Slashdot Mirror
Crusoe: new benchmarks
C'T has published some new TM5600 benchmarks. Sony's new Vaio notebook uses 10W per hour to power 128 Mb of RAM (112 Mb useable), a 12 Gb Hard-drive, an ATI Rage Mobility Gfx controller and a 9 inch display (resolution: 1024 x 480). This compares with 15-22W for a normal notebook with a bigger display. Intel's Pentium III was usually 50 percent faster at a given frequency, but sometimes virtually no faster and sometimes twice as fast. Code-morphing's impact was measurable: some programs (Quake III and pov run on desk.pov) ran 10-20 percent faster the second time they were run.
wtf does 10 watts per hour mean? Does it really mean that after 5 hours the thing draws 50 watts?
this is not a real 9" display with the 4:3 aspect ratio we are used to. this display is 1/2 the size of a real 9" display at 4:3.
i want to see power specs on a real laptop with a real display, not a palmtop with a 9" x 3" display.
My gaming is the lightest task the notebook does (mame really isn't processor heavy), but compiling code *is* processor intensive. I do it on the plane to try and get things right by the time I get to a customer site.
Both battery life and processor power are important to me, thus I've resorted to carrying extra batteries and plugging in my laptop while at the airport. I'm the luser sitting on the floor in the corner because I found an outlet.
Well, if it helps, I'm not forgiving. I think that TransMeta's technologies are very unimpressive and that their results are disappointing.
Hard drives already have a "power saving mode", which is the reason why if you leave your computer unattended for more than a few minutes, it porks out making you wait while the disk spins up. You can hear it. And of course the machine sits there like a moron while you're waiting for it to do something. This is why most people disable power managment in the computer's bios, to stop this annoyance.
I understand that laptop hard drives are supposedly designed with more efficient motors, which was requested by Intel about two years ago when they wrote some standard paper about how component makers can make their components more energy efficient. While at the same time spitting out CPUs that consume upwards of 35 watts. It must be good to be the king.
These are my friends, See how they glisten. See this one shine, how he smiles in the light.
um - the near future, you're talking Geological time-scales right?
Face it, the 1ghz VAPORWARE announcement was a pathetic attempt to soften the blow they all knew was coming the day before they had to announce their profit warning. Didja see the stocks today? Oh! the humanity!
Face it, if Motorola would just get out of the AIM partnership, PPC could be a great chip again. Right now, IBM has faster chips than Motorola is capable of producing, and CAN'T SELL THEM due to a clause in the contract that forbids, well, actual competition.
These are my friends, See how they glisten. See this one shine, how he smiles in the light.
While I'm initially disappointed; most geeks know that a very large chunk of laptop power goes into the screen, and another into the hard drive. Silly as it may sound, considering the performance hit you take by using this chip over the real-deal (Pentium), and considering the power savings is watered down because of the screen and HD, maybe MP would be the way to go. . .
Of course, I've heard wonderful things about PPC-based laptops, and how very much more efficient they are with batteries. If only Moronola would get off their butts and ramp clock speed. ("twice as fast" argument doesn't wash when clock speed is half as fast).
These are my friends, See how they glisten. See this one shine, how he smiles in the light.
Just wondering, but why were you using beowulf? PVM is good enough to do pvmPov renders. I have done it with straight pvm setups.
-I just work here... how am I supposed to know?
What seems to be forgotten in this discussion, is the fact that there is more than one component in laptops, that consumes large amounts of energy. The CPU and the chipset are certainly a significant contributor to the overall balance, by they are by no means the only one.
You still have to deal with (at least) RAM, graphics card, backlit LCD display, mass storage devices, NIC/modem, etc. Of those, the display is likely the worst offender.
I remember in the late 80's that Atari demoed a laptop that supposedly ran 12h on one set of batteries. Unfortunately, it was never marketed and I can therefore not verify the claim. The interesting bit of information though is that Atari opted for using a monochrome 640x400 _reflective_ display. While nobody could sell that type of hardware some 10+ years laters, it does show where we need to make changes before extended battery live is feasible. Maybe, OLEDs will someday mature to fill this niche.
