Slashdot Mirror
Cell Chip Coming To the PC Via a PCI Express Card
arcticstoat writes with an excerpt from Custom PC: "After developing a brand new CPU architecture from the ground-up, you'd expect that Toshiba, Sony and IBM would have more uses for the Cell architecture than the PlayStation 3, and Toshiba has been quick to make use of the architecture's HD video transcoding abilities in its new Qosimo laptops. However, Leadtek is now taking Toshiba's efforts a step further by putting the chip onto a PCI-E card for desktop PCs. The WinFast PxVC1100 is based on Toshiba's SpursEngine SE1000 processor, which is a cut-down version of the Cell chip. The SpursEngine chip features four SPEs (synergistic processing elements) based on 128-bit RISC cores, along with H.264 and MPEG-2 codecs, but it doesn't contain its own CPU as the chip in the PS3 does. The chip is capable of encoding and decoding H.264, MPEG-2 and MPEG-4 video streams in hardware."
this + mythtv = interesting possibilities
#1: Is there going to be a Mac Version? I would love to put this in my Apple Tower, I have 3 PCI-E x16 slots sitting around doing nothing. #2: When is this actually going to come out? I mean, I keep reading things on "fantastic pieces of tech" and they either never come out, or they come out everyone forgets about them. Anyone know what this should retail for, or if software can even take advantage of it yet?
This spurs engine sounds just like an extra GPU...
Why not just go with CUDA or some other GP-GPU platform and avoid the hassle?
I know nVidia and AMD/ATI are doing H.264 decoding in hardware using their GPUs... I'm sure you can get software for encoders too.
Sigs are for the weak.
...can it play Crysis?
Because if not, seeing as modern graphics cards all feature hardware MPEG, I'm kind of underwhelmed by this announcement.
The mandatory "does it run Linux ?" boils down to "do they provide enough documentation to write drivers for it ?".
I RTFA, but I didn't find an answer in it.
::checks case::
Ooh, awesome! I have one more PCI-E slot left, right next to my PhysX accelerator! Where do I pre-order?
Wise men say, "Forgiveness is divine, but never pay full price for late pizza."
-in linux, no. only mpeg2 decoding
-in any OS, not really. There is a brand new ENCODER for h.264, but reviews show it to be crap and limited
Windows does have full GPU decoding of h.264 with modern nvidia (not sure about ATI, but it is likely), but that's it.
Decoding .264 isn't really such a big deal. The ability to do low-cost multi-pass 1080 h.264 encoding at greater than real-time is something that would be EXTREMELY welcome for my company. We're a video post production house and we burn *LOTS* of CPU cycles encoding video for delivery to clients. A sub $500 card that greatly streamlined that process would be VERY welcome. Especially if it's something you could do as a background process that effectively didn't interfere with the operation of the edit suite.
-Steve http://www.stevennicholson.com
The fate of this device hinges pretty much exclusively on the quality of its software and documentation. If all you get is some gaudy half-broken-and-all-ugly fixed purpose video encode decode app(in the fine tradition of graphics card shovelware, remember the bad old days when the card vendor was responsible for the driver?) then this thing is dead in the water. A few will sell to Netflix pirates looking to rip and encode 3 times as much video as they could ever watch, instead of just twice as much; but that'll be about it.
If it has good general purpose support(I'd really prefer that this mean "good documentation" and properlinux support; but I suspect a proprietary sdk would do alright as well) then it could be a killer in certain lower end computing scenarios. Since the cell is produced in nontrivial bulk, and this thing is only about 1/2 the complexity of a full cell(does that mean that this card is "spursengine on the half-cell?) it should be cheap, cheap, cheap compared to FPGA boards or custom ASICs for such purposes as the cell architecture is useful.
I hope the do the right thing, and get rewarded(and I hope so, surely somebody looking to sell computational hardware would see the virtues of making it as useful as possible for as many customers as possible?); but if they don't, I suspect that they'd be lucky to do as well as physX, and will probably do worse.
