2015: Discrete = 10%

As far as the low/low-mid end discrete graphics is concerned, the main advantage they have is probably not allowable power draw. For portable systems it's the total system power draw that matters anyway, and even for stationary systems there are advantages to being able to power manage CPU and GPU in a coordinated fashion with APUs. For a given cost of power supply and a given level of system noise, an APU will probably be able to be fairly competitive in terms of computational power, (and superior in terms of needed volume, potentially allowing some innovation and diversification in form factors).

Rather, the main advantage with a dedicated GPU lies in it having it's own, fast, memory pool. For instance, even today some HD5550 cards have 128-bit GDDR5 interfaces, giving them some four times the theoretical bandwidth of the upcoming APUs from Intel and AMD, where additionally the main memory interface has to be shared between CPU and GPU needs. That's a major differentiator.

Also, since the PC processor business model hinges on being able to sell the same or similar part at a multitude of tiers, it is unlikely that the main memory channel will move very far from the low end CPU needs. Indeed, AMD seems to have no interest in offering anything beyond 128-bit DDR3 for their upcoming processors no matter the price tag, and Intel will only offer 256-bit memory on their highest end processors, and those who are prepared to pay for that are unlikely to want that bandwidth shared with on-die graphics. So, for the foreseeable future, bandwidth is the argument to be used to promote stand alone graphics.
 
I wonder... since DDR4 is limited to one DIMM/channel, might we see 192-bit become more mainstream both to meet IGP bandwidth demands and to allow 3 DIMMS in total? That could get us to GTX460-level performance for IGPs in, say, 2014 on 16nm. I suppose whether that is realistic in the mainstream also depends on the die size of DDR4 PHYs/IOs.
 
Maybe this is an opportunity for SKU segmentation. You can have low end Bobcat based cores and higher end Llano but between these two you can have an option for say on package frame-buffer for both. That way you can have low end Ontario / higher end ontario with frame buffer and low end Llano and higher end Llano.

If they use something like Z-Ram / Ed-ram they should get good density especially considering laptops/desktops seem to be stagnating at HD 1080P resolutions. It shouldn't cost too much relative to the higher potential margins to offer say 16/24MB frame-buffer. That should enable the APU to perform like a descrete chip.
 
Addendum for correctness: Both AMD and Intel are said to support quad-64bit memory channels on (at least some of) their workstation/server versions of their future 32nm CPUs. Intel is rumored to offer an "enthusiast" platform that also supports four memory channels. There are no such rumors about a four channel consumer platform from AMD.

My earlier statement about memory support referred to consumer platforms only.
 
Back
Top