No, but they could have compared it to Intel version of the picture book.
The only time the display was really just inconvenently too small was when a dialog "knew" it would have enough height, but my display was too short. Bloody pain. Happened failry offen in Windows, but pretty much never in BSD. The X apps mostly had scroll bars, or did something else when denied the height. When not I could at least drag the window up and down with ease, Windows made it really hard to do that.
It is a pity Sony didn't support BSD (or Linux) on that system, it was much more usable.
What machines would these be? While the StrongARM uses less power then the Crosue, most of that 10W isn't the CPU. I don't think we will get a 1W laptop until hard drives are replaced by something that sucks less power, DRAM gets replaced by MRAM, and most importantly we can make the big power sucking LCD backlight go away, or at least make it much smaller (like an eyeglasses backlight....).
Until then our power sippers will be palm pilot like displays with no backlight most of the time, very little RAM, and no hard drive...
Oh, I didn't mean to pick on the SA because it sucks a watt. I actually thought it sucked more like half that. I was trying to say that replacing the crosue with another CPU in a 10W box will only give you at best a 8W box (assuming the Crosue sucks 2W, and the new CPU zero).
So to suck dramitically less we either have to go the Palm Pilot route and drop the hard drive, most of the ram, and lots of other stuff...or find a way to get all that other stuff to sip power rather then gulp it.
They tend to run slower as well, I picked the StrongARM because it was in the same ballpark (even if it's integer performance is likely to be halfish the Crouse, and the FP will be abysmal because it has no FPU). The Xscale would have been better, but I didn't think of it at the time.
Yep, as long as you don't need x86 compatability it is better then the Crosue. Then again something almost allways beats the x86 if you don't need x86 compatability (well it has price/performace going for it in some price and performace bands, but I'm wondering offtopic...).
Hard drives already have a "power saving mode", which is the reason why if you leave your computer unattended for more than a few minutes, it porks out making you wait while the disk spins up. [...]
This arrangement is decidedly sub-optimal for certain sparse workloads, e.g., mine, which I suspect is why Xeger said:
Why not produce a hard drive that "idles" at low RPM [...]
It sounds like he was thinking the same thing I am, namely that instead of stopping when idle and spinning all the way up to 4800 RPM (or whatever it is) when needed, it should idle at an extremely low speed, like 120 RPM (no, I didn't forget a zero), so it can handle occasional small tasks slower than normal but without needing to spin up at all.
Here's the thinking: my typical workload when I'm mobile tends to generate a single I/O once every 70 seconds or so. As a result, my HD spins unnecessariy for about a minute, spins down, and, about five seconds later, forces the application to freeze for two or three seconds while it spins back up to answer a single I/O request, and then continues spinning unnecessarily. Rinse and repeat until battery is drained. Meanwhile suffer a three-second latency for common actions.
Okay, it's not quite that pathological, but I'm not exaggerating by much. I'm mostly reading and editing, which involves hardly any I/O at all once my files are open. So my HD should be idle all the time, right? Well, sorta. The thing is, every now and then I click a control, switch windows, or save a buffer, which generates a small but nonzero amount of activity.
This is where the extremely-slow-but-not-stopped mode would come in: 120 RPM = 1/2 second per rotation = 1/4 second per half rotation, so the mean latency would be a quarter of a second. Pretty horrible compared to the typical 8 milliseconds at normal speed, right? Well, yeah, but since it takes two to three seconds to reach normal speed, this would actually be much better. Of course, you'd still want it to spin all the way up when you do something big, like a compile or launching or quitting an application. Also, maybe you'd want to double the speed to a whopping 240, or even 300 RPM to keep the max latency under the quarter-second annoyance threshold, but I think that with some work, this Could be a really good idea.
