Kutaragi Ken:Want a teraflop? You have to buy a rack from us

[DeadmeatGA]

http://www.theregister.co.uk/content/3/33791.html

Sony Cell CPU to deliver two teraflops in 64-core config
By Tony Smith
Posted: 05/11/2003 at 12:10 GMT

'Cell', the massively parallel processing chip currently being designed by Sony and IBM, will scale from single-chip systems through to entire server rooms packed with thousands of them, Sony's executive deputy president Ken Kutaragi told attendees of the company's Transformation 60 conference yesterday.

That's always been the goal, of course, since Cell was first announced back in March 2001. But yesterday Kutaragi put some numbers onto the chart.

A four-core chip home server system will be able to deliver one billion floating-point operations per second, apparently. Move up to a 32-core chip - in, say, a blade server module - and you'd get 32 gigaflops of processing power, while a 64-core slab of silicon inside a rack-mount unit doing graphics work would churn out two teraflops, according to Kutaragi's presentation foils.

Ultimately, Kutaragi suggested, we'll see 16 teraflop supercomputer 'cabinets' and one petaflop (a million billion flops, in other words) server rooms - the latter delivering enough raw power for true AI systems, he said.

Kutaragi likened a single Cell chip to IBM's 32-node RS/6000-based chess supercomputer Deep Blue. The exponential scaling rate suggests Cell really doesn't come into its own until you use lots of them together.

That's certainly the design philosophy: "With built-in broadband connectivity, microprocessors that currently exist as individual islands will be more closely linked, making a network of systems act more as one, unified 'supersystem'. Just as biological cells in the body unite to form complete physical systems, Cell-based electronic products of all types will form the building blocks of larger systems," was how Kutaragi described the Cell concept back in 2001.

Since it takes 2200 PowerPC 970 chips - aka the G5 - to yield just over ten teraflops - much the same as you get from 2000 Athlon 64s - getting similar performance out of just 64 Cell cores is impressive, if Sony and co. can deliver.

Right now, they're just a little way past half way through the five-year Cell research project, so they have a few more years yet to demonstrate the device in action. We'd expect the successors to today's chips to have got a little closer to those kinds of figures by 2006, but not that close.

Others are not so far behind. ClearSpeed's recently announced CS301 chip, for example, can deliver 25 gigaflops peak, the company claims. The CS301 is a 64-way parallel processing co-processor designed to work alongside an x86 or other general purpose CPU. The downside is that it's expected to cost over $16,000 per chip. We suspect Sony and IBM are aiming for something a little more mass-market. ®

You heard what the man said; you want a teraflop? You have to buy a CELLRack(TM), not a chip, from SCEI.... We can finally put the "Teraflop" chip non-sense behind us and discuss something productive instead...

I was proven right once again.

 

[notAFanB]

I guess that pretty much settles it.

btw where can I find this article DM, I like to lookup the presentation if I could.

 

[chaphack]

Right now, they're just a little way past half way through the five-year Cell research project, so they have a few more years yet to demonstrate the device in action. We'd expect the successors to today's chips to have got a little closer to those kinds of figures by 2006, but not that close.

2006? :?
Which figures are they talking about? 10TFLOPS or 2TFLOPS?

 

[London Geezer]

huh???

there must be something wrong here.

a 4 core chip delivers 1Gflop??? thats LESS than the EE in PS2....

missing something?

 

[DeadmeatGA]

...

I am not sure whom we should blame for this "Teraflop chip" non-sense that has haunted us for two years, the mistranslation of what Kutaragi said or Kutaragi Ken himself for giving a mispresentation to press; when he said a Teraflop CELL, he should have clarified it as a mega-expensive MCM rack, not as a single chip inexpensive enough to go into something like PSX3.

Since a MCM rack is not going into a $399 device like PSX3, that makes it official; PSX3 won't be doing a teraflop, but more like 100~200 GFLOPS in the best case scenario.

 

[London Geezer]

mmmmmmm nope that still doesnt explain how a 4 core ship can only do 1Gflop....

 

[DeadmeatGA]

...

nope that still doesnt explain how a 4 core ship can only do 1Gflop....

Ask Kutaragi san, that's what the man showed in his pdf. And the scary thing is, Kutaragi has been showing the same roadmap over and over for past two years.

 

[cybamerc]

Sony's chart is very odd. Obviously, the home server version (which is incidentally what PSX is labelled as) is less powerful than even the EE in the PS2/PSX. While I'm certainly sceptical when it comes to Cell and Cell applications I wouldn't read too much into that chart as it simply makes no sense on its own.

 

[chaphack]

Yes, it is very strange...it is the same thing for the past few years...in which is the same from IBM blue gene chart...

Is that a SERIOUS oversight or what...? whatta lack of professionalism or something..? confusion? lack of true Cell understanding by Business man Mr KK? :?

