Slashdot Mirror
C'T visits Transmeta
The german technical magazine C'T recently visited Transmeta's office in Santa Clara. Transmeta's roster is impressive, including not only Linus but also Robert Collins and Christian Ludloff both well known for their work on finding undocumented instructions and registers.
Transmeta's LongRun technology (reducing CPU power by varying frequency and voltage) only works with APM, and without it the TM5400's net consumption is 5W at 43 degres Celsius and 600Mhz. At 700Mhz the TM5400's performance is slightly under that of a 500Mhz PIII. The TM5400 will be the first processor to use IBM's new CMOS8S copper process. In the interview David Ditzel denies having used Elbrus technology in Crusoe. For non german speakers, there's always Babelfish.
I'm not so concerned about Elbrusian technology as I am about Elbonian technology.
If you're not part of the solution, you're part of the precipitate.
A static compiler can also do a "fast case" and "worst case" version of code. The advance load and check instructions on the IA64 are built for that sort of thing, and the speculatave load as well. The DEC/Compaq GEM compiler can do an alias and noalias version of a function and do an alias check on entry.
Existing CPUs don't have many features that encurage this, but it can definitly be done. It will bloat the code some, but if the common cases is common enough the extra i-cache traffic will be minimal.
Anyway, I can't wait untill this article gets out. I have a subscription to the Dutch version of C'T (note the ', if we don't use it some lame Dutch "magazine" will think we mean them) and I buy the German one from time to time. The C'T itself is one of the more clued magazines around. It's one of the very few who are doing Linux articles (and more offcourse) which date even back to before the major hype. Personally I think you can best compare the C'T with a Byte but with less advertisement & more clued stories.
Therefore I think its only fair that they get to do such a big (guess it will be big) story about Transmeta. And be assured that if C'T does a story it will be a fair story. Stupid things will be called stupid and vica versa.
If you don't have a subscription I'd really advice you to try an issue. You won't be dissapointed. Unless you can't read German (or Dutch) offcourse :-).
1. the entry level pc is a celeron 500 / k6-2 500 right now, not a P3/K7 700, that's more of a midrance pc.
2. the transmeta cpu is competing on the MOBILE market, not the desktop market, so it's not competing with the latest athlon or sledgehammer.
_______________________________________________
There is no statute of limitation on stupidity.
>At 700Mhz the TM5400's performance is slightly >under that of a 500Mhz PIII This isn't going to be anywhere near good enough. The entry level for a new PC *right now* is a 700 megahurtz processor. By the time Crusoe is being sold, you will be able to get a socket version of the Athlon with on-die level 2 cache which will *totally* smoke this thing, and likely won't be much more expensive. Plus Willamette and SledgeHammer will be just around the corner. The power consumption had better be *sweet*, or else Crusoe will never get off the ground because of its woeful lack of performance.
by Justin Osborn
To the tune of: Big Bad John by Jimmy Dean
The mystic Transmeta announced it that day
It was faster than Intel, as some may say
Conserved batteries, gave heat the slip
And everybody knew it was a superior chip - Crusoe
The folks at Transmeta had slaved day and night
To make a chip that ran just right
It ran on software, instructions optimized.
It wasn't at all advertised - the Crusoe
A guy named Linus, from Finland they employed
He made a living doing what he enjoyed
Hacking the kernel, making it mobile
Spewing out code and letting it compile - for Crusoe
The OEMs quickly made alliances
For laptops, handhelds, and web appliances
It seemed that they would rake in the dough
The folks at Transmeta would get the biggest take though - from Crusoe
I walked to the store to get a laptop
Crusoe was cheap, they were $2G a pop
I imagined all the time from rechargin I'd save
As I read the review the magazine gave - the Crusoe
I pulled out my laptop flying overseas
The project wasn't finished, couldn't be at ease
The hard drive clicked as I typed and typed away
But the battery just didn't seem to drain away - from Crusoe
My presentation was going well but it was going long
My boss began to talk, blather on and on
The Crusoe made it through, and here's my one quip
In the middle of my laptop is one heck of a chip - the Crusoe
Sheepdot: Open Source good, Closed Source baaaaaaad!