David Gould
David Gould
main(i){putchar(340056100>>(i-1)*5&31|!!(i<6)<< 6)&&main(++i);}
Thanks. I suspected as much, but I wasn't sure, and I figured I'd already rambled enough... Any idea what minimum speed is required by the air cushion? Maybe there could still be an intermediate state between off and on, to better handle the usage characteristics I described, even if it couldn't be that slow. Or is the whole idea just ridiculous? Oh, well.
David Gould
David Gould
main(i){putchar(340056100>>(i-1)*5&31|!!(i<6)<< 6)&&main(++i);}
Don't get me wrong, I love my dual-monitor 3200x1200 setup at home, but 500K pixels is easily enough to get useful work done. (Besides, standard laptops aren't much better at 1024x768. Only 50% more pixels.)
There is a large and growing market for subnotebooks, and it's obvious that's the part of the laptop market where a power-frugal processor would be most valuable. Most of the other Crusoe-powered laptops coming out have similar displays (Loox is 1024x480, TP240 is 800x600). It makes perfect sense to evaluate the processor in the context of the systems in which it'll actually be used, rather than big desktop-replacement laptops where nobody cares about power consumption...
Alan
The reason that a benchmark runs faster the second time is that the Code Morphing software doesn't need to retranslate, and with these short-lived benchmarks, the translation time is a significant amount of the timing. Rebooting won't necessarily result in a faster translation, as the Code Morphing software supposedly re-morphs sections of code more aggressively over time anyway if they get called often. Basically, if you rebooted your kernel, you might reboot more quickly, but the steady-state performance of the system would be identical after a minute or so.
This is the reason standard benchmarking is unreliable on a Transmeta part. Basically, the benchmark runs end to end touching many features of an application, but not really reusing many of them, so you get charged the startup and initial Code Morphing overhead on a large body of code and you don't get to see the actual steady-state performance of the device. In contrast, if a user's sitting there using Word for an hour, they'll spend 99% of there time at steady state using just a few features the bulk of the time.
So, no, you don't need to reboot to get a faster kernel on a Transmeta device, unless you just want to watch it boot 2 seconds faster the second time.
--Joe--
Program Intellivision!
Ack, you people don't get it! Here's the short, simple explanation:
When the program is run the first time, you see "Code Morphing Time + Program Running Time." To the user, this manifests itself as "Total Running Time." The second time you run it you mostly only see "Program Running Time" (and some "Code Morphing Time", but not nearly as much), and so "Total Running Time" looks somewhat smaller. The reality is that "Program Running Time" didn't really change much, if at all.
In a real world scenario (not a benchmark), "Program Running Time" is the important figure, as you typically end up using the program for quite a long period of time and so the "Code Morphing Time" ends up being in the noise, rather than being one of the dominant terms as it is in some of these benchmarks.
--Joe--
Program Intellivision!
DISCLAIMER: I am not a kernel hacker, so I might have some factual errors in the text below. Kernel hackers: Feel free to correct me.
It does bad things if the clock rate varies, as this affects micro-delay loops that are used when talking to certain (broken) peripherals. The execution speed of the instructions varies even on true Intel parts. The kernel has two mechanisms to cope with this, and the important one should work fine on Transmeta.
(Reference: arch/i386/lib/delay.c in the kernel source.)
The older mechanism is the BogoMIPS busy loop. This mechanism relies on a tight loop that fits in cache and should run with fixed behavior on a given device. This mechanism probably doesn't work real well on a Transmeta part, though I suspect Code Morphing would hit steady state real soon and so the BogoMIPS loop wouldn't be hurt too badly. Still, it's suboptimal. That leads me to the second mechanism.
The newer mechanism which is available on most modern CPUs is the Time Stamp Counter, which returns a cycle count rendered in terms of CPU clock cycles. As long as you know the MHz rate of the CPU, you can measure time very accurately. Presumably, despite the Code Morphing layer, the Transmeta CPU will return a meaningful, coherent clock count for this instruction.
