Slashdot Mirror
IBM's Plans For the Cell Processor
angry tapir writes "Development around the original Cell processor hasn't stalled, and IBM will continue to develop chips and supply hardware for future gaming consoles, a company executive said. IBM is working with gaming machine vendors including Nintendo and Sony, said Jai Menon, CTO of IBM's Systems and Technology Group, during an interview Thursday. 'We want to stay in the business, we intend to stay in the business,' he said. IBM confirmed in a statement that it continues to manufacture the Cell processor for use by Sony in its PlayStation 3. IBM also will continue to invest in Cell as part of its hybrid and multicore chip strategy, Menon said."
What business would want to give up guaranteed sales? I mean, a gaming platform is like walking into a bank, depositing one cent and then getting a cent every second until the bank closes.
Restore the madness of youth's lechery
Bring on a 12 core PS4 with raytracing games.
The basic problem with the Cell processor is that it has 256KB (not MB, KB) per processor, plus a bulk transfer mechanism to main memory. Given that model, it has to be programmed like a DSP - very little state, processing works on data streams. For games, this sucks. No CPU has enough memory for a full frame, or for the geometry, or a level map. Trying to hammer programs into that model is painful. (Except for audio. It's great for audio.) In many PS3 games, the main MIPS machine is doing most of the work, with the Cell CPUs handling audio, networking, and I/O. And, of course, Sony had to put an NVidia graphics processor in the thing late in the development cycle, once people finally realized that the Cell CPUs couldn't handle the rendering.
But if each Cell CPU had, say, 16MB, the Cell machines could be treated more like a cluster. Programming for clusters is well understood, and not too tough.
It's probably too late, though. Multi-core shared memory cache-consistent machines are now too good. It's not necessary to use an architecture as painful as the Cell. It's probably destined for the graveyard of weird architectures, along with data flow machines, hypercubes, SIMD machines, systolic processors, semi-shared-memory multiprocessors, and similar hardware that's straightforward to build but tough to program.
The story is not that IBM continues to manufacture chips but that the Cell design is not dead. This contradicts earlier stories to some degree.
In all fairness, it contradicts only on the surface as IBM only stated in the older story that Cell as separate design will end and its co-processor-heavy design will merge with future POWER iterations.
There were also rumors that IBM won't manufacture PS3 Cell CPUs any longer, leaving it to contractors.
A while back I was looking for one or two Cell CPU based machines as development boxes for inhouse geophysical software - basicly to see if it's worth going onto that platform. The three week process between contacting what appeared to be the only vendor of Cell based workstations and getting a price for an entry level machine was frustrating. It involved daily calls to a slimy bastard that appeared to just want to waste time trying to become my friend until he had carefully finished weighing my companies wallet.
In the end the time window had come and gone (the developers got bored or gave up on the idea of using the Cell) before I could get even a hint at the price but I kept going for the sake of future projects. The price for one workstation with one processor was fairly similar to that of six of our cluster nodes. You would need some sort of black-ops budget where any Accountants coming close are shot on sight before paying that sort of price. An entry point machine no much different to a playstation with more memory cost a truly insane and unjustifiable price.
You've really missed hearing about Cell?
It's a new processor architecture, IBM and Sony (and possibly others) had a hand in it. Effectively two "Power" cores and a bunch of vector processing units. It's supposed to be very very good for vector operations. For a while (a few years back now) the world's most powerful supercomputer was a machine composed of nodes containing two cell processors and an Opteron each.
It's different to other parallelisation strategies as the vector units (SPU/SPEs) allow you to parallelise stuff at an operation level, unlike just stuffing more cores into the box which is the intel/PC strategy. For games and graphics this it thought to be good, hence its inclusion in the playstation 3. It's also supposed to be good for scientific computing.
I guess you could think of it as somewhere between a CPU and a GPU, or a hybrid of the two approaches.
I have been wondering just how long it will take for the "ooohhh shiny!" factor to wear thin. Hell I fire up Far Cry I or Wolfenstein on my $36 HD4650 and the people stand around and go "oooohhh". You really don't need any higher to have decent immersion in a game, and especially with FPS if the game is worth a damn you are too busy dodging fire to just stand around and look at the shiny. Then add in the spiraling costs and delays to market adding lots of "ohhh shiny" add, and it quickly becomes "get a hit, and on time, or we'll all out of business" and that simply isn't sustainable long term.
