Llano's GPU uses AMD's older VLIW5 (Cypress) architecture rather than the latest VLIW4 (Cayman) architecture. Codenamed Sumo, it is derived from last year's 400-SP Redwood core found on the HD 5670 and HD 5570.
As you can see below, AMD devoted a good amount of die space to the GPU on Llano, way more than Intel did with Sandy Bridge. With the integration of PCIe lanes, the Northbridge is now completely contained within the APU.
On A8-series APUs, the 6550D will be clocked at 600MHz. On A6-series APUs, the 6530D will be clocked at 443MHz and have one SIMD disabled, leaving us with 320 SPs.
|HD 6550D||HD 6530D||HD 5570||HD 6570|
|Memory Clock||Up to 933MHz||1000MHZ/900MHz||1000MHz|
|Memory Bus||128-bit shared||128-bit||128-bit|
Up to 512MB of system memory can be allocated to the GPU. The GPU accesses this dedicated memory via the Radeon Memory Bus. In addition to this, there is also a Fusion Compute Link which is used by the GPU to access shared system memory.
Ultimately, the four CPU cores and the GPU will have to share 29.8GB/s (assuming DDR3-1866 speed) of memory bandwidth. Contrast this with a typical 1000MHz GDDR5-equipped GPU, which gets 64GB/s of dedicated bandwidth. It looks as if Llano's GPU will be significantly constrained by this, which is why AMD recommends using the fastest possible memory. We'll be running our graphics benchmarks with both DDR3-1333 and DDR3-1866 to find out just how much this memory bottleneck affects performance.
Next, we'll take a look at the chipset accompanying Llano.