I want to be a synergistic procesing element!
Doesnt everyone?
NO SIG
Yeah, and there's that whole "Roadrunner" thing, fastest supercomputer in the world. And IBM sell Cell bladeservers...
Only pci-e x1 and 128meg of ram? ati, nv cards have more ram at a lower cost with a pci-e x16 link.
The x1 link will slow this down. HTX is even better then pci-e for a add in cpu.
Its not meant for playback of a single video like the GFX cards do, or watch a DVD or Blu-ray, its designed for content creation and distribution. In an early demo, the Cell did 48 simultaneous Mpeg2 streams in real-time.
http://gizmodo.com/gadgets/home-entertainment/playstation/cell-processor-demos-mpeg2-x-48-100853.php
meep meep?
Probably. We already have enough information about the CELL processor on its own to make use of it under Linux; this card is just taking a cut-down CELL and tacking a PCIe bus on it.
Unless they purposefully fucked the register table to prevent it, it's probably just a matter of finding the correct PCIe offsets to access known registers/segments on the CELL. While it's possible they could "sabotage" it to prevent the first-day-out-of-the-box Linux driver, chips modified this way usually have to go under more steps of formal validation again (beyond that of just throwing a PCIe controller on, and sheering a few SPUs off), so most companies won't do it.
Before we get too confident, though, there is a history of this kind of intentional fucking. Conexant acquired some video IP from a defunct company Brooktree, the BT8x8 model, which worked fabulously under Linux, which they re-released with virtually unchanged functionality, but with a completely revamped address table. Brooktree was more friendly and released the specs for its chips to the public, so the Linux driver was fantastic. Since Conexant would not release the new specs without an NDA (and is generally is Linux's bane when it comes to hardware), it took months to get the new driver back to the shape that the old one was in (and IIRC, it was only after someone stepped forward and went under the NDA to do so).
The Cell is a perfect example of how not to design and build a multicore processor. It's a powerful processor but it's a pain in the ass to program. The worst thing that a multicore designer can do is build a processor before the programming model is designed and tested and all the chinks ironed out. But Sony and IBM are not alone. Intel is making the same mistake with Larrabee. AMD is soon to follow suit with its Fusion hybrid. It's enough to make a grown man cry. The truth should be clear to everyone by now. Heterogeneous processors are not the way to go simply because there is no easy software model that makes them easy to program. GPUs are not the answer either because they lack universality. As Tim Sweeny said recently, what is needed is a homogeneous processor. It will do wonders for productivity. Homogeneity and universality is what is called for. The Cell is anything but.
In my opinion, both the CPU and the GPU are doomed for the simple reason that they are not universal. There is only one type of parallel processor core that can handle anything you can throw at it and that's a pure MIMD vector core. None of the multicore vendors have one none are planning to build one. Why? Because they don't have the right programming model. Unless they see the error of their ways, some other organization will do the right thing and rocket past them. They won't know what hit them until it's too late. The writing is on the wall.
sweet, i can finally have my PVR record programs before they actually air!
but seriously though, how much is this card going to cost? is it just for professional video processing or will there be other uses for it as well? i wouldn't mind having one of these things for a PVR/media center, except for the fact that it needs a one-slot cooler, meaning it probably runs hot and noisy.
Lets say the PS3 retails for £300 (it's less than this, but what the hell, this is slashdot, we don't need to be accurate. Or impartial for that matter...let me start again) Lets say the shitty PS3 costs £300, which is far too bloody much, but once you take away the shitty Blu-Ray drive, the shitty Hard drive, shitty controller, shitty case, etc. the price for the shitty fully-fledged CELLs (7 of them, remember) can't be more than £100 and that's a safe overestimation, with added money for the Lube Sony will use to anally violate you with their shitty cocks. This chip has only 4 shitty cores of the shitty CELL and it's not even the full CELL, it's a shitter version of it so I'd say it's a safe bet that it SHOULD cost no more than £50-70, but since the company that makes it is so shitty, they'll probably triple that price. Cunts.
