Slashdot Mirror
RLX Gets Denser
A reader writes: "There's story about RLX Technologies shrinking their "blades" server on Linuxgram." Knowing how much we pay for our "floor space" at the colo, the notion of having multiple blade machines is pretty cool - and shrinking this to a 1U form factor with 6 blade of the Transmeta Crusoe 5800 line of chip is pretty cool.
Maybe this would be a good market for the metaclustering described a few days ago. Splitting the Linux OS into a user-runable process, and having many "virtual" servers on one physical hardware. For a hosting company, the lower power requirement of the RLX, the lower space, and the lower cost of the Transmeta hardware might make this an attractive option, especially if they are inclined to do metaclustering.
A reason which is blantantly obvious to those of us who use non-x86 processors in our computers is that software originally written for x86 processor takes a bit of work to port to other processors and then the ports aren't always as functional as the original piece of software. RLX boxes run a slightly customized Linux installation but for the most part are using off the shelf GNU software. Were a company to come out with boxes using they'd have to invest alot of money into making sure their software was up to snuff on their hardware. Transmeta and those using Transmeta chips rounded the corners by making their product compatible with a shitload of software already existing.
I'm a loner Dottie, a Rebel.
Yes, I forgot to mention that part. The rep (who was actually like a director of engineering something, I think) who we dealt with was extremely refreshing. He answered all our idiotic questions, listened to our suggestions, worked with our guys. RLX were very nice to deal with. I was sorta sad we never hooked up with them business-wise (our business was just starting to move away from smaller web hosting type customers -- which is who we were eval'ing for in the first place)
The guy there told me that they had fixed a problem with the sleep state recently
Ahhhh, so it wasn't a setting so much as an unintended feature. We had wondered about that. My bet was on setting. :-) I bodged up a couple perl scripts which kept them all alive (and sent back host data while they were at it). Kind of a hack, but it worked.
There were some other issues with power cycling they were working with.
Hmmmm. That we didn't see. Not so good to have some of the blades bounce themselves...
Actually, I really want a set of blades for home. I work with lots of OSes, and one RLX and a decent KVM would make working from home very nice. -B
Ash and Hickory, straight-grained and true, make excellent bludgeons, dandy for the cudgeling of vegetarians.
Oh, stop it. You sound like a gossipping fishwife, piecing incomplete data from speculative press into silly conclusions. "The truth is the truth." Give me a break.
The 5800 is the 5600 at TSMC's 0.13u fab instead of IBM's 0.18u fab. I agree with Hemos that the 5800 is cool. And if the 5800 is cool, it's only because it's identical to the 5600, and the 5500 and 5400 are the exact same design with half the L2 disabled (less L2 => less die area => better yield => cheaper parts). It's all the same chip design. Any claims that the 5800's design is unmanufacturable just doesn't wash with me. IBM manufactured the same parts just fine at 0.18u.
Transmeta's mistakes are myriad, but Transmeta's accomplishments are unbelievable. The press keeps covering them because they have a an incredible technology and it's a David vs Goliath story.
Yeah, the 5x00, at full bore, is probably only 60% as fast (depends very much on the load) as an Intel part at that clock-speed, but Transmeta with its 418 employees beat out AMD with its 14,400 employees in the ultra-light mobile space, and Intel with the bloat and burn of 86,100 employees may be next.
Transmeta has only had product in the channel for just over a year. 60% isn't half bad, and they are learning at an absurd rate. Intel has been at this over 20 years, and with all that experience, even they make some terrific blunders, like RAMBUS and the Itanium.
I am an avid fan of x86 microarchitectures. There are only two out there which I consider interesting, bold, and the kinds of designs which will take us to the next stage of microprocessor development, the P4 and Crusoe. Which is more advanced and has a more promising future for growth? Crusoe, definitely.
The P4 with its unscheduled trace cache is like Crusoe's inbred second cousin. The P4 lacks a sensible place to put its translations, has to reschedule its translations on every execution, has not the brains to apply even a fraction of the advanced CMS optimization techniques, and all that comes at a big cost in transistor count, yield, and heat.
It is clear that virtually every Intel engineer knew the P4 architecture was a big gamble with its much lower decoder bandwidth, puny L1 cache, huge branch penalties, and deeply retarded choice of RAMBUS. When the P4 came out, it was slower than the P3. The Itanium was way slower. That has never been true of a single release, internal or external, at Transmeta.
Even though it is changing radically, no one at Transmeta thinks the next-generation Crusoe architecture is much of a gamble. Viewed from virtually every angle, it's an almost undeniable improvement over the p95 architecture.
The challenge for Transmeta will be finding a manufacturing partner that can keep up with Intel's undeniably strong track record in advancing process technology. Intel is pushing the process and MHz hard, but their microarchitectures are going nowhere fast.