How many die shrinks were there between the original xbox 360 and when it merged the CPU and GPU onto the same die? I realize there are benefits to be had by sharing transistors, but it's also not a fair argument given the technological advantage of die shrinks and design improvements. Saying an A10 APU is superior to a 9800GTX + equivalent CPU processor is kinda
when we're talking about several years difference.
This was the second die shrink, if my memory is correct- I think it was 90->65->45nm.
Don't get too hung up on the A10. In currently available APUs, there are a few factors that limit their potential-
a) Sheer quantity of graphics hardware. Top end Trinity has a fifth of the number of shaders of the HD6970, the discrete GPU it has most in common with (VLIW4). However, if you attempt to increase this number dramatically then you run into the next problem:
b) Memory bandwidth. Dual channel DDR3 just doesn't offer that much, as we know from years of the GPU market. Even something low end like a HD6670 is bottlenecked by DDR3 memory, and that doesn't have to share the bandwidth with CPU cores.
c) Split memory spaces. Llano and Trinity offer some features for memory snooping between CPU and GPU (Onion and Garlic buses), but these are still limited. Without the two working directly in the same memory space, you lose a lot of the inherent efficiencies of an APU design as you need to copy resources between the two, waste chunks of memory by having duplicates of resources, and run into utilisation problems (CPU or GPU uses all of its partition up, while there is spare space going to waste in the other side).
d) Software design limitations. Pretty much every single current PC game is written to run well on discrete graphics cards. Even when APUs started exposing a common memory space, they will still not be used efficiently by current games. These are built on APIs which presume an object worked on by the CPU cannot be worked on by the GPU in the same spot, and vice versa. Large quantities of memory and performance are going to go to waste.
The PS4 fixes all four of these issues. Totally new frameworks built from the ground up for this platform, with buckets of graphics power and the bandwidth to feed it. Not to mention that rumours are pointing towards the next Xbox having essentially the same design.
It's interesting insofar as the cpu really, the gpu is already 30%~ or more slower than Titan (best case, a few years down the road when they go full direct to hardware), which should be offered by GM104 on 20nm when it arrives. However the more important aspect is the cpu, and how much performance they can get out of it imo.
In terms of raw horsepower, yeah, it's considerably weaker. But this design means that they can support whole new features that a standard PC design just can't do. Look at GPGPU physics- PhysX can come up with some seriously nice looking graphical effects (especially with the GPGPU power of GK110), but it can't affect the underlying simulation of the game world. But imagine if all of these thousands of simulated particles actually impacted on the world, knocking players out of place, deflecting bullets, washing bridges away in a torrent of GPGPU fluids, whatever. That's what you could realistically achieve by killing the communication cost between CPU and GPU.
Facebook
Twitter