PDA

View Full Version : Cell accelerator board


Jeff Lew
08-01-2006, 04:00 PM
http://www.tgdaily.com/2006/07/31/mercury_announces_cell_accelerator_board/
(http://www.tgdaily.com/2006/07/31/mercury_announces_cell_accelerator_board/)
This can get pretty interesting. It says it can run Linux and has 15-30 times the floating point power of regular PC's.

Jeff Lew

Fede
08-01-2006, 04:55 PM
That sounds great, obviously cause it's such new technology it's still quite expensive $7000. But i can see the benefit.

If they really put that in to the PS 3 it should help to bring the price down sooner than later, but who knows when the PS3 actualy will be released......

I guess we will have to wait and see

GreyWolf_OPS
08-01-2006, 06:15 PM
I am curious on how many working cores the Cell will have on this $7K - $8K board. If anyone recalls reading about the horrible yield that IBM was getting. They said that they were getting decent yield if you didn't need all 8 SPU's or SPE's. So the real question is, is the $7k Cell board using a gimped Cell chip, cause if it is, its not really worth it seeing how the PS3 will have the Cell chips with 8 working cores. One in reserve incase one fails, which seems to be a common problem with them from what some developers have said.

And another thing for $7k the board better have drivers to integrate with 3D software like Maya or XSI. Otherwise it would not be worth owning such an expensive card and no software to use it.

daart
08-01-2006, 08:42 PM
Wow, that card looks impressive!

http://imagecloset.net/img2/out.php/t19444_CellBECard.jpg (http://imagecloset.net/img2/out.php/i19444_CellBECard.jpg)

visualride
08-02-2006, 01:59 AM
Wow, that card looks impressive!

Hey, that looks like my old 200 meg hard disk card. :shrug:

DaJuice
08-02-2006, 03:48 AM
Err... I think you specifically need to code for these things. I don't expect that you can just drop one of these into your workstation and expect it to accelerate anything.

tozz
08-02-2006, 05:21 AM
I am curious on how many working cores the Cell will have on this $7K - $8K board. If anyone recalls reading about the horrible yield that IBM was getting. They said that they were getting decent yield if you didn't need all 8 SPU's or SPE's. So the real question is, is the $7k Cell board using a gimped Cell chip, cause if it is, its not really worth it seeing how the PS3 will have the Cell chips with 8 working cores. One in reserve incase one fails, which seems to be a common problem with them from what some developers have said.

And another thing for $7k the board better have drivers to integrate with 3D software like Maya or XSI. Otherwise it would not be worth owning such an expensive card and no software to use it.
They aim for 7 cores for the PS3, nothing else. I dunno why people have such a hard time understanding this fact when it has been stated about a million times.

Do you have actual quotes from real developers saying their dev kits has had cores breaking down?

ffear
08-02-2006, 02:17 PM
Err... I think you specifically need to code for these things. I don't expect that you can just drop one of these into your workstation and expect it to accelerate anything.

If you read the article you'ld see that it runs a Linx OS in that card so any rendering that you'ld do would work as long as it supports Linx. It's the same concept as multi core processors, one chip doing one thing and another doing another. So Maya, Renderman and Maxwell all would work right out of the box!

-=TF=-
08-02-2006, 05:47 PM
YAY! :eek:
cool gadget!

jewalker
08-02-2006, 06:16 PM
If you read the article you'ld see that it runs a Linx OS in that card so any rendering that you'ld do would work as long as it supports Linx. It's the same concept as multi core processors, one chip doing one thing and another doing another. So Maya, Renderman and Maxwell all would work right out of the box!

That's only if the version of Linux you run has been modified to support the cell processor. The cell processor has a single Power Processor Element which is a PowerPC chip. With this chip alone you can run Linux and almost all other software, but that doesn't gain you the performance of the additional processors, which people are calling cores. The SPEs (Synergistic Processing Elements) as they are called are not general purpose cores that run like a dual core processor, but are vector processing units with their own local memory. They need to programmed individually in order to be utilized. Only software written specifically for the Cell processor will notice any significant speed improvements.

http://www.blachford.info/computer/Cell/Cell0_v2.html

billrobertson42
08-03-2006, 12:17 AM
Err... I think you specifically need to code for these things. I don't expect that you can just drop one of these into your workstation and expect it to accelerate anything.