The problem with varying clock rates is that the time-base for the BogoMIPS or TSC clock change and the kernel isn't notified. In theory, the Transmeta could actually just use a fixed-rate counter for the TSC whose time-base didn't vary as the CPU's clock-rate varied, thus fixing the problem entirely. But then, that'd make too much sense. ;-)
As for HLT, I thought Linux did that already? That's how come my CPU stays nice and ice cold when I'm not running my Distributed Net client. A quick look at arch/i386/kernel/process.c shows the uniprocessor idle loop calling __asm__("hlt"); as long as the CPU supports it.
--Joe--
Program Intellivision!
I apparently don't understand the idea behind Code Morphing. I thought it was that there was a cache for the translated x86 instruction stream, and that when those instructions were run again, the processor says "hey, I have those in the cache already" and so doesn't have to translate them again. However, this would make the second time through the loop faster. I can't figure out how any technology on the processor would make a program faster after it was quit and restarted. Can anyone clear this up?
Mod down posts with a "Free Mac Mini/iPod" sig, they're spam!
Crusoe: Not the fastest one, but economically
[ 11,10,2000 17:13 ]
For some days the c't laboratory measures the efficiency of the TM5600-Prozessor von Transmeta[1 ] . After the first results to the Speicher-Performance[2 ] now further results of bench mark are certain.
The Crusoe is 12 GByte fixed disk, ATI in the Sony Notebook Vaio PCG-C1VE[3 ] with 128 MByte primary storages, rises up Mobility and a 9-Zoll-Display with 1024 x 480 points dissolution. The processor runs alternatively with 300 mc/s with 1.2 V of core voltage or 600 mc/s with 1.6 V and can be switched during operation between both frequencies. It does not have 128 KByte Level-1 and 256 KByte Level-2-Cache. x86-Code can it execute directly, but translates it beforehand into its internal VLIW instruction set (very long INSTRUCTION word). In order not to repeat this process continuously, the Crusoe stores the translated code in a code Morphing memory. In addition it zwackt itself 16 MByte from the primary storage, so that for the operating system and applications only 112 MByte remain remaining.
In the case of 300 mc/s the c't Akkubenchmark results in a run time of approximately two hours. Sony indicates the Akku capacity as approximately 20 Wh. Therefore the Notebook takes up altogether only about 10 Watts of performance - quite considerably, most Notebooks between 15 and 22 Watts goennen itself nevertheless. In the efficiency comparison the Crusoe remains certainly behind one fast clocked mobile Pentium III clearly:
Processor
Clock
[ mc/s ] BAPCo
SYSMark 2000 PovRay 3.1
chess2.pov 3DMark 2000
CCU Marks UT
[ fps ] Cinema
4d
Crusoe 300 31 124 PPS 33 8,4 1,8
Crusoe 600 50 257 PPS 56 11,8 3,7
Pentium III 500 86 347 PPS 78 14,9 5,5
Pentium III 600 92 417 PPS 81 15,4 6,6
Comparative measurements on Acer TravelMate 522 TXV with Pentium III-600 (with speed steps), 128 MByte primary storages and likewise the ATI rise up Mobility.
Some bench mark we let run several times consecutively, in order to observe the influence code of the Morphing (translate of x86-Maschinencode into Crusoe instruction). In the theory a bench mark should run with the second time faster, since the processor can fall back to the Morphing memory and again not translate the code must. In practice this effect actually shows up with some bench mark: Thus the Frame rate of Quake III of 13,5 rose fps by 10 per cent to 14,9 fps in the second run. PovRay calculated " desk.pov " in the first passage in 20 seconds and needed with the repetitions only 16 seconds. (both measured with 300 mc/s.)
However the results remained by the 3DMark 2000, unreal Tournament or the " chess2.pov"-Berechnung von PovRay constantly and also most individual values of the BAPCo Suite varied only around the two per cent usual with all systems. This bench mark execute obviously most program sections anyway already several times, so that the rate advantage enters with the repeated passing through of a code paragraph bench mark result also. For example the BAPCo Einzeltest " Elastic Reality " consists mainly of calculating 150 frames. According to the first picture the code should be situated completely in the Morphing memory, so that the Crusoe can calculate the further 149 pictures with max. rate. Code the Morphing would have to go already extremely slowly, in order to measure an influence here.
