Microsoft's Xbox 360, Sony's PS3 - A Hardware Discussion

PlayStation 3’s GPU: The NVIDIA RSX

We’ve mentioned countless times that the PlayStation 3 has the more PC-like GPU out of the two consoles we’re talking about here today, and after this week’s announcement, you now understand why.

The PlayStation 3’s RSX GPU shares the same “parent architecture” as the G70 (GeForce 7800 GTX), much in the same way that the GeForce 6600GT shares the same parent architecture as the GeForce 6800 Ultra.  Sony isn’t ready to unveil exactly what is different between the RSX and the G70, but based on what’s been introduced already, as well as our conversations with NVIDIA, we can gather a few items.

Despite the fact that the RSX comes from the same lineage as the G70, there are a number of changes to the core.  The biggest change is that RSX supports rendering to  both local and system memory, similar to NVIDIA’s Turbo Cache enabled GPUs.  Obviously rendering to/from local memory is going to be a lot lower latency than sending a request to the Cell’s memory controller, so much of the architecture of the GPU has to be changed in order to accommodate this higher latency access to memory.  Buffers and caches have to be made larger to keep the rendering pipelines full despite the higher latency memory access.  If the chip is properly designed to hide this latency, then there is generally no performance sacrifice, only an increase in chip size thanks to the use of larger buffers and caches. 

The RSX only has 60% of the local memory bandwidth of the G70, so in many cases it will most definitely have to share bandwidth with the CPU’s memory bus in order to achieve performance targets. 

There is one peculiarity that hasn’t exactly been resolved, and that is about transistor counts.  Both the G70 and the RSX share the same estimated transistor count, of approximately 300.4 million transistors.  The RSX is built on a 90nm process, so in theory NVIDIA would be able to pack more onto the die without increasing chip size at all - but if the transistor counts are identical, that points to more similarity between the two cores than NVIDIA has led us to believe.  So is the RSX nothing more than the G70?  It’s highly unlikely that the GPUs are identical, especially considering that the sheer addition of Turbo Cache to the part would drive up transistor counts quite a bit.  So how do we explain that the two GPUs are different, yet have the same transistor count and one is supposed to be more powerful than the other?  There are a few possible options.

First and foremost, you have to keep in mind that these are not exact transistor counts - they are estimates.  Transistor count is determined by looking at the number of gates in the design, and multiplying that number by the average number of transistors used per gate.  So the final transistor count won’t be exact, but it will be close enough to reality.  Remember that these chips are computer designed and produced, so it’s not like someone is counting each and every transistor by hand as they go into the chip. 

So it is possible that NVIDIA’s estimates are slightly off for the two GPUs, but at approximately 10 million transistors per pixel pipe, it doesn’t seem very likely that the RSX will feature more than the 24 pixel rendering pipelines of the GeForce 7800 GTX, yet NVIDIA claims it is more powerful than the GeForce 7800 GTX.  But how can that be?  There are a couple of options:

The most likely explanation is attributed to nothing more than clock speed.  Remember that the RSX, being built on a 90nm process, is supposed to be running at 550MHz - a 28% increase in core clock speed from the 110nm GeForce 7800 GTX.  The clock speed increase alone will account for a good boost in GPU performancewhich would make the RSX “more powerful” than the G70. 

There is one other possibility, one that is more far fetched but worth discussing nonetheless.  NVIDIA could offer a chip that featured the same transistor count as the desktop G70, but with significantly more power if the RSX features no vertex shader pipes and instead used that die space to add additional pixel shading hardware. 

Remember that the Cell host processor has an array of 7 SPEs that are very well suited for a number of non-branching tasks, including geometry processing.  Also keep in mind that current games favor creating realism through more pixel operations rather than creating more geometry, so GPUs aren’t very vertex shader bound these days.  Then, note that the RSX has a high bandwidth 35GB/s interface between the Cell processor and the GPU itself - definitely enough to place all vertex processing on the Cell processor itself, freeing up the RSX to exclusively handle pixel shader and ROP tasks.  If this is indeed the case, then the RSX could very well have more than 24 pipelines and still have a similar transistor count to the G70, but if it isn’t, then it is highly unlikely that we’d see a GPU that looked much different than the G70. 

The downside to the RSX using the Cell for all vertex processing is pretty significant.  Remember that the RSX only has a 22.4GB/s link to its local memory bandwidth, which is less than 60% of the memory bandwidth of the GeForce 7800 GTX.  In other words, it needs that additional memory bandwidth from the Cell’s memory controller to be able to handle more texture-bound games.  If a good portion of the 15GB/s downstream link from the Cell processor is used for bandwidth between the Cell’s SPEs and the RSX, the GPU will be texture bandwidth limited in some situations, especially at resolutions as high as 1080p. 

This option is much more far fetched of an explanation, but it is possible, only time will tell what the shipping configuration of the RSX will be.