Is there something that CELL can still do better than modern CPU/GPU

[gongo]

This thought just came to me and i like to consult the experts on it... I start by acknowledging the Cell microarchitecture is ancient in computing time, the last and only refresh.. IIRC... was the PowerXCell 8i with DP upgrade....some would call it a deadend one. With the rise of GPGPU...PS4/720 would not be odd to use a Fermi or Larrabe derivative...but call me a dreamer (and a computing idiot)...Cell was born out of one anyway..to make a processor that accelerate dynamic content uncommon at that time.....Cell in PS3 did make some great content in the weakness of the GPU...the stream processors runs at 3.2Ghz...something that neither Nvidia or AMD SP has that raw speed today...I heard Cell was great in FFT and Cryptograph/Security...is it still as good?? PS3 so far has not been hacked by pirates...not an intelligent analysis by any means.....has the modern i7 and HD5800 grown so powerful to be able to emulate...brute force or smart designs..whatever Cell solution you throw at them?

I love the idea that PS4 will use a Cell2 CPU and a GPU...i know Fermi looks so much like the next gen dynamic content processor (what ever that meant *lolz*)...still like my discrete microprocessors.

 

[Arwin]

There are some forum posters here that are much more up to speed, but from what I've read, the Cell CPU design still holds up pretty good imho. But the most important consideration at this point is probably power vs performance / transistor count. Modern GPUs use so much power and have so many transistors, you could probably have a Cell processor with two PPEs and 32 SPEs for the same budget.

I however don't know what the current power consumption is of an SPE and how this holds up for instance against a Larrabee core (which is at 6w or less). But overall, the way the SPEs have their local store as well as how they connect to the EIB, still feels like a pretty good solution and I can't say I've seen many designs that are that much better.

Having said that, I don't know much about i7, or what, say, ATI's latest chips are.

 

[Neb]

.....Cell in PS3 did make some great content in the weakness of the GPU...the stream processors runs at 3.2Ghz...something that neither Nvidia or AMD SP has that raw speed today...

Single Precesion FP or MHz wise?

Either way old GPUs already leaps around the perfomance in these regards. However they might not be as flexible*.

*Excluding 5xxx/Fermi series as I dont know their new capabilities.

 

[gongo]

I am not an expert so the only thing i see is that of the stream processors of Nvidia and AMD do not run any where close to 3.2ghz.

 

[Neb]

MHz/GHz is not a realiable metric. If it where a 3.0GHz P4 Dual-core Celeron CPU would be 50% faster than a 2.0GHz Core2Duo instead of being brutally slower. Or you could say a 4870 has 800SP (units) @ 750MHz and GTX280 has 256 units at ~650Mhz!

About Cell in PS3 it does IIRC ~230Gflops SP perfomance. You have something like 8800GTX back in 2006 with about ~~500GFlops SP. And then you have 4870 with 1200GFlops SP and 5870 with 2700GFlops SP. Not the best metric either but far better than using clock speed. Now about flexibility regarding non graphics is another thing as in what type of calcualtions.

 

[liolio]

I think this is an interesting insight that somehow answer your question

 

[rekator]

About Cell in PS3 it does IIRC ~230Gflops SP perfomance.

If I'm remember well there a ID presentation where they said CELL in PS3 is 192 Gflops, but don't know if it's for all the CELL or for the Glops for devs, so 5-6 SPU +PPU.
If your 230 Glops is for 7 SPU+PPU, the 192 Glops can be for 6 SPU+PPU, don't remember PPU Gflops.

 

[zed]

I think this is an interesting insight that somehow answer your question

except its incorrect

I believe even with todays latest GPUs u still cant throw any code at them like u can with a cpu, no doubt this will change in the future

 

[Panajev2001a]

If I'm remember well there a ID presentation where they said CELL in PS3 is 192 Gflops, but don't know if it's for all the CELL or for the Glops for devs, so 5-6 SPU +PPU.
If your 230 Glops is for 7 SPU+PPU, the 192 Glops can be for 6 SPU+PPU, don't remember PPU Gflops.

PPU is 8 FP ops (MADD) per cycle. So 8 * 3.2 GHz = 25.6 GFLOPS (Single Precision).

The same peak value is valid for SPU's which are worth 8 FP ops/clock cycle (Single Precision).

 

[upnorthsox]

PPU is 8 FP ops (MADD) per cycle. So 8 * 3.2 GHz = 25.6 GFLOPS (Single Precision).

The same peak value is valid for SPU's which are worth 8 FP ops/clock cycle (Single Precision).

I believe that's per thread on the PPU so that makes it about 204.8 gflops for 6spu + ppu.

 

[Arwin]

Here's an interesting list that shows supercomputers and their power-to-watt rating:

http://www.green500.org/lists/2009/06/list.php

I'm not sure how good this list is, but it's been running for a while, and 536.24 MFlops per Watt seems pretty impressive! And that's the #1 ranking machine which sports PowerXCell 8i 4.0 Ghz processors.