Further results follow in the c't output 22/00 (starting from 23 October in the trade). ( jow[4 ] / c't)
********* sig: If you don't like the law, get filthy stinking rich, and buy a better one.
No... though you could say 'if on for one hour, the laptop will consume 10 watt-hours'.. but that's the same as saying it's a 10W device.
The amount of power used is measured in kilowatt-hours... meaing the amount of energy equivalent to drawing a kilowatt for an hour.
The best way to say it is not to say it at all.
Saying 'this laptop sucks 10 watt-hours per hour' would be correct, but can be reduced to 'it's a 10 watt laptop'.
10wh/h = 10w...
An amp-hour is meaningless unless you know the voltage. That's why batteries are usually measured in amp-hours.. the voltage is known and fixed.
(there is no need to specify watts as you already know the voltage, and watts = volts * amps). It's saying the same thing.
Yes. How does that make what I said wrong?
A '10 watt-hour' is a '10 joule/second' load for an hour.
The hard drive doesn't suck as much power as you'd think, under the right conditions. If you provide an extremely generous disk cache with tons of read-ahead, and spin down your drive after a minute of disuse, it won't be such a big problem. Why not produce a hard drive that "idles" at low RPM and then kicks into high gear when it receives an I/O request. A variable-speed hard drive would be significantly harder, but probably still doable.
Lithium polymer batteries have started hitting the market--the iPaq uses them, for one, and my new cell phone does as well. With their significantly higher energy density and the Crusoe's power-saving, we'll be seeing laptops with a running time of 6-8 hours--or to put it another way, laptops with a running time of 4 hours that have virtually no battery. This still isn't enough, and definitely isn't worth the 50% performance hit reported on some applications.
In my experience, the biggest power drains on a mobile system are (of course) the display and the CD-ROM/DVD drive. I'm waiting for a new display technology (light-emitting polymer, for example) that will make more difference than the Crusoe or lithium polymer combined.
So, the chip is much slower than the p3, and extrapolating from the provided numbers, if you fit it with a real screen (not the ridiculous PDA size this benchmark has been run with) it sucks almost as much power as the aforementioned vanilla p3 notebook, and probably more or less the same as a p3 notebook at half of the Crusoe's clock speed (since that's a comparable speed given its performance).
/. community is much more forgiving.
Can anybody explain me what's the point of all the hoopla that has been going on about this ? If Intel or AMD created a processor like this they would be fried and grilled here, but since it's Linus' employer I have the feeling that the
Don't get me wrong, from a company that has never produced CPUs the Crusoe is an excellent first product, but I fail to see why this should be hailed as the second coming or something.
Am I being too cynical ?
-- the cake is a lie
How do you use 10W per hour? Considering that is a rating of power (energy per unit time), not energy.
I could care less about how fast the clock on the computer is -- it matters how much it can do. It's like that 'hard working' incompetant employee that every office seems to have -- they'll be there for 12hrs/day, and still not get shit done.
If a 400 Mhz machine can do something in 4 cycles that a 600Mhz machine takes 8 cycles to do, then technically, the 400Mhz machine is the 'faster' machine for that process. You start seeing that behavior on a few dozen/hundred instructions, and it makes real sense to get the lower clock speed chip.
Build it, and they will come^Hplain.
Umm, all the tests conducted ARE real apps. Thankfully, we seem to have evolved past the Winstone era. Unreal Tournament is a real FPS measure, Cinema4D and Povray test actual rendering times, and Bapco Sysmark is a script of around a dozen common Windows applications (Photoshop, Office, Netscape, etc.) So all these results ARE valid, and the loop thing really wouldn't change anything. (The score reported by Bapco on the Photoshop segment is going to be damn close to actual photoshop performance, because it IS actual photoshop perfromance.) The performance is still pretty good though, just not as groundbreaking as Transmeta would have you believe.
