Support NeoGAF

[iapetus]

Actually, CELL is based around the 970. Expect about 80-90% performance compared to an equivalently clocked 970. Where it goes nuts is that there's a number of vector units attached that are basically "VMX on steroids" to quote one of the main guys at IBM behind this. The vector units (or Data Plane Processors as they're calling them) can also communicate between each other as well as with the central core. The workstations are actually headless server blades, each of which will have 2 CELL's on them and they'll be running Linux.

This stuff isn't bullshit, it was all disclosed Thursday at the Australian Game Developers Conference. I didn't sign a NDA so it's all good. I also fondled a PSP =]

 

[gofreak]

Sweet!

Two cells = Two PEs?

Any word on the number of APUs/Data Plane processors attached to each 970?

If it's 8, then those workstations are probably clocking in at about ~500Gflops (at 4Ghz)?

edit - also, do you have a link?

 

[iapetus]

The main Slashdot story

The post quoted above

 

[McFly]

If it's 8, then those workstations are probably clocking in at about ~500Gflops (at 4Ghz)?

And I'm sure there's a fiber connection on the back to connect a second blade to get a full teraflop or if you connect 32 blades together you get that 16TF workstation.

Fredi

 

[gofreak]

We really just need a few more bits of information here..

One "Cell" is one PE, I'm guessing?

How many APUs are attached to each 970 in these workstations?

What's the clockspeed?

And did they claim any specific floating point op performance?

This info would go a long way (and I see iapetus has asked some of these questions)

 

[Fafalada]

It would sound nice if true - but somehow I have a very hard time believing it.

Anybody feel free to dispell my skepticism please

 

[McFly]

From B3D right into this thread: Data Plane Processing with Configurable Architectures

Fredi

 

[Panajev2001a]

At the AGC ? The conference program was changed at the last minute there and the new one did not seem too specific on CELL architecture.

They divulgated that kind of info without requiring an NDA in August and they wait to basically "repeat" it next February at the ISSCC 2005 ?!?

I do not think the PUs are "based" on the PowerPC 970 as in they area modification of it: the PowerPC 970 is not what IBM chose for Xbox 2/Xenon and I do not think that for a "clean sheet" design like they called CELL they slapped-in a huge core such as the PowerPC 970 (which is a bit too over what you would need for a PU).

Also... >4 GHz on 90 nm ? The PowerPC 970 does not go anywhere close to that speed (they are struggling now to push the design to 3 GHz).

Also, if that was the PU and VMX on steroids was the APU then the SDK would have been in developers' hands (a hardware one) a long while a go: they would have shipped dual G5 based systems, the same ones used for Xbox 2/Xenon development.

Why ? PowerPC 970 instruction set... VMX instruction set (the VMX units on Xenon is likely "on steroids" as the "leaked specs" indicated)... it would make sense. Yet until October there was nothing even close to a hardware based PlayStation 3 SDK for third parties.

Maybe the PU shares some "tidbits" or ideas with the PowerPC 970 as it does with the POWER5 (OS Virtualization for example), but it is an over-simplification to say that it is "based" on those cores as if it were a modification of either of those two processors.

 

[Fafalada]

Also... >4 GHz on 90 nm ? The PowerPC 970 does not go anywhere close to that speed (they are struggling now to push the design to 3 GHz).

In all fairness, we don't know yet if the PU runs at same clock speed as the APUs.

Anyway, as you pointed out, the post is very likely to be BS either way.

 

[iapetus]

Well, FWIW someone else followed up saying that they were also present.

 

[Fafalada]

Well, FWIW someone else followed up saying that they were also present.

Well, like I said, feel free to dispell my doubts.
But until I see something more official I don't remain very hopefull.

And check your PMs

 

[Panajev2001a]

In all fairness, we don't know yet if the PU runs at same clock speed as the APUs.

Anyway, as you pointed out, the post is very likely to be BS either way.

The APU part is more suspicious... it takes a lot of steroids to change a unti that can only work in co-processor mode like Altivec/VMX and change it into what the patents show us as the APU.

We go from VMX with its three separate sub-units (simple Vector ALU, Complex Vector ALU and Vector permute block) to this:

http://appft1.uspto.gov/netacgi/nph..."International+Business+Machines"+AND+Hofstee



The PU being 64 bits makes the PowerPC 970 come to mind, but then again it sounds a bit wrong to assume that it is the PU...

Even if they were using that core, I'd think that it went under quite serious modifications that made a G5 based dev-kit non-reppresentative of the CELL architecture (in terms of performance, functionality, ISA, compiler and other tools, etc...) quite useless.


As I said, I think that a core based on the PowerPC 970FX (with various modification from IBM in critical and non critical areas) is way too much for a PU (at 2.2 GHz the PowerPC 970FX eats between25-39 Watts alone), but stretching things quite a bit I could almost see it (even though it sounds too much like a wet dream).

I do not see a VMX unit with registers increased from 32x128 bits to 128x128 bits being the APU.

I think that, if the conversation took place, that pairing a hypotetical VMX unit on steroid to the APU was the closest the IBM guy could give the "idea" of what the APU can do in terms of performance without disclosing its internal architecture.

 

[Fafalada]

I do not see a VMX unit with registers increased from 32x128 bits to 128x128 bits being the APU.

Well that part would kinda classify as on STEROIDS at least.
But before you say too much - remember there's another nexgen CPU out there which actually seems to have real VMXs on steroids in it... and I heard little birdies talking about 128worms in there too...

I think that, if the conversation took place, that pairing a hypotetical VMX unit on steroid to the APU was the closest the IBM guy could give the "idea" of what the APU can do in terms of performance without disclosing its internal architecture.

Well it's all about what target audience is familiar with I think If you talk to PS2 developers(or Deadmeat) they will tell you Cell is a PPC with 8VUs :lol

... I can't believe I just drew a paralel between DM and a PS2 dev ...
O_O

 

[Panajev2001a]

Well that part would kinda classify as on STEROIDS at least.
But before you say too much - remember there's another nexgen CPU out there which actually seems to have real VMXs on steroids in it... and I heard little birdies talking about 128worms in there too...

As I said:

Why ? PowerPC 970 instruction set... VMX instruction set (the VMX units on Xenon is likely "on steroids" as the "leaked specs" indicated)

Well, but that CPU (whatever it is I really do not know at all) while maybe sharing some similarities to CELL in "core" terms does not use APUs.

Can those real VMX units on steroids of that unnamed and unknown CPU access RAM directly through DMA without being spoon-fed ?

... I can't believe I just drew a paralel between DM and a PS2 dev ...
O_O

Bad Fafalada, bad !!!!

 

[Fafalada]

Can those real VMX units on steroids of that unnamed and unknown CPU access RAM directly through DMA without being spoon-fed

Oh I wasn't trying to compare S|APUs to VMX, I was just referring to your comment about number of reg... er... worms.

Bad Fafalada, bad !!!!

Ouch... I'll be good I promise.

 

[Panajev2001a]

Oh I wasn't trying to compare S|APUs to VMX, I was just referring to your comment about number of reg... er... worms.

Well, you kinda scared me there .

 

[deadhorse32]

ok guys could someone be a little bit clearer about those wormzzz, 970 pcc and nextgen CPU ?

Thanks for the answer (PM are welcome too).