That is why I wouldn't be surprised if the next gen gaming consoles don't do something similar to the original Xbox, which I thought was a damned good idea at the time. You could take a cheap ULV Phenom II Quad, add a 5xxx Radeon GPU and some decent controllers and have the average Joe drooling at the "ooohhh shiny" for a long time, and the combination of cheap hardware, the ability for developers to easily code with tools they already have, and the quick time to market would probably make it a hit.
I just don't see the incredible amounts required to bring a new gen of consoles not seriously hurting any companies bottom line. With a more off the shelf approach all they have to do is cook up the DRM and a close to bare metal OS for it and let the economies of scale keep the price low out the gate and drive prices even lower as time goes on. While MSFT could blow the cash simply because they have twin cash cows in Office and Windows, I doubt Sony will be able to afford the needed capital, and Nintendo has made it pretty clear they aren't gonna play the "ooohhh shiny!" game at all with targeting the Wii to casual gamers. I just don't see a never ending ooohhh shiny arms race being good for anybody. Just look at how ATI is using Eyefinity to push new GPUs and Nvidia looking at HPCs with CUDA, even they know the "ooohhh shiny" can only go so far. Hell I figured when I got the HD4650 it would just be a stopgap until I could get a $150+ GPU, but now? Hell it plays Bioshock II and everything else I throw at it with plenty of ooohh shiny and doesn't turn my apt into a sauna bath, so why bother? I used to be a serious graphics whore, but even I got tired of the ooohh shiny and now prefer games that are actually...what's the word?...oh yeah FUN. I'm starting to wonder if the whole graphics race is starting to hit a dead end.
ACs don't waste your time replying, your posts are never seen by me.
I found this article interesting. They write about Valves approach to multi-core CPU's and game engines.
The programmers at Valve considered three different models to solve their problem. The first was called "coarse threading" and was the easiest to implement. Many companies are already using coarse threading to improve their games for multiple core systems. The idea is to put whole subsystems on separate cores; for example, graphics rendering on one, AI on another, sound on a third, and so on. The problem with this approach is that some subsystems are less demanding on CPU time than others. Giving sound, for example, a whole core to itself would often leave up to 80 percent of that core sitting unused.
The second approach was fine-grained threading, which separates tasks into many discrete elements and then distributes them among as many cores as are available. For example, a loop that updates the position of 1,000 objects based on their velocity can be divided among, say, four cores, with each core handling 250 objects apiece. The drawback with this approach is that not all tasks divide neatly into discrete components that can operate independently. Also, if some entries in the list take longer to update than others, it becomes harder to scale the tasks evenly across multiple cores. Finally, the issue of memory bandwidth quickly becomes a limitation with this method. For certain specialized tasks, such as compiling, fine-grained threading works really well. Valve has already implemented a system whereby every computer in their offices automatically acts as a compiler node. When the programmers were getting ready to demonstrate their results on the conference room computer with the big screen, they had to quickly deactivate this feature first!
The approach that Valve finally chose was a combination of the coarse and fine-grained, with some extra enhancements thrown in. Some systems were split on multiple cores using coarse threading. Other tasks, such as VVIS (the calculations of what objects are visible to the player from their point of view) were split up using fine-grained threading. Lastly, whenever part of a core is idle, work that can be precalculated without lagging or adversely affecting the game experience (such as AI calculations or pathfinding) was queued up to be delivered to the game engine later.
Valve's approach was the most difficult of all possible methods for utilizing multiple cores, but if they could pull it off, it would deliver the maximum possible benefits on systems like Intel's new quad-core Kentsfield chips.
To deliver this hybrid threading platform, Valve made use of expert programmers like Tom Leonard, who was writing multithreaded code as early as 1991 when he worked on C++ development tools for companies like Zortech and Symantec. Tom walked us through the thought process behind Valve's new threading model.
http://arstechnica.com/gaming/news/2006/11/valve-multicore.ars
Teasing the nobles, and rightfully so!