How Sony's Development of the Cell Processor Benefited Microsoft

The Wall Street Journal is running an article about a recently released book entitled "The Race for a New Game Machine" which details Sony's development of the Cell processor, written by two of the engineers who worked on it. They also discuss how Sony's efforts to create a next-gen system backfired by directly helping Microsoft, one of their main competitors. Quoting: "Sony, Toshiba and IBM committed themselves to spending $400 million over five years to design the Cell, not counting the millions of dollars it would take to build two production facilities for making the chip itself. IBM provided the bulk of the manpower, with the design team headquartered at its Austin, Texas, offices. ... But a funny thing happened along the way: A new 'partner' entered the picture. In late 2002, Microsoft approached IBM about making the chip for Microsoft's rival game console, the (as yet unnamed) Xbox 360. In 2003, IBM's Adam Bennett showed Microsoft specs for the still-in-development Cell core. Microsoft was interested and contracted with IBM for their own chip, to be built around the core that IBM was still building with Sony. All three of the original partners had agreed that IBM would eventually sell the Cell to other clients. But it does not seem to have occurred to Sony that IBM would sell key parts of the Cell before it was complete and to Sony's primary videogame-console competitor. The result was that Sony's R&D money was spent creating a component for Microsoft to use against it."

Comments

It also helped MS

[Sycraft-fu]

Because it was a really misdirected effort when it came to a console. Sony really had no idea what the hell they were doing as far as making a chip for their console. Originally, they thought the Cell would be the graphics chip. Ya well turned out not to be near powerful enough for that, so late in the development cycle they went to nVidia to get a chip. Problem was, with the time frame they needed, they couldn't get it very well customized.

For example in a console, you normally want all the RAM shared between GPU and CPU. There's no reason to have them have separate RAM modules. The Xbox 360 does this, there's 512MB of RAM that is usable in general. The PS3 doesn't, it had 256MB for each CPU and GPU. Reason is that's how nVidia GPUs work in PCs and that's where it came from. nVidia didn't have the time to make them a custom one for the console, as ATi did for Microsoft. This leads to situations where the PS3 runs out of memory for textures and the 360 doesn't. It also means that the Cell can't fiddle with video RAM directly. It's power could perhaps be better used if it could directly do operations at full speed on data in VRAM but it can't.

So what they ended up with is a neat processor that is expensive, and not that useful. The SPEs that make up the bulk of the Cell's muscle are hard to use in games given the PS3's setup, and often you are waiting on the core to get data to and from them.

It's a neat processor, but a really bad idea for a video game console. Thus despite the cost and hype, the PS3 isn't able to outdo the 360 in terms of graphics (in some games it even falls behind).

I really don't know what the hell Sony was thinking with putting a brand new kind of processor in a console. I'm willing to bet in 10 years there are compilers and systems out there that make real good use of the Cell. However that does you no good with games today.

Thus we see the current situation of the PS3 having weak sales as compares to the 360 and Wii. It is high priced, with the idea that it brings the best performance, but that just doesn't bare out in reality.

Re:It also helped MS

[Sycraft-fu]

Specs on it I see show the system bus as being around 2GB/sec. That's comparable to PCIe (about the same as an 8x connection). While that's fine, it isn't really enough to do much in terms of back and forth operations. You'll find on a PC if you try that things get real slow. You need to send the data to the graphics card and have it work on its own RAM.

Now that isn't to say that you can't do things to the data before you send it, but then that's of limited use. What I'm talking about is doing things like, say, you write code that handles some dynamic lighting that the CPU does. So it goes in and modifies the texture data directly in VRAM. Well you can't really do that over a bus that slow. 2GB sounds like a lot but it is an order of magnitude below the speed that the VRAM works at. It is too slow to be doing the "read data, run math, write data, repeat a couple million times a frame" sort of thing that you'd be talking about.

You see the same sort of thing on a PC. While in theory PCIe lets you use system memory for your GPU transparently, in reality you take a massive hit if you do. The PCIe bus is just way too slow to keep up with higher resolution, high frame rate rendering.

So while it's fine in terms of the processor getting the data ready and sending it over to the GPU (which is what is done) it isn't a fast enough bus to have the SPEs act as additional shaders, which is what they'd probably be the most useful for.

Re:It also helped MS

[Zixx]

For example in a console, you normally want all the RAM shared between GPU and CPU. There's no reason to have them have separate RAM modules. The Xbox 360 does this, there's 512MB of RAM that is usable in general. The PS3 doesn't, it had 256MB for each CPU and GPU. Reason is that's how nVidia GPUs work in PCs and that's where it came from. nVidia didn't have the time to make them a custom one for the console, as ATi did for Microsoft. This leads to situations where the PS3 runs out of memory for textures and the 360 doesn't. It also means that the Cell can't fiddle with video RAM directly. It's power could perhaps be better used if it could directly do operations at full speed on data in VRAM but it can't.