In fact, LongRun samples all traditional power management activity, which is:
- AutoHALT (ACPI: "C1")
- x86 HLT instruction;
- Quick Start (ACPI: "C2")
- Internal CPU clocks off (responding to a STPCLK# signal from the south bridge);
- CPU is responsible for maintaining cache coherency;
- Deep Sleep (ACPI: "C3")
- External bus clock removed from the CPU (by the south bridge);
- System is resonsible for maintaining cache coherency.
LongRun effectively samples the time spent in these power management states, and converts it into a fraction of the overall sampling interval. This fraction directly determines the target CPU performance percentage.APM is the older, BIOS-based implementation to exercise the above CPU power management states. ACPI is the new OS-based implementation to exercise the CPU power management states.
Neither APM nor ACPI are necessary for LongRun to function efficiently. In fact, LongRun works perfectly well with all OS's that are smart enough to invoke HLT whenever their scheduler (or GUI event loop) idles. This is true at least for all modern operating systems (all recent Windows versions, Linux and OS/2).
As a result, LongRun transparently works without any BIOS or OS modifications on pretty much all OS's you care about.
Cheers,
Marc
I can't sleep, so I instead of running around th house I decided to translate this article for a change.
Crusoe's islandJourney reports are rather rare in c't. In the last one that I can remember a certain Rob.S.Pierre described a press journey from Apple to Ireland [1] - and whirled up a lot of dust with it. Following an Intel press journey (to IDF) I found myself in Santa Clara, not at Intel's headquarters, but at Transmeta.
Hardly a stone's throw away from Intel's headquarters the startup company Transmeta resides in a beautiful park, in three one- to two-story houses. From the street next to the park with the characteristic name "Freedom Circle" the relatively ugly concrete/glass-towers of the "big brother" are directly visible.
Instead of open-plan offices with lots of cubicles the more than 200 Transmeta employees sit mostly in small offices, although there are also a bunch of big labs. I was assured that I was the first European journalist in their holy halls, and above all the very first to be allowed to spy for a whole day in the offices and labs.
In Linus Torvald's office I felt at home from the start. Torvalds - in proper style in a SuSE-T-shirt - confirmed that he started at Transmeta because he wanted to do something else than always only Linux, Linux. He didn't say it explicitly but indicated that he had been sick of Linux at some point. And so the task of cooperating as a chief architect in the Code Morphing software (CMS) was a welcome change and a big challenge.
But the ghosts one called... at the moment Torvalds is nevertheless working mainly on Linux again, specifically on "mobile Linux" that is intended as an accessory for the Crusoe processor for small diskless systems (in the 32 MB of flash)
By the way, Mobile Linux runs as x86 sofware and is morphed dynamically. There had also been experiments with a native Linux version, but it proved less effective than the emulated x86-version. Transmeta also does not want to publish the the complex native instruction set of the Crusoe processors because the company can keep the option of making arbitrary changes. The CMS versions of the Crusoe processors TM3120 and TM5400 are already different.
InspectionsMarc Fleischmann - a German, by the way, who is also responsible as a development leader for the virtual Northbridge - could convince me of the effects of the power saving technology LongRun in the Transmeta-lab. A TM4500-Crusoe running between 266 and 600 MHz in five steps (the minimal step is 33 MHz) and between 1,1 and 1,6 V consumed 6 W as a peak value, but on average and with usual software it rested at roughly 1 W and got - without radiator - only tepid. LongRun requires idle states for its regulating. Without APM under DOS LongRun does not work and the 600 MHz TM5400 consumes - when running doom for instance - constantly 5 W but still only warms up to 43 degrees. Included in these power consumption values is the Northbridge - compatible in part with Intel's BX - which is integrated into the Crusoe processor or respectively virtualised by the CMS. As a comparison: the original BX Northbridge consumes above 2 W alone - and doesn't know neither PC133 nor even a second DRAM interface for DDR-266 as the TM5400 does.