That would be half of the fun. :bounce:

$7000? :banghead:

Gurki
08-06-2006, 11:59 PM
If you read the article you'ld see that it runs a Linx OS in that card so any rendering that you'ld do would work as long as it supports Linx. It's the same concept as multi core processors, one chip doing one thing and another doing another. So Maya, Renderman and Maxwell all would work right out of the box!

Hi,

it's not as simple as that. Cell is no symetric multiprocessing, it's an
PowerPC with 8 SIMD units (single instruction, multiple date, also known
als AltiVec on G4 or MMX/SSEn on x86).

So even if there are commercial render applications on (non-x86) Unices
- for example Linux on PPC, they won't work on the SPEs of the CELL
architecture. Even if you recompile the source, it will fail to run, because
the SPEs are quite different from typical general purpose CPUs. No simple
lod/store with a L1/L2 cache hierarchy, no branch prediction, no out of order
execution, etc.

Explicit caching in the 256k private ram of each SPE is used and explicit
DMA transfers to main memory ! You eveb have to do your own threading
(scheduling) as an application programmers. So it's hard work, to use the
(high) potential of this architecture (as software developer).

And most people think wrong, if they think raytracing is FPU bound.
This was true 5-6 years ago (for x86, for PPC/MIPS/Alpha even 8).

On a Xeon 3,6 or Opteron 2,6 the real problem is not MFLOPS but
memory latency and branching (instruction cache misses and pipeline
stalls in traversal code for e.g kd-trees).

In a kd-tree you only need 1 fp-multiply and 1 fp-add to traverse a
cell. On single ray traverses hundreds to thousands (depending on scene
complexity) kd-tree nodes before it finds some geometry to do an
intersection test (much more to do for the fpu) and in most cases only
3-8 intersection tests have to be done (if the kd-tree is well balanced).

So to look at GigaFLOPS and think "this will raytrace in seconds" was
also the wrong idea when people thought the raw power of modern
programmable GPUs was suited to do things like raytracing.

but the memory architecture is most important for any non trivial
3d scene. cache coherent traversal algorithms and ray bundles
(with SSE use) were the main targets for "realtime ray tracing"
on x86. Read the PhD-thesis' of Ingo Wald and Carsten Benthin for
more details!

They founded a company to sell their RTRT software called intrace.
But don't think an Core 2 Quadro will be enough. Car manufacturer
VW bought 50 Dual Opterons for an RTRT cluster to get descent
quality for their PowerWall.

So inTrace is investigating this Mecury CELL Board and they wrote an
technical report on it (warning: it's more an scientific paper than some
marketing fuzz with nice pics, if you don't know kd-trees, go and
read Matt Pharr's Physically Based Rendering Techniques first):

http://graphics.cs.uni-sb.de/~benthin/


Greetings

Gurki

Thalaxis
08-07-2006, 01:25 PM
They said that they were getting decent yield if you didn't need all 8 SPU's or SPE's. So the real question is, is the $7k Cell board using a gimped Cell chip, cause if it is, its not really worth it seeing how the PS3 will have the Cell chips with 8 working cores.


It only has 7 SPE's, plus one PPC core.


One in reserve incase one fails, which seems to be a common problem with them from what some developers have said.


It's not a spare in case of failure, it's a spare in case of manufacturing defects. Its purpose it to increase yields, not address reliability issues.

While it is possible that there were failures in Cell processors delivered to developers, it's almost certain that they were pre-production parts delivered to give the developers a head start on their learning curve -- think "beta".

In a shipping model, as long as the system integrator builds the system's power and cooling to the proper specs, I seriously doubt that reliability will be a problem. The problem will be living up to the hype. :)

CGTalk Moderation
08-07-2006, 01:25 PM
This thread has been automatically closed as it remained inactive for 12 months. If you wish to continue the discussion, please create a new thread in the appropriate forum.