 

[Joe DeFuria]

mmmmmmm nope that still doesnt explain how a 4 core ship can only do 1Gflop....

It doesn't have to.

The "thing" about cell is it's scalability. There's no reason why a 4 core chip shuold be more or less powerful than past silicon of a completely different architecture that was not meant for massive parallel scaling.

In fact, if Sony wants the Cell core to scale downward into really inexpecive devices, then a "single core" cell would very VERY small (cheap).

Cell gets its power from slapping on multiple cores together, not from one particular core / chip being oh-so-powerful.

Personally, I just pretty much IGNORE any talk about teraflops, cores, chips etc, and specific applications like PS3. It's all about as clear as mud right now, and Sony doesn't make anything clearer with their own statements. I'm willing to bet they don't even really have a good idea how things will settle out in terms of numbers.

 

[London Geezer]

well according to the venerable Deadmeat, the PS3 Cell will have 2 cores. that would output 500Mflops.
sorry guys but that is

Dreamcast could push more floating point operations per second...


yeah i guess we'll have to wait until the fog settles...

 

[DeadmeatGA]

...

mmmmmmm nope that still doesnt explain how a 4 core ship can only do 1Gflop....

When you have a very simple scalar processor clocked at 500 Mhz.

Purhaps there are two versions of CELLs? The "general" CELL for the entire Sony group and the "custom" CELL for the PSX3 with lots of custom APUs not shared by other devices? This explains why Kutaragi Ken classified CELL as a "network" processor and why Sony Electronics needs a dedicated "media" processor as the core of future consumer electronics.

Under this platform strategy, CELL is nothing more than a glorified ethernet chipset that automatically distributes data packet evenly among themselves. Each CELL would then have its own attached processing units(APUs, GS3) and process data in its own unique way.

 

[notAFanB]

well if the DM post is anything to go by I can't see how the calims of leapfroggin current industry leader is gonna happen. bar putting an liquid cooling unit in there and cranking up the clock.

 

[DeadmeatGA]

...

This "Teraflop" processor misunderstanding happened because Kutragi's definition of a "processor" is different from our accepted standard definition. To Kutaragi, a MCM rack module is still considered a processor, whereas the standard computer science terminology does not accept such definition.

CELL really is exactly the opposite of what we thought it was; it is not some 450 mm2 mega processor chip burning 200 W of power to attain 1 teraflop per chip, it is rather a collection of very low-power small processors small and cheap enough to use whole lot of them to construct computing modules. If you were to contruct a rack of CELL chips, then each CELL chips has to burn very little power, or else you cannot construct one due to power consumption and heat problems.

 

[chaphack]

If anything, i hope to see 1TFLOP(of TXB famed) reaction when PS3 aint no one'er!

 

[London Geezer]

mmmmmmmm..... doesn't convince me.........
something is missing, and i think it's the APUs

i mean, attaching the APU's to the cells would be the whole point to get the high floating point performance... in the circumstances, what would have been the point of this thread?

oh wait u don't have to answer that...

 

[marconelly!]

I was proven right once again.

Oh man... IF you are proven right, that would be one of the VERY few times you've been proven right )

Anyways, I think DM with his last two posts may be onto something. I don't know how else to understand that the NEW 'home server' made in 2005/6 will have 6x less flops than TODAY's home server (PSX). Perhaps cell mentioned in this presentation is not the 'main CPU' cell after all. That would also explain how can a 32 core chip deliver 32 gigaflops, and chip with just two times more cores can deliver hundred times more performance. Perhaps 'cpu cells' are those where each core has their own set of APUs that add to the full number of flops.

 

[London Geezer]

ok so let me get this straight.....

Deadmeat was bragging about how he KNEW that the PS3 version of Cell would have 2 cores.

no he's saying that a 4 core 500MHz Cell chip would output 1Gflop (that is 1/6th of the theoretical raw output of the present EE).

even at 4GHz (which aint gonna happen IMO) that would give 4 Gflops. that is still 30% slower than the EE.

let's say that we attach and APU for each Cell core (not really getting this but i digress). even to get to the 200Gflop rating Deadmeat is giving, each APU would have to output 50Gflop EACH on its own.

now, Deadmeat, r u going to explain this and entertain all of us?

 

[Clashman]

I could be wrong, but I think he may have been saying two CHIPS, not cores. Which would put output, (I think), at 64 GFlops.

 

[Megadrive1988]

forget FLOPs forget TeraFLOPS


Playstation pushed nicely faked polygons around with texture maps, gouraud shading, lighting (100k - 180k polys/sec, no z-buffer)

Playstation2 pushes lots of true 3D polygons (z-buffered) with a few effects like filtering, alpha blending, etc (upto about 20M pps)

Playstation3 should push gobs of geometry around with some new rendering techniques like pixel shaders, procedural textures, perpixel & sub pixel lighting.
(probably 1B pps with lighting, shaders, effects, in games)