IBM: Cell chip in PS3 underutilized

If it's High Moon, then I reckon it's likely to be Mike Acton doing most of the talking. He of www.cellperformance.com fame..

Dean

 

Well Mike certainly springs immediately to mind when I hear High Moon, but I don't think it'll be him speaking; Mike went over to Insomniac to head up their engine development team in December. I'm going to try getting in touch with him over this Vivendi event actually, but it shows what he was talking about in that interview before - Clinton Keith (High Moon CTO) really is dedicated to developing around Cell it seems. Interesting to see that sort of focused effort from a third-party, but to quote Mike:

CB: Shifting gears back to your place at High Moon Studios, it seems surprising that a multi-console developer like High Moon would set up an R&D group focused so specifically on drawing out the power of the Cell and the PS3 in the first place. Are such internal research teams common in the industry as far as multi-console development houses go, and if not, what led to its creation within High Moon?

Mike Acton: Clinton Keith, CTO of High Moon Studios, is an established evangelist for raising quality of life in video game development by using better approaches. About 18 months ago, Clint recognized that it's not just our development practices that need to change, but also the core strategy for dealing with dynamic technologies. What we needed was not just another ad hoc engine team, but a center of competency… a unique department dedicated to understanding the platform and discovering, or perhaps rediscovering in some cases the best practices to apply to performance-critical game code. From that, we started our PS3 research team. It is with Vivendi's ongoing support that we have continued to build on our base of knowledge and expertise right up to today.

Our goal at High Moon is to bring the overall experience of our games to the players that expect a best-in-class product regardless of their chosen platform. Our production team wants to deliver nothing less. In order to produce AAA product on multiple platforms, the trade-offs for each platform must be well-understood. The most straightforward way of accomplishing this is to create a small team dedicated to the platform. In the industry as a whole it isn't uncommon to find individuals who are relied upon to be platform experts. Here, our CTO Clinton Keith simply codified this practice for us.

 

Here is some more on the topic.

"eight years down the road", Huh? Some perspective there.

 

I love their vision. Thats what I wanted to see in next gen.

Also thats the major reason that I was so excited by the E3 2005 target renders and had so much hope. It wasnt much the visual detail, but the flow and interactivity of everything in them. That was the most unbelievable part in these videos. If they kept the vidual detail the same but didnt have the physics, AI and interactivity I wouldnt have been as much as impressed.

Unfortunately though they were PR work most likely, and neither 360 or PS3 have shown the level of interactivity I was hoping for. Although I must say that PS3 seems to be slightly closer towards it from what I see in games like Motorstorm, Heavenly Sword and MGS4

 

I'm not sure what you mean by that. You understand that's the lifespan of the platform he's talking about? For actually making useful performance of the SPEs, quote the next question :

Already they're creating unique content, which'll likely show up as a mini-game, which is one key aspect of the eDI. These development strategies enable returns to be made on RnD. Rather than wait for the release of a major title to get returns on the water physics engine you've been exploring, turn that engine into a mini-game.

Realizing the potential of Cell is a progressive thing that will improve over time, like always. The best of PS2 has only happened 6-7 years in. That's what you expect in any platform with lots of room to grow. A platform that's tapped in 2 years is either bottlenecked by APIs or hadn't anything interesting to work with. The 8 year figure is that they'll still be finding out how to do things in 8 year's time, and not that it'll take 8 years to figure out Cell to any useful amount.

 

Did you catch the "remainder of the decade" comment? I had to read that sentence like three times, before posting, WTF? There is only three more years left in this decade, 8 more years puts you halfway through the next decade. So what is he saying exactly?

 

Seems like jumbled speak to me. 8 years to the end of the decade makes no sense. I think he meant 8 years of development progress in a 10 year lifecycle.

 

Yeah, I actually did, but I assumed he was refering to the PS3s ten year life-cycle as a decade, strange I know. Maybe he thinks the PS3 is a start of a new era and people will have a new time count and refering to the time before and after the PS3.

BTW I thought this sounded fun:

I´d love to have that freedom in my work.

Edit: Shifty beat me.

 

I thought it was remarkable that he predicts they will keep improving the use of Cell for 8 years more, that was my interpretation. Sony has admitted the PS2 is currently running out of steam, so if what he says is true the PS3 will be honking on a few years more. Well we'll see about that.

