Now, how do the CPU and GPU fit together? You might want to read The AMD A8-3500M APU Review: Llano Is Unleashed for an explanation of how AMD plumbs its APUs, since the following covers Trinity specifically.
The new APU comes packaged on a 226 square-millimeter die with 1.303 billion transistors, which is almost exactly same size as Llano’s 228 mm2 die. Like Llano, Trinity is manufactured on GlobalFoundries’ 32 nm SOI process. Default frequencies scale much higher than Llano did, with 3.8 GHz CPU and 800 MHz GPU ceilings for the top-end desktop parts. TDPs on the low-power mobile models dip down as far as 17 W. We’ll delve more into the specific productization shortly, but it’s at least apparent right away that Trinity scales both higher and lower than Llano. As a side note, the new APU comes with either one or two modules, equaling two or four cores, so there won’t be any six- or eight-core APUs on their way to the market in this generation.
A quick peek at the die makes it clear that AMD is devoting more die space to graphics—an observation we made comparing Intel’s Ivy Bridge to Sandy Bridge. This isn’t much of a surprise, though, considering the Piledriver architecture makes generous use of resource-sharing.
The New Unified Northbridge
Llano utilizes a somewhat standard CPU-to-Northbridge link and the Fusion Control Link for GPU-to-Northbridge memory exchange. In Trinity, a new Unified Northbridge (UNB) replaces the older hardware. This is AMD’s first application of a UNB outside of the server space; the company’s product CTO Bob Macri likens it to a traffic cop at the heart of the machine. It’s the component that ties everything to the memory and I/O subsystems. It must cope with various functional units, each making unique requests with specific attributes.
For example, the CPU hands relatively few requests to the UNB, but these are given a high priority since any added latency can have a significant impact. On the other hand, the GPU makes thousands of outstanding references and needs a lot of DRAM utilization to perform well. In this case, the UNB optimizes and reorders requests so the GPU can access the memory subsystem in the most efficient way possible. The UNB also has an important role in keeping the power down, and will modulate memory frequency depending on the load to get the optimal power usage per unit of work.
The New Memory Controller
An APU’s memory subsystem is especially important because so many functional units contend for bandwidth. Trinity’s memory controller is a new design with support for 1.25 V DIMMs that introduce potential power savings. As for data rates, DDR3 modules are limited to 1866 MT/s on the desktop and 1600 MT/s in mobile configurations, similar to Llano. The platform addresses up to 64 GB in desktop form, with peak system bandwidth as high as 29.8 GB/s. The mobile config accommodates up to 32 GB of memory and bandwidth as high as 25.6 GB/s.
Accelerated Computing And The IOMMUv2
Llano is rated for up to 572 gigaflops of compute performance, but the top of the Trinity stack has the potential for 736. The increase is due mostly to improvements inherent to the Devastator GPU, but something that catches our eye is the new IOMMUv2 block.
This unit adds virtual address access discrete graphics, allowing an external GPU to directly access the same virtual address space as the CPU through page tables. As you can imagine, this is a key part of the programming model for AMD’s Heterogeneous Systems Architecture (HSA).
Turbo Core 3.0
Third-gen Turbo Core is the latest version of the feature first introduced in AMD’s Thuban-based Phenom II X6 CPUs, and it’s now purportedly dynamic in the way it balances CPU and GPU clock rates. The above slide from AMD’s presentation suggests that power and frequencies can be shifted from CPU to GPU on demand to meet the system load.
Unfortunately, this isn’t the behavior we witnessed in AMD’s system monitor utility, which reported CPU performance at a 2.3 GHz nominal rate, regardless of load. We did, however, see the GPU speed fluctuate and drop as low as 334 MHz during a single-threaded Cinebench run. We’re not sure if Turbo Core 3.0 isn’t working properly or if AMD’s system monitor is reporting the CPU clocks incorrectly. We asked AMD about it utility and were told that there’s currently no tool able to accurately monitor its chip’s Turbo rates. Incidentally, that’s the same answer we received a year ago when the company introduced Llano. You can be sure we’ll be diving into more depth on this in the days to come.
As with Llano, Trinity is designed with low-power applications in mind. You can see the main power improvements AMD made to its APU in the diagram above, highlighted. This includes an improved CC6 power state able to turn off individual Piledriver modules when three or four execution cores aren’t needed. Both modules can be shut down on a package level.
Under normal use, the APU aggressively employs both DDR3 memory channels for graphics data. But when the screen is static, the APU detects more modest activity and remaps the output to a single memory channel. It can even put the unused channel to sleep and reduce the clock rate of the active channel to the minimum required for the display.
Engineers also increased the on-die buffering of memory so that DRAM isn’t queried as much, meaning power-hungry memory is left in a low-power state as long as possible.
- AMD’s Next APU: Trinity
- The CPU Side: An All-New Piledriver Core
- The GPU Side: VLIW4 > VLIW5
- All Together Now: The Trinity APU
- Products And Platforms
- Test Setup And Benchmarks
- Graphics Benchmarks: 3DMark
- Benchmark Results: DiRT 3
- Benchmark Results: The Elder Scrolls V: Skyrim
- Benchmark Results: Crysis 2
- Benchmark Results: Metro 2033
- CPU Benchmarks: PCMark And Sandra 2012
- CPU Benchmarks: Productivity
- CPU Benchmarks: Content Creation
- CPU Benchmarks: Media Encoding
- APU Enhanced Software And The User Experience
- Power Consumption
- Tripling Your Pleasure With Trinity?