Trinity… but Better!
AMD has finally refreshed their upper APU lineup with Richland
Richland. We have been hearing this name for a solid nine months. Originally Richland was going to be a low end Trinity model that was budget oriented (or at least that was the context we heard it in). Turns out Richland is something quite different, though the product group does extend all the way from the budget products up to mainstream prices. We have seen both AMD and Intel make speed bin updates throughout the years with their products, but that seems like it is becoming a thing of the past. Instead, AMD is refreshing their Trinity product in a pretty significant matter. It is not simply a matter of binning these chips up a notch.
Trinity was released last Fall and it was a solid product in terms of overall performance and capabilities. It was well worth the price that AMD charged, especially when compared to Intel processors that would often be significantly slower in terms of graphics. The “Piledriver” architecture powers both Trinity and Richland, and it is an improved version of the original “Bulldozer” architecture. Piledriver included some small IPC gains, but the biggest advantage given was in terms of power. It is a much more power efficient architecture that can be clocked higher than the original Bulldozer parts. Trinity turned out to be a power sipping part for both mobile and desktop. In ways, it helped to really keep AMD afloat.
It turns out there were still some surprises in store from Trinity, and they have only been exposed by the latest Richland parts. AMD is hoping to keep in front of Intel in terms of graphics performance and compatibility, even in the face of the latest Haswell parts. While AMD has not ported over GCN to the Trinity/Richland lineup, the VLIW4 unit present in the current parts is still very competitive. What is perhaps more important, the software support for both 3D applications and GPGPU is outstanding.
Digging into Richland
Richland is not a totally new product. It is in fact still based on Trinity, but it is a slightly different revision. A few hardware changes were made, but nothing that would radically change overall performance or characteristics. Instead, AMD activated parts of Richland that were not doing anything of great import in Trinity. The Trinity die was originally designed with quite a few sensor sites and a pretty robust controller that handled power needs depending on both workload and temperature.
Firmware and software both were heavily tweaked to enable greater overall turbo performance while keeping TDPs and power consumption in check. Tweaks were also made to the 32 nm HKMG/PD-SOI process to get a little more performance. Finally Richland is a new, minor revision of Trinity that enables the hidden functionality and helps to clock the APU up without eating a whole lot more power or creating more heat than previous parts at lower clockspeeds.
Richland comes in 100 watt and 65 watt TDPs for the desktop. Previously we saw the latest Richland mobile parts which could go down to 15 watts TDP in some circumstances. The high end unit is the A10 6800K. This part has a base clock of 4.1 GHz (up 300 MHz from the A10 5800K) and a max turbo clock of 4.4 GHz. It has a full 384 shader cores running at a maximum speed of 844 MHz. This again is up from the previous 800 MHz of the 5800K. This chip also increases the memory speed from previous models. The new APUs can now run officially at DDR3-2133 MHz speeds, which should give a nice little boost to most graphics workloads. The only downside to this speed is that AMD only officially supports 2133 speeds with one DIMM per channel. So those RG2133 4 x 4 GB packs that AMD is selling seem sort of useless when they may not entirely be stable with 4 DIMMS occupying 2 channels at 2133 speeds.
Motherboards will need to have their BIOS updated to be able to fully utilize the Richland design. There is a tremendous amount of firmware work that went into to make Richland what it is. The hardware is now much more aware of where potential bottlenecks are, and it will dynamically speed up and slow down different parts of the design to adjust for performance issues. Power, temperature, and utilization are all measured throughout the core by many sensors. If the CPU is bottlenecked, it will raise the voltage and clockspeed on that unit until a thermal limit is reached, then it will start clocking that part down. At the same time the APU will clock the GPU unit down and lower the voltage, so that it will act more as a heatsink for the CPU portion. It is not as efficient as what Kabini showed us in terms of providing a “heatsink” for the other portion, but it does make a small difference. A couple of extra seconds at a high turbo state might help to complete the work quicker, and allow the APU to be clocked down until it needs to be under heavy load again.
Richland does seem to stay under turbo mode longer than previous versions. This can be a blessing or a curse. In small chunks it is a fast processor, but the longer it is under load the lower the clockspeed goes. For example, when running Cinebench and watching the watts being measured, it would start out at near the highest recorded wattage and about 30 seconds later we would see power being lowered over time. There can be a good 5 to 7 watts difference between the peak at the beginning of the benchmark to when the test ended. From top to bottom there are also more frequency and voltage levels available to the processor. These again are entirely dictated by power consumption, load, voltage, and temperature.