 

You do know that a decade is defined purely as an arbitrary 10 year duration and not constrained specifically to any starting point..?

http://en.wikipedia.org/wiki/Decade

 

This is where eDI is truly great, and I think rewards the super-geeks. For all their years creating technologies, now they have a chance to create some games as well. eDI's become the coder's playground, and all those bizarre experiments can be floated to see what the reaction is.

It looks like if you want to play too, you have to be on the cutting edge of software development. Perhaps many from the demo-scene could find there way to producing technological marvels as software downloads? I recall Stardust on Amiga that had some very nice demo effects. Heck, even demo downloads would be awesome! A console that opens up the hardware to demos will get lots of free advertising!

 

Eh, no.

My interpretation was restricted to this part:

 

Well as much as it may have been.. My (and more specifically the usage of the term in the article posted) interpretation still hold.. Therefore your initial comment:

Is moot..

 

It can if they strip off some SPEs.
Cell may turn out to be a bit over designed for the task, but the stream architecture is definitely a design win over DSPs, properly implemented that should become blatantly obvious. At the same power consumption level and a slightly higher transistor budget, a stream processor should beat out a DSP every time. From my understanding of what a stream processor is, it's basically a collection of one or more DSP like processors with some additional trafficing logic implemented, it shouldn't lose to a DSP, and under the right circumstances should beat a traditional general purpose processor, though its potential design wins are much more limited if the dataflow requires a lot of management. I believe a stream processor should win in a performance per watt comparison, though it may lose in cost.

 

I'm curious - what are you basing that on? I would be under the impression the SPEs would be decent coprocessors to a more traditional consumer electronics CPU (such as an ARM), but I am very skeptical the PPE makes sense in those markets. This doesn't reduce the necessity of fixed-function blocks for specific workloads (such as all or part of media/video processing), however.

 

I was insinuating that cell, once its size is reduced, could be quite a competent processor against the higher end embedded chips. The PPE doesn't take up that much die space, and mass production of Cells across many devices should make them more affordable.
Something similar has already happened with other IBM microprocessors, and the markets were G3 based processors are currently seen I'd expect Cell processors to take over first. Blu-ray and HDDVD players would also be prime targets for Cell processors, as well as high end scalers.

There certainly have been situations in the past where more general purpose processors have supplanted DSPs almost completely, and even a single PPE + SPE offers quite a bit of FPU crunching ability.
Beyond that, per die size (and eventually cost once mass production is reached) Stream architectures are a large conceptual win over DSPs. An 8 core Cell is definetely overengineered for just about any market, but a 1 PPE-SPE mix may just have the right performance and cost for higher end consumer electronics. Who knows, with mass production and if Moore's law still holds, maybe it could see its way into just about every device under the sun.

BTW, Wikipedia states Cell's die size is 235mm^2, the PPE core makes up 11.1% of that, and each SPE makes up 6.2%. Assuming just 7 of the SPEs are stripped off, that's a die size of 133mm^2. Die size should probably be reduced to a bit lower as well, since the interconnections between all those SPEs won't be necessary. Besides that, maybe some L2 cache could be stripped off. In the end, the die size probably would rest around 100mm^2, which is really a bit large for consumer electronics, I can't think of many devices using something of that size. I'm not sure even high end consumer electronics that already use more general purpose processors use processors that big/expensive, and Sony is losing their window of opportunity to shove Cells into expensive Blu-ray players.

Yeah, I'd say looking at those die size numbers that even a stripped down Cell/SPE config may still be too big, even if it would offer excellent performance.
Oh, and wikipedia states that IBM has listed the Cell processors with 1 PPE and 1 SPE (SPU?) as only utilizing 11W of power at 5ghz, and 2 at 3ghz. I'd say at least the power consumption ratios would be more than competitive with any performance competing processor, Cell just needs to overcome the problem that there's very little that needs even a stripped down Cell's performance.

 

And what other markets uses G3 processors?

 

Well, not G3s per se but a similar IBM processor (from the PowerPC 440 or 460 line) is used in high end routers and some high end consumer electronics, though even a 1 SPE cell is probably an order of magnitude more expensive (both PPC 440 and 460 are less than 1/10th the die area of a 1 SPE cell). Ok, based on that I'd say Cell is really only a natural fit for high end scalers and video processors (to go in something like an HD Tivo, blu ray players).

 

IBM has a presentation at Cool Chips X in the next month.
http://www.coolchips.org/cool10/advance1.html

 

A 65nm SPE 1PFLOPS supercomputer? What's that supposed to mean? Link doesn't work btw.