IBM Optical Chip Zips Huge Files Using Little Power

An anonymous reader wrote to mention that IBM has unveiled a new prototype chip that can transmit data at up to 8 TB/sec, or about 5,000 high-def video streams. While this might not be entirely amazing, the fact that they did it using the same amount of juice required to light a 100-watt lightbulb, is. "The resulting total bi-directional data transfer rate is 300 Gb/s, nearly doubling the performance of a version IBM introduced last year. Compared to current commercial optical modules the transceiver provides 10-fold greater bandwidth in 1/10 the volume while consuming comparable power, IBM said."

Comments

What's with the summary?

[macslas'hole]

A hundred watts, that's all good and well, but what does it have to do with zipping huge files? Or am I reading impaired?

IBM Rocks!

[The_Dougster]

Even though a lowly peon like myself can barely aspire to ever own much real IBM hardware, I have to say they really make some great stuff. Since my P20 monitor finally died, all I have now is an IBM Z50 Workpad, which is a pretty sweet little thing.

I had a RS/6000 briefly, I experimented with running Debian on it. It was some impressive metal, but AIX ran circles around Debian and the graphics was unsupported in Linux. I sold it for more than I payed for it and kept the P20 monitor for free. I ran that monitor for about 5 years.

IBM hardware has always been esoteric, fantastically expensive, and of supreme quality; however, they are just a bit out of touch with regular lusers. For instance, why can't we buy a workstation with a CELL chip even now? We know it could run Linux, easily. Why are we forced to fool around with PS3 consoles when Big Blue could be making the next best thing since the IBM PC?

I'd seriously consider spending $5k for a spiffy IBM cell box running AIX or Linux as long as it could run a PCIe OpenGL card. Heck, I'd take it if it came with OS/2 even!

Re:IBM Rocks!

[BiggerIsBetter]

Software that's custom made for the cell chip barely exploits all the parallel compute power, so I doubt gcc would compile Linux to make use of it (if it can even compile to cell, which I'm not sure of). IOW, you have no idea what you're talking about. :-) There are out-of-the-box distributions of Linux for Cell platforms (Yellow Dog, Fedora, Ubuntu even), and the gcc supplied in IBM's SDK is quite happy to compile for the PPE and SPE cores. Yes, I've played with all this on the PS/3. PCI and Blade hardware is available from Mercury and IBM, but it's pricey... you could drop one of Mercury's Cell PCI cards into a small IBM xSeries...

Anyway, I agree with the OP, this is a killer chip for many many of the applications we use today, and IBM should talk Lenovo (or, oh please, SUN) into selling a Cell-based Linux (or, oh please, Solaris) workstation. That would be ridiculous for software development if it had a Java SDK to go with it.

Re:IBM Rocks!

[Oddster]

Software that's custom made for the cell chip barely exploits all the parallel compute power, so I doubt gcc would compile Linux to make use of it (if it can even compile to cell, which I'm not sure of). IOW, you have no idea what you're talking about. :-) There are out-of-the-box distributions of Linux for Cell platforms (Yellow Dog, Fedora, Ubuntu even), and the gcc supplied in IBM's SDK is quite happy to compile for the PPE and SPE cores. Yes, I've played with all this on the PS/3. PCI and Blade hardware is available from Mercury and IBM, but it's pricey... you could drop one of Mercury's Cell PCI cards into a small IBM xSeries...

Anyway, I agree with the OP, this is a killer chip for many many of the applications we use today, and IBM should talk Lenovo (or, oh please, SUN) into selling a Cell-based Linux (or, oh please, Solaris) workstation. That would be ridiculous for software development if it had a Java SDK to go with it. You're both half-wrong. Yes, you can get Linux running on a Cell just fine. No, software that was not written specifically with the Cell in mind (read: almost everything) uses the co-processors (SPU/SPE whatever) anywhere near capacity. And in fact, almost all of the software that is written for the Cell processor still doesn't use the co-processors anywhere near capacity. It is a very difficult platform to program for, and because of the inherent design of the Cell, it simply performs poorly compared to SMP for a large class of problems. And by difficult, I mean that you have to sit there trying to figure out how to fit your 17k of code and 512k of data into a unified 256k buffer (information theory comes in handy), because going outside the local buffer and using DMA is not only a huge pain in the ass to code up, it is also a huge performance hit. Programming for the Cell is a step backwards from the ideal computer science goal of abstracting the hardware from the code.

I work in the games industry, and I recently saw the performance timer graphs of a very popular racing game that was very recently released for the PS3, a second-generation PS3 title. It was using the co-processors at about 10% of their capacity, and that only came in regular-interval spikes. And this is a piece of software that the Cell was specifically designed to run.

Trust me, you do not want to be programming for the Cell. Stay happy with SMP on the desktop and above, leave the Cell to die in the PS3 as it should. Yes, it's a bit of a technical marvel, but quite frankly, it is not worth all the secondary costs.

How to stops the retarted submissions?

[Anonymous Coward]

Someone who confuses "TB/s" vs "Tb/s" and "zips" vs "transmits" doesn't deserve to be posted.

'Zips huge files'

[malakai]

I can't be the only one that clicked on this expecting some sort of hardware based compression acceleration. I expected some sort of optical take on compression.