Though of course this gives us little information on the efficiency of GPUs as I don't think there will be many supercomputers based on GPUs on this list (I vaguely recall that Google is experimenting with that).

@upnorthsox: yes, I think you are correct.

 

[Grall]

I believe that's per thread on the PPU so that makes it about 204.8 gflops for 6spu + ppu.

It only has 1 FP unit; the PPU doesn't have two full hardware threads...

 

[Panajev2001a]

It only has 1 FP unit; the PPU doesn't have two full hardware threads...

It does AFAIK, but that does not mean having duplicated execution resources.

 

[Panajev2001a]

I believe that's per thread on the PPU so that makes it about 204.8 gflops for 6spu + ppu.

No, while the PPU has two HW threads and thus two sets of VMX registers, it only has a single VMX unit (the same goes for the VMX-128 units of the Xbox 360 CPU cores).

 

[upnorthsox]

No, while the PPU has two HW threads and thus two sets of VMX registers, it only has a single VMX unit (the same goes for the VMX-128 units of the Xbox 360 CPU cores).

Yes I went back and checked and it's integer ops that both threads can run not FP. Sorry.

 

[Billy Idol]

what interests me: why not the other way around!?
why don't you game dev guys use the GPU for physics calculation?!?!?
I suppose that there is a lot of stuff a GPU is faster in compared to CELL (I am not only talking about graphics stuff)

We in our research group are evaluating scientific computing using GPU...and they have a far better potential than every known CPU (i.e. GPUs are damn fast CPUs!!!)?

We also evaluated CELL for scientific computing...but lets put it this way: never (at least for us)

why even bother with a fast CPU, if all the physics can be calculated using the GPU....just replace the CPU with a moderately fast GPU ...

 

[liolio]

what interests me: why not the other way around!?
why don't you game dev guys use the GPU for physics calculation?!?!?
I suppose that there is a lot of stuff a GPU is faster in compared to CELL (I am not only talking about graphics stuff)

We in our research group are evaluating scientific computing using GPU...and they have a far better potential than every known CPU (i.e. GPUs are damn fast CPUs!!!)?

We also evaluated CELL for scientific computing...but lets put it this way: never (at least for us)

why even bother with a fast CPU, if all the physics can be calculated using the GPU....just replace the CPU with a moderately fast GPU ...

For me the answer is simple, it seems like programming gpgpu style takes a lot of time. So far there were no standard for programming, now you have openCL and direct compute, it's just a matter of time before middle-ware vendors jump in. It will be easier to recover cost on PC and next generation systems.
Sebbbi pointed other issues, synchronization issue, it takes sometime for CPU to recover calculations done on the GPU. Fusion may be a huge win in this regard or the other way around you may run a lot more stuffs on the GPU.

Also this post is interesting in this regard (posted twice but it looks like nobody noticed).

 

[Arwin]

what interests me: why not the other way around!?
why don't you game dev guys use the GPU for physics calculation?!?!?
I suppose that there is a lot of stuff a GPU is faster in compared to CELL (I am not only talking about graphics stuff)

Sure, but it depends on whether you want to compare apples to oranges. Like I said before, it's probably wise to consider power-to-watt ratio if you want to compare the efficiency of various processors. But also, this varies according to the type of calculations you wish to perform. If you look at the Folding@Home project, you will see that for some calculations, a PC is more efficient, for others, the Cell, and for yet others, a GPU.

why even bother with a fast CPU, if all the physics can be calculated using the GPU....just replace the CPU with a moderately fast GPU ...

Good luck running a modern PC without a CPU.

 

[Crossbar]

Good luck running a modern PC without a CPU.

Couldn´t Larrabee pull off this trick?
I mean the Arm cores in the NetPCs are not really any power horses, still they work pretty well.

Couldn´t one or a couple of the Larrabee cores be assigned to run some OS and app code on a similar level?

 

[Billy Idol]

For me the answer is simple, it seems like programming gpgpu style takes a lot of time. So far there were no standard for programming, now you have openCL and direct compute, it's just a matter of time before middle-ware vendors jump in. It will be easier to recover cost on PC and next generation systems.

Maybe CUDA is the solution? ... who knows

Sure, but it depends on whether you want to compare apples to oranges. Like I said before, it's probably wise to consider power-to-watt ratio if you want to compare the efficiency of various processors. But also, this varies according to the type of calculations you wish to perform. If you look at the Folding@Home project, you will see that for some calculations, a PC is more efficient, for others, the Cell, and for yet others, a GPU.
Good luck running a modern PC without a CPU.

I am not so sure. I mean, what I really wanted to ask: is there really anything that CELL can do better than a GPU with respect to non-graphics stuff?

What we concluded: if the graphics guys sort out their performance for double precision and introduce ECC memory, a standard GPU is a real challenger in super computing.