September 20, 2006 9:00 PM PDT

AMD chip socket makes room for others

Advanced Micro Devices will let computer makers pop specialized coprocessors into sockets that otherwise would house an AMD primary processor such as Opteron or Athlon.

It's an expansion of the Torrenza initiative, introduced in May, which provides a way for others to connect their technology directly to AMD chips, via the company's HyperTransport interface. The program initially let companies plug in their coprocessors via an external connection called "HTX." Now it is licensing the processor socket design as well, said Marty Seyer, senior vice president for AMD's commercial business.

That means a computer maker doesn't have to develop a separate design if it wants to use a coprocessor--a special-purpose chip for handling tasks such as graphics, mathematical calculations or security. "Now an OEM (original equipment manufacturer)--say, IBM--only has to develop one infrastructure," Seyer said.

Manufacturers that have licensed the socket technology include IBM, Sun Microsystems, Cray and Fujitsu-Siemens, a Fujitsu subsidiary, Seyer said. Among other things, the technology includes details about how different processors keep track of what data is stored in each others' cache memories, through a feature called "coherent HyperTransport."

"We are working with five to 10 coprocessor partners that are at various stages" in development, Seyer said.

AMD is pushing HyperTransport as one of its competitive advantages against rival Intel, and it has made significant market share gains. But Intel processors once again are competitive, and Intel is working on a HyperTransport rival called the "Common System Interface." It's not yet clear whether Intel will let others directly connect chips to its own via CSI.

IBM wouldn't comment on its specific plans, but it has shown a general interest in coprocessors in the design of the "Roadrunner" supercomputer at Los Alamos National Laboratory, said Bernie Meyerson, the chief technologist in IBM's Systems and Technology Group. Roadrunner uses more than 16,000 Cell processors to accelerate calculations running on more than 16,000 AMD Opteron chips.

IBM welcomes the chipmaker's idea. "You can take plain vanilla hardware that has the open socket and turn it into an appliance where it may have a hundredfold increase in a particular capability," Meyerson said.

AMD expects the first socket-based coprocessors to arrive in 2007, spokesman Phil Hughes said. He declined to reveal payment terms for the licensing.

Although the coprocessors are expected to first arrive in servers--for example, in those used for high-performance technical computing--the idea also will be useful for personal computers, Seyer said.

For example, graphics processors that today communicate over the relatively slow PCI Express conduit could be plugged directly into a socket instead, Seyer said. Other possibilities include physics engines that could help video games, or security processors that could deal with virus attacks.

See more CNET content tagged:
coprocessor, HyperTransport, socket, AMD Opteron, AMD

12 comments

Join the conversation!
Add your comment
I might be wrong....
I'm not all so super on my computer history (that was a 9am monday morning @uni), but wasn't this the kind of thing we used have right back when? A chip for the normal stuff, a chip for maths, a chip for this and a chip for that? Then we integrated them. Now we're going the other way? Let me know if thats stick, wrong-end-of!

Don't get me wrong. Sounds super. No anti-trust cases for you, AMD! I like the idea. For example, we can start bumping some of the graphics processing onto one of these to reduce the size of the gargantuan graphics cards we're getting lately! Or, perhaps, SLI without an actual second card? There appears to be some exciting applications of this, including the bog-standard ability to bung another number cruncher in there!!

Nice one AMD!
Posted by djcaseley (85 comments )
Reply Link Flag
Wow I'm having Amiga Flashbacks!
Seperate processor, seperate Audio Chip, Not just one but 3 Graphics chips!
Fat Agnus, Denise, Paula (A2000)

:)
Posted by vission (1 comment )
Link Flag
7950gx2
I have a 7950gx2 card and while it's 2 cards in one, it accomplishes SLI off of one PCIe slot instead of two and is SLI ready so if I was sick and crazy enough, I could slap another 7950gx2 in.