A deep unwavering belief is a sure sign you're missing something...
-p.
On the other hand, the article failed to mention the size of the screen on the P!!! laptop, along with what effect the 9" screen had on the Crusoe's power consumption. Considering the power the screen consumes, and how small 9" is compared with the size of a normal notebook screen, that could be very relevant.
-----------------------
Nicotine free Amish .sig.
Let's take a look at those numbers
50% faster
10 W vs 15-22 W
Since 15W = 50% greater than 10W, the Crusoe
technology has a really meager effect.
It seems the speed penalty is too great for the power savings. 50% less speed, and 50% less power consumption !!
We all see how many watts those processors are supposed to consume. But I can't understand why nobody has ever given us real numbers. Let me make myself more clear. The question for the average user, is HOW many hours your notebook can be actively used (meaning that you actually do something that might be battery consuming, unlike manufactorers' usual numbers which tend to be 3 times more). I'd really applaude someone finally explaining what good is any power saving schema, and if it can do something more than a 10% better amount of time online. Just my 2 cents here
---- Fear the mighty TsEA
You don't and it doesn't. The Crusoe is an advanced microprocessor that has a software layer which provides the x86 instruction set. That's the whole point.
Now hiring experienced client- & server-side developers
-- @rjamestaylor on Ello
It seems that now other components of this notebook are guilty of the most power usage, so the focus should turn to other high-energy devices. What are these? My first couple guesses are the display screen and the hard drive. Here's a good application for those goggles that have a videoscreen built into them (so that a regular-sized monitor appears before your eyes). Those things can't use very much energy, do they? So use that for a display. I'd be interested to see how long the battery lasts then. Carl hi mom
Vote Libertarian
Thats all well and good that your bunch of 486s and p100's beat a k6 and celeron, but how much space did these boxes use up? How much energy do they consume? What is the network throughput? I hope you are running it switched otherwise hubs would suck.
Only the State obtains its revenue by coercion. - Murray Rothbard
I seem to recall something about Linux developing a kernel tuned specifically for the Crusoe processors. It runs native, if I remember correctly. But if you were running a non-native instruction set instance of the kernel, it would run faster the second time around. Frequently used instructions have their translations cached and are optimized. So it would run faster the second time around. ;)
Somehow, testing this processor in a system with a tiny display doesn't seem like a very good way to compare it to a realistic real world notebook. (Sure, some people may buy this, but the display seems too small for many real world applications.)
The NY Times today ran an article about Intel's PR counter offense (essentially), laying assault on Transmeta as being erroneously knighted the low-power mobile-CPU provider. Intel claims: Anyway, with direct regard to your last point on CPU battery draw, The biggest power consumer is the LCD display
--
Me pican las bolas, man!
Thanks
--
Me pican las bolas, man!
Thanks
Jaco
Would the fact that it is faster the second time around have a positive effect on recursive programming, AI, lisp, etc...?
OK, we're dealing with laptops and webpads.
Based on batteries coming out of B.C. right now, I'd say Transmeta can cut the power on a system with color video screens and CD to about 80-90% of current usage. If you beef up the RAM to the gills, probably 70-85% power usage.
Main drain is monitor power for most people - fastest method to cut this is better screen technology. Time to market of useable low-power high-res screen is probably 2002-2003 product cycle. This will still leave you at 50-60%.
Not a lot of hope on the CD power usage.
On the other hand, a non-CD Ethernet laptop, with honking big RAM and improved hard drive could probably cut power consumption down to 25-30% of current usage. This is with a total redesign. Expect to see these babies in late 2002. Price mark will be high until early 2004.
[Note - I am expecting to put in an indication of interest for TMTA IPO shares - I am biased]
--- Will in Seattle - What are you doing to fight the War?
Half the display size (vertically), half the processor speed, half the power consumption... I think they should ship OS/2 with that thing, just for the bad pun.
Stop asking: My Karma is nowhere near -100 and will never get there with a default score of -1. I lost.