Math co-processor

I was thinking, with the various CPU's and instruction sets out there, its pretty obvious the PC is lacking in floating point power due to the X86 and x86_64 instruction sets; however, they are great instruction sets for general purposes.

With the game consoles taking a reduced instruction set processor that is specifically good at floating point operations, I'm curious as to why there is not some method available to get a cell processor and attach it to your PCI or PCI express slot (or some other slot bus), to see a performance boost in floating point operations. Surely these technologies can be brought to the PC as well.

If IBM created a card that would have a cell or a few cell processors on it, a driver could redirect all floating point operations to a device address with bus a broadcast which contains the instruction codes and the data to be worked on. The cards logic could contain all the necessary logic to do the work and then send the completed information back to the CPU. With the driver, the CPU would now be able to understand the results and use them accordingly. I am not sure what the bandwidth requirements of this would be, but PCI-Express x16 slot could probably handle it well. It sounds like a promising way to improve the floating point performance on a PC; however, what is going on with these companies, that they are hesitant to bring technology to the PC the way they did in the 90's?

The x86 / x86-64 instruction set does not restrict FPU performance. SSE1/2/3/ mmx, 3dnow etc are all additional instructions which are a cheap (i.e. lower precision) way to achieve fast FP with little die space.

Your never going to see any kind of Cell processor on a PCI card or even PCI express due to latency & memory bandwitdth issues.

It's not all bad news though, with IBM now saying Power will be Opteron socket compatible it won't be too hard to shoehorn a cell into an AMD based system.

Basically forget PCI/PCI Express for any kind of additional horsepower - Within the next few years we'll have multi socket boards that can take any number of different CPU's/GPU's/PPU's etc.

thingi

There is a cell pci-e card:

http://www.mc.com/cell/products/view/index.cfm?id=106&type=boards

Also with AMDs torrenza initiative, they are encouraging co processors that connect via HT through additional CPU sockets.

Haha, so wow I just came up with the idea last week and it is already on the market! j/k

The x86 / x86-64 instruction set does not restrict FPU performance. SSE1/2/3/ mmx, 3dnow etc are all additional instructions which are a cheap (i.e. lower precision) way to achieve fast FP with little die space.

Your never going to see any kind of Cell processor on a PCI card or even PCI express due to latency & memory bandwitdth issues.

It's not all bad news though, with IBM now saying Power will be Opteron socket compatible it won't be too hard to shoehorn a cell into an AMD based system.

Basically forget PCI/PCI Express for any kind of additional horsepower - Within the next few years we'll have multi socket boards that can take any number of different CPU's/GPU's/PPU's etc.

thingi

Click to expand...

Yeah, I thought about what I said a little and initially decided that there would be problems with PCI for sure. As far as PCI-E, one exists but it is not really usable for a typical user. It would probably be difficult to make most programs make use of it without rewriting the software likely.

I think i was much thinking much more about having additional sockets on board; however, was considering the possibility of PCI-E. I suppose using PCI-E for it is not quite a great idea unless you are a business and have programmers who can make use of it. That would be awsome, if in the future you could build a PC that would make use of different CPU instruction sets etc.

I'd like to see multi-socket motheboards and the emergence of task specific processors. Like, instead of buying a whole video card, you buy the video processor and a generic processor in to a multi-socket motherboard. I'm sure that doing so would require a major architechture overhaul, but it seems to be worth it in the potential for performance increase.

I know my proccesor has a floating point unit, which wroks fine for me.

>HyperlogiK< said:

There is a cell pci-e card:

http://www.mc.com/cell/products/view/index.cfm?id=106&type=boards

Also with AMDs torrenza initiative, they are encouraging co processors that connect via HT through additional CPU sockets.

Click to expand...

so what exactly does that card do? is that like having another 2.8ghz processor?
will 32-bit operating systems be able to address all that memory? thought it was 4gb max.

>HyperlogiK< said:

There is a cell pci-e card:

http://www.mc.com/cell/products/view/index.cfm?id=106&type=boards

Also with AMDs torrenza initiative, they are encouraging co processors that connect via HT through additional CPU sockets.

Click to expand...

Hmm this invention would be great for folding

I think this is something for the future, not now. It will propably be great for thetransition when these multi-socket boards hit the market.

We've had these for a while...

They're called ATi Radeon X1900's.

But I would love to see the return of a socketed math co-processor.

The current spin of the Cell CPU (a la in the PS3) isn't exactly a wonderful option as its only single precision and its SPE (essentially a cut down PPC970 core) only supports in-order operations. Unless Sony/IBM made another spin for use in that card, it should probably only be used in multimedia uses and not in scientific calculations. Also to use that card your software would need to be Altivec (a.k.a. VMX) aware and be compiled against that just like software on the Mac must be in order to take advantage of it.

Also every Mac with an Altivec unit is capable of double precision unlike the Cell. This means that it wouldn't be too hard for IBM/Sony to make one more apt for use in a workstation.

i sent a request for information regarding to product asking if it would help in the performance of games and high resolution video editing...

if so what would be the price tag? lol

waiting for their e-mail response

Eman, it's not allowed to put quote blocks in your signature.

They also have the physics processor.

I was thinking the same as One Bull, more options for folding. I would like to get into folding but dont have the spare resources to put together a folding rig.

Also I have wanted to know this, can you fold on multiple processors at the same time in one machine, such as GPU, CPU and (now that this exists) the cell? Can you fold on them all at once or would you need seperate machines for each?

ok they called me with their response

first off for the board and the 1 year developers license $25,000USD

they told me to thermal and power reasons this board has only been approved for certain systems

the HP XW9300 AND THE XW9400

sun ultra 40-m2

and the

ibm intellistation or something like that

i was also informed that the power consumption was about 200 watts or so

it doesnt look like ill be using this or anybody else will be using this for a while LOL

External coprocessors went mia in the early-mid 90's.
Cost & latency is why they went the ways of the do-do. It's cheaper, and much faster to intigrate it on the CPU itself.