The Physix Engine would be awesome for this kind of application and I was looking at purchasing one of the Physix cards once more games become available to use it's technology.

I do remember the older PCs using math co-processors... and back when I was into the whole cyberpunk thing, I actually took 2 math co-procs and snapped them to the side of some boots. Kind of retarded in hindsight, but they no longer served any use.
Posted by SeizeCTRL (1333 comments )
Link Flag
Really a good idea
Well, i think its a good idea. Think about as faster will be the video cards :)
Posted by jorgevilela (1 comment )
Reply Link Flag
bad idea - dongles for drm/copy protection
please install you sony/ms drm decode on your chip

oh yeah, you'll have to swap them out when you want to view something from ms vs sony

f* that
Posted by baswwe (299 comments )
Reply Link Flag
If they were going to do that...
They would most likely use some sort of a key that you plug into a USB port. It would not be proctical for them to sell something that would require the adverage user to open their case so use a piece of software.
Posted by startiger (50 comments )
Link Flag
Thats crazy
Why would need to develope and liscense a new interface for DRM when the standard usb already works for tha. I agree I hate DRM but to go the path you suggest is way off due to the costs involved when you already have so many other ways to accomplish it.
Overall I think the slot idea is fine but only for specialized applications with specific software. It provides and additional avenue for those with very particular needs. I doubt it will ever be used main stream for the same reasons stated above.
Posted by mssoot (169 comments )
Link Flag
i don't think you'll be swapping
it's not likely that coprocessors will be swappable or removable in an easy fashion. operating systems have to be custom compiled and engineered for single processor or multi processor use, as will many applications.

i have a mod chip in my xbox that let's me run a custom verison of linux instead of the normal microsoft dashboard. i can't swap it in and out (it's soldered in place) but it does come with a killswitch, so i can shut it off if i want to.
Posted by chris__anderson (23 comments )
Link Flag
This has to be for the graphics..
I mean, they just bought ATI right? They are probably planning on plugging in the GPU right there on the motherboard. That could be a significant speed boost...nice.
Posted by M A (51 comments )
Reply Link Flag
Don't expect this on the desktop
I really wouldn't be holding my breath to see this sort of thing in desktop computer systems any time soon. Just adding a second socket to a system board tends to add $100 to the cost before you ever plug anything in to that socket!

It also isn't likely to help too much, certainly graphics would be fairly pointless since the key to improving graphics performance is increasing memory bandwidth. This chip would top out at about 10GB/s of bandwidth vs. 50+ GB/s of current top-end video chips. Other desktop uses just don't present enough of a computing challenge to be of major use.

Now, for high performance computer clusters and maybe even servers, that's a whole other story. Things like a network and encryption off-load engine might help in servers, or a fairly custom (probably FPGA) chip to handle specific calculations for special-purpose servers. The latter is probably more likely for high-performance workstations and HPC clusters. Specific task offload chips (ie what Cray is already doing with non-socketed chips) or high performance computing engines (what IBM and ClearSpeed are doing) all have some very nice applications in the HPC world.
Posted by Hoser McMoose (182 comments )
Reply Link Flag
Your numbers are off
PCI-E has a maximum bandwidth of 2.5 Gb/s (that's gigaBITS pe second) each way, whereas Hypertransport 3.0 has a maximum bandwidth of 41.6 GB/s (that's gigaBYTES per second). That's an incredible increase in bandwidth. Even Hypertransport 1.0 has 6.4 GB/s bandwidth, and they're really looking at putting it to use after they roll out the hypertransport 3.0 systems.

Hypertransport is used to communicate between processors, of course it's higher bandwidth than PCI-E. It would also be much lower latency.
Posted by HoboMaster (4 comments )
Link Flag
 

Join the conversation

Add your comment

The posting of advertisements, profanity, or personal attacks is prohibited. Click here to review our Terms of Use.

What's Hot

Discussions

Shared

RSS Feeds

Add headlines from CNET News to your homepage or feedreader.