The Kaveri family of processors will be able offer up to 856 gigaflops of compute power. This power will be delivered by four "Steamroller" core CPUs – with each pair of cores sharing 2MB L2 cache — and multiple Radeon GCN cores.
The biggest change to the processor design introduced by Kaveri is the move to a Heterogeneous System Architecture (HSA).
The architecture allows the CPU and GPU to work more closely together in tackling different compute workloads. CPUs and GPUs are suited to different types of computing tasks, while CPUs are good at handling serial workloads GPUs are suited to processing compute jobs in parallel. By matching workloads to the appropriate processor AMD predicts Kaveri will increase performance and reduce power consumption
"We've always been producing APUs but essentially they've been two asymmetric blocks and the whole idea with HSA is to find ways of allowing these two to interact more closely together. The idea is to unlock more gigaflops than was possible before," said Adam Kozak, AMD senior product marketing manager.
The HSA introduces a heterogenous unified memory architecture, which allows the GPU and CPU to address the same memory space. Traditionally, CPU and GPU have to address different blocks of memory, which means data has to be copied between these two blocks. By eliminating the need to copy data between these different blocks the CPU and GPU can both work on the same data without waiting for it to be copied back and forth.
Another HSA feature is heterogeneous queueing, which rebalances how the GPU and CPU interact with each other to give them more of an equal footing. While the CPU assigns workloads to the GPU in most current PCs Kaveri's queueing feature allows both processors to dispatch work to each other and create work for themselves.
High definition video playback will also be eased by the AMD Unified Video Decoder, which supports hardware decoding of H.264 and VC-1 video codec standards. The chip also has built-in support for video encoding for H.264 via its video compression engine.
AMD will launch the first Kaveri processors, on an FM2+ socket, for desktops. Shortly after AMD will launch processors based on the same architecture as Kaveri aimed at the notebook, embedded and server market.
Besting microservers with Berlin
For servers AMD will launch the Berlin family of processors, an APU that shares the same Heterogeneous System Architecture as Kaveri targeted at the microserver market.
Microservers are small servers with low power consumption that are packed into dense clusters designed to carry out workloads that need to be scaled across many processors.
When Berlin launches in the first half of this year it will likely be competing against Intel's family 64-bit Avoton S1200 Atom processors, or a 14nm die shrink of the Avoton.
Berlin is the sucessor to AMD's Opteron-X series, which were its first single processor APUs designed for dense server clusters.
AMD says the additional performance, flexibility and parallel processing capabilities of the Berlin's HSA will allow it to be used for a wider range of workloads than its predecessors. These workloads include big data processing — analysing huge amounts of text and visual data, high performance compute jobs — tapping into the parallel processing power of the built-in GPU, running hosted desktops, as well as streaming video and other online multimedia delivery.
"We have introduced smaller, more power efficient processors that are for what we call web or enterprise services clusters," said Margaret Lewis, director of server ISVs and software planning at AMD. "We've learned from our customers that servers have to be scalable and adaptive."
AMD will also offer a new unified SDK for its APU and improvements to the CodeXL Developer tool suite.
More details will be released about the Kaveri and Berlin family of processors at the AMD Develeop Summit over the next couple of days.