Wednesday, March 05, 2008

PlayStation 3 Cell Came Before Its Time

The Microsoft Xbox 360 platform is now profitable and Sony's PlayStation 3 is gaining parity with the Xbox 360 in the marketplace. We're about halfway through this console cycle (these cycles typically last about five years) and we can reflect on technology choices with the benefit of some hindsight. The ramifications of some of the more obvious choices, for example PS3's Blu-Ray vs. 360's plain-old-DVD, are well known. Microsoft's background as a operating system and compiler developer provides the 360 with perhaps the best development tools ever to grace a non-desktop system. We'll look at that in another entry. We'll also want to compare the Xbox 360 GPU and the PlayStation 3 GPU.

For now, however, I'd like to look at the Cell. In doing so, I'd like to ask and try to answer the question, "Why has the quality of PlayStation 3 titles lagged behind that of the Xbox 360 titles?"

The Cell Processor is Amazing

From the info I gathered, the PS3's Cell should have a peak numerical performance of about 205 gigaflops. From Wikipedia, I gather that the 360, on the other hand, has peak performance of 116 gigaflops. The PS3's Cell has one regular PowerPC core (with two hardware threads) called the PPU. It also has seven SPUs (one is reserved for the OS), which are streamlined vector processors that are fast as hell, although functionally limited compared to the main PowerPC core. They have their own memory at 256K apiece. Half of that is available to the developer. The 360's Xenon processor, on the other hand, consists of three conventional PowerPC cores almost identical to the Cell's PowerPC core, for a total of six hardware threads. So, if you're like most people, you're probably thinking, "Wow, the PS3 blows the 360 away!"

Except that it doesn't.

Developing Games is Hard

I just read this at .mischief.mayhem.soap about the size of the Assassin's Creed team. Actually, I think that's about the same for Madden NFL. Modern games require an enormous amount of content, some of which must be reflected in code. Games on multicore consoles are much more difficult to debug. Making games is simply much more expensive that it used to be. I heard a colleague remark that he saw that one PlayStation 1 version of Madden had 30 files. Now, it's thousands of files, millions of lines and scores of developers.

Developing for the Cell is Hard

When we first saw the specs for the Cell, I thought, "it looks like it was designed to play movies." And in playing movies, the PS3 will blow the 360 away. The Cell was designed for performing the same kind of vector calculation on large numbers of small chunks of data. Essentially like a GPU, but more flexible although less parallel. There's a lot of this sort of thing in games. Some physics, audio decoding/encoding, movie decoding/encoding, some graphics/geometry processing. Still, there are a lot of things that aren't well suited for this model. Among those would be language interpretation (we do a lot), gameplay code, sparse data manipulation, and a lot of really mundane and common software engineering things that are inherently sequential. Sorting a list of pointers that must be dereferenced to be compared, for example.

Explanations of why the 360 performs so well compared to the PS3 usually start out like, "Well, the Cell looks better on paper, but.." Developing for a 360 is like developing for a multicore PC. Take your existing code, multithread it and know that it'll work on the 360 and upcoming PC-like multicore machines with little changes.

You can build PS3 software exactly the same way using only the PPU and its two threads, but that will yield almost exactly one-third of the power of the 360's processor. In fact some early titles that were developed first for the 360 did just this. This was not as bad as it sounds though. A lot of early code for these machines was not thoroughly multithreaded because much of that code was directly ported from older Xbox and PS2 code. In these cases, it used just a portion of the 360's power anyway.

Now we're pushing the capabilities of the 360 and trying hard to get the PS3's six fast but special-purpose SPUs to do what we can do with the four remaining general-purpose hardware threads on the 360. It's hard, though certainly possible to achieve parity. Still, it requires a lot more code and expertise just for the PS3 (usually from our best developers), which typically means much less overall developer productivity.

The Cell Looks Like the Future

Unfortunately, as delightful as it is to develop for, the approach taken with the 360's Xenon processor simply can't scale. At some point, with some number of traditional cores, memory access and synchronization are going to break down or become so contentious that any benefit of a larger number of cores will be nil. So the Cell looks like the future. A small number of traditional cores with a large number of stream processors can yield phenomenal performance in the right proportions. I predict that some member(s) of the next console generation will have a processor with 8 traditional cores and perhaps 256+ stream processors. 256 sound ridiculous? Consider that an ATI HD 2900 XT GPU has 320 stream processors.

Why do I think the Cell was premature? It did not fulfill the possibilities of traditional multicore processor before moving on to a hard-to-develop-for, though potentially more powerful, architecture.

In any event, it's clearly a foreshadow of processor architecture to come. Still, the difficulty of achieving parity between the Cell and 360 Xenon when game development is already at its most taxing really isn't doing the PlayStation 3 any favors.


DEADC0DE said...

I totally agree with you. First time I saw the cell I thought, that thing is perfect for what it was designed for, scientifical computations, small kernels on huge amounts of non-sparse data, but in games we don't do that often. Overall the PS3 seems to be flawed to me, it's a LOT less balanced than the 360, probabily it could be more powerful, but it's hard to use it fully. Also, the 360 GPU is way nicer. We are not ready yet for massive parallelism. It's something that it's really in its infancy for our industry. And while it _could_ be true that the Cell is going to be the future, I'm not so sure about it. It's hard to make predictions for me, it will depend also on the future of the GPUs.

NarcolepticNerd said...

The ps3 OS uses 1.5 SPU's. It uses one spu for small amounts of time, but very often during the frame.

Don't get me started on the requirements of Sony's OS.

Joshua Smith said...

Thank you for the nice tips. It was easy to read, but I'd like to add that if your software company needs to be updated try software development service.

Joshua Smith said...

Thanks for mentioning this great info. For those who is searching for new cars, check auto quotes from top auto insurance companies.

Joshua Smith said...

We appreciate your help. Let me tell you something about low cost home insurance to save your money on house policy.