Being a (former) PS3 and 360 dev, I have to say this is not true. Let's start with the memory split. Both consoles have about 20GB/s of memory bandwidth per memory system. Only the PS3 has two of them, giving it twice the memory bandwidth. The 360 compensates for that by having EDRAM attached to the GPU, which removes the ROP's share from your bandwidth budget. Especially with a lot of overdraw, the bandwidth needed by the ROPs can get huge (20GB/s, anyone?), so this would be a nice solution where it not for two things: the limited EDRAM-size and the costs of resolving from EDRAM to DRAM.
The RSX can also read and write both to XDR (main memory) and DDR (VRAM), giving it access to all of memory. The reason it is tighter on texture memory is because the OS is heavier.

About access to VRAM, it is true that reading from VRAM is something you don't want the Cell to do. It's a 14MB/s bus, so it is of no practical use for texture data. Writing into VRAM is actually pretty ok, as it's at 5GB/s, which is more or less achievable without trouble. At 60fps that's more than 80MB per frame.

In general, both design teams made sound decisions. The 360 has a significant ease-of-use advantage to PC developers with DirectX experience. The PS3 on the other hand is a lot more to-the-metal, but allows for some pretty crazy stuff. Sadly, most development these days is cross-platform, so you won't see a lot Cell-specific solutions. It's just not cost-effective.

Re:a few facts please?

[TheRaven64]

They are very different approaches. The 360's CPU is basically a 3-core, 6-context, in-order variant of the POWER4 with a vector unit. In terms of pure number crunching ability, it's pretty pathetic next to the Cell. On the other hand, it is based on a model that we have spent 30 years building compilers for. You only need to write slightly-parallel, conventional code to get close to 100% of the theoretical performance out of it.

In contrast, the Cell has one PPU which is roughly equivalent to one context on the 360's CPU (somewhere between 1/3 and 1/6 of the speed). It also has 7 SPUs. These are very fast, but they're basically small DSPs. They have very wide vector units and are limited to working on 256KB of data at a time. You can use them to implement accelerator kernels for certain classes of algorithm, but getting good performance out of them is hard.

In terms of on-paper performance, the Cell is a long way out in front, but it is a long way behind in ease of programming, meaning that you generally get a much smaller fraction of the maximum throughput.

Re:a few facts please?

[Michael Hunt]

Do you have even a vague understanding of what 'transform' and 'lighting' actually mean? Allow me to elucidate.

'transform' refers to the act of converting vertex positions in model space (the coordinate system used in the vertex buffers) to clip (camera) space. This is typically one matrix * vector multiplication per vertex; the vertex's position in model space is multiplied by the world-view-projection matrix. On modern hardware, this is generally done in the vertex program (other things may be done to the vertex's position before or after the co-ord transform, mind you, such as multiplication by a set of bone matrices for hardware animation, etc.)

'lighting' refers to the process of deciding the colour of each fragment ('potential pixel'). Before programmable graphics hardware, this was done by taking the dot product of the vertex normal and the light vector (or position, depending on light type), and multiplying it by the light's diffuse colour. The resulting colour intensity was then linearly interpolated across the face between vertices, and used to light the texture in conjunction with the ambient term. With modern programmable hardware, lighting is usually done per-fragment based on a normal map, which is input as a second texture to the fragment program. The light position is converted from object (or world) space into 'tangent' space, which is a coordinate system whose basis vectors are parallel and orthogonal to the plane being lit, and the surface is lit based on the dot product of the light vector in tangent space and the normal from the normal map.

Back in the bad old days, when men with beards owned IRIX boxes and everybody else had a TNT2 or worse, transform and lighting were done in software for most folks, by a client of the rendering system, before the primitives were submitted as draw calls to the rendering system. Post-about-2001, cards with hardware T&L, such as Geforce 256, showed up in the PC space. These cards were the first consumer 3D hardware to perform fixed-function transform and lighting (roughly as I've described it above) in silicon. The API didn't change much, although there was a DirectX version bump (6 to 7). OpenGL programmers didn't really notice; the library itself, obviously, had to know if it was talking to a fixed-function card or a dumb card, but most OpenGLs were provided by hardware vendors in any event, so this wasn't a factor.

Fast forward to today, everybody's using hardware which allows parts (most, these days) of the rendering pipeline to be replaced entirely with programs written by the engine developer (or even the artist, in some cases.) Transforming vertices can be done in conjunction with all manner of other crap, and lighting can be done using whatever model the programmer/artist desire. Regardless, however, it's all done in the same pipeline on the GPU. If the SPUs, as you suggest, were pre-transforming and pre-lighting vertices before writing them to 3d hardware's vertex buffers, then all you'd get is some really confused 3d hardware. RSX (the 3D chip loosely based on nvidia's G70 architecture) has 8 vertex pipelines and 24 fragment pipelines, all programmable. This is more than enough power to do significantly more with each vertex than simple transformation, and enough power to perform even complex effects such as steep parallax-mapped lighting in the fragment pipeline.

In conclusion, while Xenos (360's GPU) may or may not be better than RSX, RSX is CERTAINLY more than powerful enough to handle its own T&L. Cell's SPEs are, at least on some level, a compromise between the massively data-parallel yet somewhat braindead pipelines of a GPU and the more-or-less serial yet significantly intelligent threads of a modern CPU. They'd be great for accelerating physics (Bullet, i believe, has a Cell backend) or AI, but really add fuck all to the 3D rendering side of the console.