Teeth make smiles, and smiles make sales - Unidentified Harrods person in Alan Sugar's The Apprentice
CHIPMAKER Intel will put DRAM on the same package as its Haswell processor to act as a cache that is visible to both the CPU and GPU.
Intel has said that it will reveal its next generation Haswell processors at the Computex trade show next month. However the firm revealed arguably one of the biggest features ahead of time, that it will put DRAM on the same package as the CPU and GPU and offer both processing units read and write access to the memory.
As Intel was revealing branding and its own 3DMark performance figures for a selection of the GPUs that will ship in Haswell processors, the firm mentioned the GPU will have embedded DRAM. Later Intel confirmed that the embedded DRAM would effectively be a cache that is available to both the CPU and GPU and that it will be on-package, rather than on-die, something that Nvidia talked about for its Volta GPU.
John Webb, product line manager at Intel confirmed that the DRAM will be on-package, rather than on-die. He said, "It is a technology that Intel developed, it is on-package and it is actually accessible by both the graphics execution unit of our processors as well as the core CPU component. And it does benefit the performance on both the CPU workloads as well as the graphics workloads."
Webb told The INQUIRER that the memory that will be on-package will not be standard DDR and said the firm will use the same cache coherency mechanisms that it uses for multi-core CPUs. He said, "The memory is a custom DRAM made by Intel. It is designed to work as a cache and as such access by CPU and graphics engines are handled by the same mechanisms which arbitrate cache access between the CPU cores so race conditions are avoided."
Intel's decision to stick memory on the same package as its cores highlights the firm's lead in manufacturing capabilities over its rivals. However it also shows that by giving the CPU and the GPU access to memory, albeit considerably smaller than system memory, the firm is not far from extending it to what AMD, ARM, Texas Instruments and Qualcomm are doing with their heterogeneous system architecture (HSA).
AMD and its band of HSA members can rightfully claim that the upcoming Kaveri chip can offer developers easier access to more than just a cache, but Intel has shown that it too can offer the CPU and GPU access to the same memory, and should perhaps reveals why it didn't bother signing up to become a member of the HSA Foundation. µ
Sign up for INQbot – a weekly roundup of the best from the INQ