Since CMS is also controlling the Northbridge it cannot only slow down the processor but also the memory if necessary, which makes the potential for savings even bigger. A small, recently discovered bug in the Northbridge software was quite useful actually to show the advantages of a processor realised mainly in software: In front of my eyes software engineer Peter Anvin loaded a fixed version of CMS to the processor, and that was it.
Responsible for the Crusoe BIOS is Robert Collins. He came together with many other employees from the secret Pentium project that was suddenly abandoned four years ago from Texas Instruments to Transmeta and with his well-known website www.x86.org and Intel Secrets he already fought once with Chipzilla. Collins explained to me that Crusoe first loads the 2 MB (decompressed) CMS in some 100 milliseconds when booting, then installs the translation buffer (8 to 14 MB) and after that, as any proper x86, starts execution of the code at address F000:FFF0. The BIOS itself is coded completely in x86 code and passes system information, e.g. the SDRAM parameters read out of the SPD-EEPROM, through a shared memory range to the CMS software. For the chipset and processor independent part Transmeta licensed the Phoenix BIOS.
InstructionsThe "x86-validator" Christian Ludloff also came to Transmeta vi TI; he was discovered by TI back then by his article about Pentium secrets in c't [2,3]. Ludloff has literally hacked his fingers sore with millions of assembler lines to produce ugly, malicious, !#@#$# code which traps the poor processors over and over. Like that he discovered loads of bugs and anomalies not only in former versions of Crusoe/CMS, but of course also in the concurrents. Many entries e.g. in Intel's Specification Updates should actually say "thanks to CL". Ludloff is planning to make a part of his assembler software, e.g. all the processor structures, exceptions and so on publicly accessible on his website www.sandpile.org popular among insiders - that's gonna be a tidbit!
I confronted Transmeta boss Dave Ditzel with the sometimes uttered suspicion that he took some of the ideas of Boris Babaian during his time as a Sun employee in Moscow. Boris Babaian has been designing for quite a time a processor named E2K at the Russian Sun distributor Elbrus. Babaian was at his's place only a week ago, Ditzel answered. They had a good conversation. He pointed out that the designs of E2K (high end VLIW without dynamic optimisation, very high floating point performance) and Crusoe were very different. A certain inspiration came from there, but not more than from HP, Sun or other designs.
InspirationsDitzel also contradicted publications announcing that products with Crusoe processors shall come out in April. At the earliest at the end of the second quarter, probably in June, the small Crusoe TM3120 will be ready. The TM5400 with more performace and the ingenious LongRun power saving technology is to be expected in the third quarter. It's gonna be the first processor on the market made in IBMs new CMOS8S copper process.
A possible OEM, besides S3/Diamond and FIC, could be Quanta Computer, one of the biggest notebook manufacturers who produces for many others who only print their Dell, HP or other Logo on the notebooks. Sybase as a big software company has already announced that their SQL Anywhere Studio (mobile databases and data synchronisation) is to be extended especially for Crusoe systems. And the company Infomatec from Augsburg announced that they have built a partnership with Crosstainment AG and Transmeta to develop reference models of internet capable consumer devices with crusoe processors together, based on their Java Network Technology (JNT).
Transmeta is stil reluctant to publish absolute performance data, yet the values don't look too bad. A 700 MHz TM5400 reaches about Pentium III 500 level. But when it comes to the so important Ziff-Davis benchmark Winstone it is a bit under that level; the reason is according to Transmeta that this benchmark "touches" a lot of routines only once so that dynamic optimisation can't work. The ZD lab has recognized the deficiency in a discussion with Ditzel and will ensure a more realistic use of the routines in the next version.
And finally there is the most frequently asked question if Transmeta enters the stock market this year. Ditzel's answer: "It wouldn't surprise me..."
and now for something completely different