Mercury Computer Systems announced a USD 7999 accelerator card Tuesday that uses the Cell Broadband Engine processor that plugs into a computer’s PCI slot. The Cell Accelerator Board, which will be generally available in the first quarter of 2007, can speed tasks such as signal processing or image rendering, Mercury said in an announcement at the Siggraph computer graphics show.
wow thats expensive.
And just PCI? no PCI-Express or anything? pci is what, 133MB/s? I hope that card has a lot of onboard memory.
… that this $8000.00 monster of a card using *PCI* slots also has some kind of AI…
For 8 grand I can build 4 really fast graphic stations. Or even a 4 way smp, dual core (8 core) rendering node. This looks like more sony-style (STUPID) pricing, just like the ps3. We don’t even really have a clue just how fast the cell is. Lots of talk, no show.
Idiot.
http://www.mc.com/cell/
Where does Sony enter into the equation? IBM designed the chip, Mercury manufactured the card.
In any case, whether the price is worthwile really depends on what you’re doing with it. Where I work (a wireless comm. startup) some of the RF-engineers do signal processing on boards with a big FPGA and a G4-class PPC core. These boards start at double the price of the Cell board mentioned here, and go up to $30k+ for a model with dual FPGAs and dual G4s.
If this board can achieve comparable performance to some of these products, it’ll do quite well. Not only is it relatively cheap for this particular type of product, but programming a GPP in C is a hell of a lot cheaper (in terms of engineer time), than programming an FPGA in VHDL.
Edited 2006-08-02 01:28
I was equating the pricing model to sony’s, because it’s in parallel (same chip design even..)
Well sure, it depends on what you’re doing with it. Maybe for fluid dynamics it’s great. Again though, this is going back to the SGI workstation era, with the 30k+ workstations. You’ve seen how well that panned out for them.
About where you work, somebody is getting ripped off. You’re funding somebody huge amounts of money for their design time, which while I’m sure is substancial (since you’re working for a wireless company I’m sure you’re quite aware how substancial), but paying 30k for maybe 1k worth of materials and fab time is a bit excessive. Again, see 80s/early 90s era workstations.
I was also under the impression there was no GPP programming on the cell. A lot of the complaints I’ve heard by developers on cell-based systems is how hard it is to program for them. Maybe I’m wrong and IBM/etc are putting out a gcc patchset (yuck compiler) that allows you to compile code into something the cell can use, but given the cell’s design – it’s going to run like dogcrap unless you write your code with the cell in mind. It’s going to be the same headache as VHDL, just of a different sort (have fun learning another new arch….)
I guess we’ll see how it all pans out, there are always buyers for products, no matter how insane they are, because some people/companies will pay top dollar for a 1% increase in speed. I just don’t see how the cell revolutionizes anything, and I especially don’t see how a company can churn out a board with a cell on it for 10x the cost of a company which is churning out a video game system with multiple cells, and all kinds of other stuff. Blame low production runs/development costs, but it’s a bit absurd. I don’t see it being a success, just like I don’t see the cell architecture itself being a success, at least not in the near future. Maybe I’m wrong though, never know!
Cheers,
David
About where you work, somebody is getting ripped off. You’re funding somebody huge amounts of money for their design time, which while I’m sure is substancial (since you’re working for a wireless company I’m sure you’re quite aware how substancial), but paying 30k for maybe 1k worth of materials and fab time is a bit excessive. Again, see 80s/early 90s era workstations.
By your logic, software is a rip-off because you’re paying hundreds or thousands of dollars for a few dollars worth of materials. That’s ridiculous. In many low-volume products, the cost of design far outweighs the cost of materials. Somebody has to design the product, because the product fills a need that needs to be filled, and if the market for that product is small, the price is going to be high. What Mercury is actually bringing the benefits of volume to this market with this Cell-based design. Cell is going to be a lot higher volume than the huge FPGAs you usually put on a board like this, which explains why this board is relatively cheap compared to other boards like it.
I was also under the impression there was no GPP programming on the cell.
Cell is completely GPP. It runs Linux, and has GCC as a compiler. Each SPE can be treated as a general purpose computer with a small local memory, and a really fast interconnect to its neighbor SPEs and a fast I/O link to a main data store. It’s a little bit different than your average PC, but it’s a far cry from an FPGA, or even a DSP.
