A Transmeta Couplet

Godfather writes: "According to the heise-people the 600-MHz-TM5600 performs somewhere in between a Pentium III-400 and a Pentium III-600. It seems to be amazingly fast in memory access. The article is in German so you have to try the fish." A better translation would be appreciated, too, since Babelfish still leaves certain things murky. And if you've heard enough about the upcoming Picturebook, Timothy Brown writes: "Fujitsu is releasing (in early November) the Loox-T laptop, with a 500Mhz Crusoe chip. It's only available in Japan, but Dynamism, a company which sells Japan stuff to purchasers in the U.S., is accepting preorders." Here's that link.

Comments

Human Translation of the german article.

[Ashran]

Well, my english is not perfect, so please bear with me =).
I hope this translation is better than what the fish blurps.

<Translation>
The TM5600-Processor of Transmeta (used in the Sony Vaio) has to prove itself in the c't-Test-Facality right now.
First Benchmark results show that the performance of 600-MHz-TM5600 is between the
performance of a PIII 400 (while using the Appleman-testprogram, with about 15 Million iterations / sec)
and a PIII 600.

The Memory performance is pretty good for such a small Notebook,
it starts with 60 MByte/s (MemCopy) , and goes up with Cache-Hits in the Translation buffer (to 170 MByte/s).
Writing to the Memory (using Memset) is at about 280 MByte/s.
To compare:
Pentium III 500 (Via-Apollo-II-Chipsatz) writes with 70 Mbyte/s and 150 Mbyte/s.
Pentium-III-Coppermine-800 (Solano-i815-Board) is about as good, with 190 MByte/s and 255 MByte/s.

An interesting thing is, that the Crusoe-Processor responds to the CPUID command. (Cannot be disabled in the Vaio.
But the Serial-Number seems to be calculated by the Code-Morphing-Software.
The "real" CPUID command, (compatible to the AMD-Athlon) returns no Serial-Number back.

There is no correct working CMPXCH8 command in the official specs,
because Windows NT depends on the wierd behavior of a non correct CMPXCH8 if it detects a Pentium-Family CPU (Family -ID 5)
(my note: AFAIK Intel first implemented a non RCF CMPXCH8 command)
</Translation>

Full translation

[pnambic]

Crusoe in the c't labs

Transmeta's TM5600 processor (built into a Sony Vaio) is currently undergoing c't labs' scrutiny. The first benchmark results put it into the range between a Pentium III-400 (e.g. the c't Mandelbrot fractal at about 15 million iterations per second) and a Pentium III-600. Memory performance is impressive for a small notebook, starting with about 60MB/s for MemCopy and increasing to about 170MB/s for cache hits within the processor's translation buffer. Write performance (measured via Memset) peaks at about 280MB/s. In comparison, a Pentium III-500 with Via-Apollo-II chipset won't reach more than 70MB/s and 150MB/s, respectively. A Pentium III-Coppermine-800 on a Solano i815 board comes in close at 190 and 255MB/s.

Another interesting fact is that the Crusoe processor supplies a serial number via CPUID, which cannot be disabled, at least in the Vaio configuration. This serial number seems however to be generated exclusively by the code-morphing software. The "true" CPUID command (compatible with the AMD Athlon) does not yield a serial number. Additionally, Transmeta declared the otherwise flawlessly functional CMPXCHG8 command as not available in the list of official features, since Windows NT can't cope with a Pentium-class processor (family ID 5) supplying this feature.

Further benchmarks and power consumption measures will be published by c't during the next week.