Interview: AMD on Intel, parallelism, DirectX 10.1 and more

Atomic: When hybrid graphics are released, do you have any plans to let the integrated graphics card do some GPGPU work, like physics?

David: I’m saying not yet, because the IGP doesn’t have enough power to be a dedicated GPGPU at this point -- but there’s no reason why it can’t be done in the future when it’s more powerful .

It’s a research area. There are so many things that can be done with GPGPU. Our initial focus and research in that area was on intensive, high performance processing tasks like modelling and simulation for oil companies or financial analysis.

That doesn’t mean that you can’t also use GPGPU for more consumer oriented tasks, it’s just an area of future research, and until we nail down the algorithms for that sort of thing it’s hard to know how much processing power you’re going to need to make it interesting.

So while our IGP chips support the full feature set of our discrete graphics, for GPGPU to be interesting you have to able to exceed the performance of a CPU by a significant margin, otherwise you might as well just do it on the CPU.


Atomic: How close to a CPU is an IGP in terms of raw processing power?

David: Our IGP can do about 20 Gigaflops, which is probably in the realm of single, possibly dual core CPUs. But that’s just raw throughput, you don’t always get your peak performance when you try to do GPGPU stuff. I think the theoretical processing power of a 3GHz quad core is about 96 gigaflops, assuming SIMD is enabled (or else the CPU would be very slow).


Atomic: What kind of discrete graphics card would be powerful enough to dedicate itself to physics calculations through GPGPU?

David: Physics is a very broad problem. If you look at how physics is done today, the only things GPGPU is being used to compute physics for are very simple special effect physics like the foliage moving back and forth in Crysis. It doesn’t affect the gameplay in any way, but it makes things look a little better and can add more objects into the scene and calculate their movement.

What the GPU is not currently used for is calculating how a character animates and responds to its environment; things like collision detection and line of sight. Currently it’s done on a CPU, but if you look at just how these algorithms work, you could theoretically say that certain types of physics would actually run well on a GPU. I’m not saying that they are currently today.

I’ll give you an example. This isn’t really physics, but it’s related. Say pathfinding: you have to determine, for a whole bunch of objects in the scene, whether they have visibility in the scene or you have to calculate the most optimal path through some obstacles. Each path that you’re calculating is a serial task; something that would run well on a CPU. But imagine if you have to calculate hundreds or thousands of these tasks per second. This is something that a GPU could potentially handle.

Now of course your GPU is also doing the rendering, so you have to very carefully look and ask, ‘even if we could do that on the GPU, does it make more sense to burden the GPU with that task, or should we use some free CPU core that would otherwise sitting idle?’ These are all the balances you have to make, so to say would physics run better on a GPU is a very broad question; it’s what kind of physics are you talking about and what else is the GPU doing and how free is the CPU.

There’s also the communication between the GPU and the CPU to consider, because if you use the GPU to start doing some of the processing, you have to communicate the results back to the CPU so that it can do something. ‘Now that I’ve determined that these two objects are colliding, what does that mean?’ The CPU has to go and figure that out. So you have to have efficient bandwidth.

Actually, bandwidth is less of a problem now, it’s actually latency. If you’re doing something like physics, if the physics affects how you interact with the game world, you need the response to happen immediately. So if I run into an object, and the detection that I’ve hit that object gets back to the CPU three frames later, now there’s a lag in what’s happening in the game versus how I’m controlling it, so we need a low latency connection.

This is something where having everything on one die, like Fusion, can potentially help.