enlightened.
Why? Well three serious IT publications have run benchmarks for Sony's new Crusoe laptop in the past week, published them... and disavowed the results immediately.
The problem is that Transmeta's new chip confounds the traditional one-pass benchmarks that their propeller head authors have spent years perfecting. Since the Transmeta architecture allows for dynamic and smart execution of code in software, the first pass of any test sample is guaranteed to be if not crap, then not quite optimal. With the instruction set and performance heuristics in software - a compromise that allows the chip itself to throttle back its frequency and voltage - judging a Crusoe on first-pass benchmarks is like judging a marrow competition based on the seeds. It isn't going to give you the full picture.
So unsurprisingly, the three benchmarks we've looked at pronounce the Crusoe a dog. To the magazines' credit (or at least most of them), that's qualified by a note that their benchmark shouldn't be taken seriously.
PC Magazine (US) stated: "The PictureBook is a bit anemic [sic], even when set to maximum performance," but then qualifies this by pointing out "the Winstone benchmark test does not repeat tasks, which minimizes opportunities for the Code Morphing technology to have an effect."
The British publication IT Week splashed with a screaming headline 'CRUSOE FAILS TO DELIVER ON PROMISE' using similar benchmarks - but then torpedoed its own assertion with a comment from Microprocessor Report's Peter Glaskowsky pointing out that... um... a 30 per cent speed improvement could be gained from running the benchmark again.
Reviewing the Transmeta picture book, PC World (US) failed to point this out at all, noting that "its score of 66 is shockingly low". Why? They don't even tell us.
Transmeta insists "Our best benchmark is the fact that major notebook manufacturers are seeing the benefits of our chip and shipping systems based on Crusoe."
Now clearly the labs guys have a duty to make like-for-like comparisons with what they receive. These are smart folk (we claim some insight here for all three publications) with a professional duty to provide empirical evidence for their readers. But plainly, sticking to a benchmark written for one architecture is rendered meaningless another - particularly when it's as radical as Crusoe. Both IT Week and PC Mag acknowledged this, in their own way, but still couldn't quite resist the urge to scratch an itch. "These benchmarks are probably worthless," they're saying, "but here they are anyway."
In the week that The Nation published an account of how the New York Times' persecuted of Los Alamos scientist Wen Ho Lee (an innocent man branded a spy) and compounded it with a now notorious non-apology, we were reminded of how the most august publications can continue to propound information they acknowledge is worthless. And that everyone, but themselves, knows to be worthless too. No, Crusoe benchmarks don't morally equate to the Lee case, but you wonder how long the public prints can show the same ambivalence, superiority and essential stupidity as the NYT, without conceding a little tweak to reality.
And how Transmeta thinks benchmarks really should be, we hope to tell you tomorrow. ®