The Inquirer-Home

AMD touts hardware level memory coherence in HSA architecture

Says developers don't need to worry about CPU and GPU
Tue Apr 30 2013, 05:00
AMD logo

CHIP DESIGNER AMD has said the memory architecture in its heterogeneous system architecture (HSA) will move management of CPU and GPU memory coherency from the developer's hands down to the hardware.

While AMD has been churning out accelerated processing units (APUs) for the best part of two years now, the firm's HSA is the technology that will really enable developers to make use of the GPU. The firm revealed some details of the memory architecture that will form one of the key parts of HSA and said that data coherency will be handled by the hardware rather than software developers.

AMD's HSA chips, the first of which will be Kaveri, will allow both the CPU and GPU to access system memory directly. The firm said that this will eliminate the need to copy data to the GPU, an operation that adds significant latency and can wipe out any gains in performance from GPU parallel processing.

According to AMD, the memory architecture that it calls HUMA - heterogeneous unified memory access, a play on unified memory access - will handle concurrency between the CPU and GPU at the silicon level. AMD corporate fellow Phil Rogers said that developers should not have to worry about whether the CPU or GPU is accessing a particular memory address, and similarly he claimed that operating system vendors prefer that memory concurrency be handled at the silicon level.

Rogers also talked up the ability of the GPU to take page faults and that HUMA will allow GPUs to use memory pointers, in the same way that CPUs dereference pointers to access memory. He said that the CPU will be able to pass a memory pointer to the GPU, in the same way that a programmer may pass a pointer between threads running on a CPU.

AMD has said that its first HSA-compliant chip codenamed Kaveri will tip up later this year. While AMD's decision to give GPUs access to DDR3 memory will mean lower bandwidth than GPGPU accelerators that make use of GDDR5 memory, the ability to address hundreds of gigabytes of RAM will interest a great many developers. AMD hopes that they will pick up the Kaveri chip to see just what is possible. µ

 

Share this:

blog comments powered by Disqus
Advertisement
Subscribe to INQ newsletters

Sign up for INQbot – a weekly roundup of the best from the INQ

Advertisement
INQ Poll

Dead electronic devices to be banned on US-bound flights

Will the new rules banning uncharged devices be effective?