Maybe I’m wrong and IBM/etc are putting out a gcc patchset (yuck compiler)
What’s wrong with GCC? Its the only compiler in the world that’ll target everything from a supercomputer to a router board, is well-understood, and has good error messages. Heck, it even generates decent code on a lot of architectures, though for many of the uses of GCC, code performance is really far down on your list of concerns.
It’s going to be the same headache as VHDL, just of a different sort (have fun learning another new arch….)
Hardly. Cell lets you program regular old C, with the proviso that the directly-addressable memory (the local store), is small and the main memory is more like a fast central storage accessed via DMA operations. It’s got a different concurrency model (self-contained batch jobs versus shared-memory threads) from a convential PC, but then so do clusters (message passing), and lots of people program those just fine.
I guess we’ll see how it all pans out, there are always buyers for products, no matter how insane they are, because some people/companies will pay top dollar for a 1% increase in speed.
Very few companies pay top dollar for a 1% increase in speed. For the tasks these types of products are usd for, the question usually is “I’ve got to 200 msamples/sec to FFT in real time, what can I find that’s fast enough?” The answer is, in most cases, a piece of hardware that costs a lot more than your average PC. However, it really doesn’t matter how much cheaper the PC is if it can’t keep up the data rate you need, does it?
Edited 2006-08-02 02:22
It’s PCI express.
You’re right:
http://www.mc.com/cell/products/view/index.cfm?id=106&type=boards
Not much info in that short article IMO
It is really expensive but if this turns your tame little PC into a hugely capable workstation then it might be worth the price .
And of course – spare money – you can allways slot in a few more .
Im severly missing information on how many Cell processors this board will have – for 8000 US Dollars it has to be more than one lousy processor .
Why does it say “can speed up” – how about “will” ?
“The board can run YDL” – So ?
Are they suggesting it does .. possibly will or did they just want to remind that Linux is the only OS to run on Cell ?
Okay .. yes it does according to another link.
Actually the link behind “Mercury Computer Systems” IMO interesting.
IMO this is even more trouble for SGI as Mercury are invading exactly their market as far as I think – although a PC can not handle the huge data amounts SGI machines can – but very much same market .
IMO
BTW anyone know if this can speed up games ?
It’ll speed up anything that you write software for it to run. If you buy Epic one then maybe it’ll speed up UT2K7 (and all the other games that are based on the Unreal eninge) 😉
If they had to sell one million the price would be half. 2 millions, one fourth of the original price and so on. But If they spend some more design time, it would be ideal for everyone. No just a few companies. But I cannot understand why we couldn’t have a lot cheaper solution with a mother board consisting of 4 processors and 4 cores (like upcoming solutions of AMD). MIPS64 also offers exactly the same capabilities for I/O . PMC Sierra has a great portfolio of products. The problem is design decisions
I believe that there should be cheap motherboards taking 4 or 8 processors. VIA also has very small processors consuming little power, with its upcoming C7 that could be placed on multi CPU motherboards. The end user should have the ability to put 8 C3 or 8 C7. However the interconnect is not of the CELL level but it is not negligible and the release of their DP product suggests that it can be done. Maybe the V4 bus has many enhancements and it is ready for multi CPU motherboards.
MIPS64 from Sierra have these capabilities and could provide excellent multi CPU motherboards. It seems that they have different plans.
There are two problems with using something like four Opterons in this setup. First, a quad Opteron setup is likely to be just as expensive, when you factor in the cost of quad-core motherboards and Opteron 8xx’s. Second, it’s likely to be slower. SPEs are specially made for stuff like signal processing. IBM has benchmarks showing one SPE outdoing a 2.7 GHz G5 by a substantial margin, even though the G5 itself outdoes an Opteron by a large margin in that benchmark. Moreover, the SPEs have a much faster interconnect between them, and access to a single pool of very high bandwidth memory. Third, the power and space requirements of that setup would be a lot higher. I don’t know how other people use their signal processing cards, but if they’re like us, they’ll be stuffed in tiny, poorly ventilated enclosures or makeshift racks, and manhandled as they’re moved around the lab. A quad-core Opteron setup wouldn’t be nearly as amenable to that as a PCI-E card you can stuff in a small computer case.