+1 IDisagreeSoHeMustBeATrollOrAnAstroturferOrAShill
What does it have over a normal multi-core processor, like say a Core 2 Quad?
The problem I've been seeing with the Cell both in terms of how it performs in the PS3 and the researchers tinkering with it at work (I work for a university) is that it doesn't really seem to have something that it is great at. A lot of the tasks people tout for it are highly parallel tasks, like Folding@Home. Ok, wonderful, except a GeForce crushes it. A GTX 280 using the CUDA client is much faster than a Cell. Ok so, not for tasks like that. You say it is more applicable to general purpose tasks. Fine, but we've got that already. Intel's Core and AMD's Athlon processors are some amazing general purpose processors for some amazing prices. From what I've seen, at regular CPU tasks, it can't keep up (in the PS3's case, the CPU core that has to dispatch everything to the cells gets swamped). So then what is the market?
You'll note that here one of the things listed is H.264 encode/decode. Well that IS something that GPUs do and quite well. The decode functions ship with newer drivers. As for encoding, there's a program called Badaboom that uses the GPU to do the encoding. So thus far a lot of the things I've seen the Cell marketed for (video, physics) are things GPUs with CUDA kick ass at.
The real question isn't what can this card do, it is what can it do better and/or cheaper than either a CPU or GPU? Doesn't matter if it can do everything that they can do, if it is turning complete it can pretty much by definition. What matters it can it do it for either less dollars or in less time (or better yet, both)?
Also in the video domain it has to compete with ASICs. I don't know about H.264 but there are very cheap MPEG-2/4 codecs. Each chip does 4 realtime encodes and decodes, and they aren't pricey. Computer based CCTV systems use them all the time.
Mercury had a PCI-e cell expansion card for over a year now.
Unlike the leadtek one, the mercury version has the full version of the cell processor, with 8SPEs. Dont think it comes with any prebuilt codecs though.
SpursEngine is not a partial good Cell; it's a different chip.
i wouldn't mind having one of these things for a PVR/media center, except for the fact that it needs a one-slot cooler, meaning it probably runs hot and noisy.
Look at the pictures : the cooler looks rather small, and seems to be of the standard type that you find over most low-end GFX cards and some chipsets.
As long as there's sufficient air-flow in your HTPC, you could probably swap if for on of those heat-pipe based monstruosities that you can fit over standard GPU and use passive cooling or low noise big fan. (something like this)
Of course, given the standard shape, you could also put a water cooling block on it.
but seriously though, how much is this card going to cost? is it just for professional video processing or will there be other uses for it as well?
Well, I think this is going to be the tricky part.
On one hand IBM and company have gone through great effort to diversify and push their chip to as much as possible different usage (more product sold = cheaper to build) so I don't think it'll have as much problems as the PhysX chip had.
But on the other hand, the main usage for which this card is marketed for (accelerating HD decoding) is already supported by the hardware acceleration inside latest graphics card on the biggest market (Windows machines).
Most HTPC builder running Windows Media Center, will probably prefer to put in one of the latest HD-enabled GeForce or Radeon (or even the latest onboard Chrome from VIA), rather than having to buy two separate cards, one graphic card AND one HD accelerator.
Thus I don't know if there's going to be enough demand to drive the price low enough.
And that's bad for us Linux users, because we don't get such a good support for hardware decoding in graphics card, and developing it will be slow. Whereas the CELL based solution would be much more easy to tackle as there's already tons of code and tools for the CELL's SPUs.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
This card is supposed to do HD Encoding. Not just decoding.
Show me a current application that uses your GPU to do good H.264 or MPEG2 encoding in realtime and I'll bite.
- It's not the Macs I hate. It's Digg users. -
After looking briefly at this, it is an entirely different beast. It is a full Cell processor, with 8 SPE's and a PPE and 5GB of RAM, as well as Flash memory that allows it to boot Cell Linux. It's really a system on a PCI-E board, rather then an accelerator board.