SAN FRANCISCO: CHIP DESIGNER AMD said lower prices for GDDR5 memory will lead to its use as part of chip memory hierarchy.
AMD has long been promoting its idea of unified memory access where the CPU and GPU both have access to main memory. What the firm is now looking towards is making that main memory comprised of both types of DRAM, that is, both DDR memory and significantly higher bandwidth GDDR memory.
John Gustafson, chief graphics product architect at AMD told The INQUIRER that firms have been looking to increase bandwidth between the CPU and system memory but the price has been prohibitive.
He said, "Everybody is looking to use that extra speed memory in combination with the lower cost memory, just like you do with all memory tiers. It's a new tier of memory. A long time ago in HPC you always wanted to have static RAM right next to the processors and the DRAM far away and we're back to that situation again."
Gustafson's reference to static RAM (SRAM), is a nod to the well documented history of high performance computing (HPC), where SRAM is preferred for its performance characteristics but is naturally limited by its price. However Gustafson said that gamers' demand for high performance graphics cards has lowered the cost of high performance memory such as GDDR to a point where it becomes feasible to use gigabytes in consumer machines such as Sony's Playstation 4.
Gustason said, "Everyone's been using conventional DRAM because if you walked even a tiny bit away from conventional DRAM the price went up by a factor of 100, but who broke the back of that problem was the video gamers, because they are demanding this stuff by the millions of parts and they are engineered for a reasonable price.
"They are still more expensive, but it's not crazy expensive so for all the applications that are so memory bound that if you make the memory five times faster they will run at least four times faster, there's a lot of people that will pay for that."
Both AMD and Nvidia have said that the GPU's biggest challenge is memory bandwidth, as the chips cannot be fed with enough data to make use of its computing capabilities. Gustafson called memory access the "GPU's biggest problem".
While Gustafson doesn't see DDR memory going away anytime soon, he sees that processors such as AMD's APUs will have access to GDDR memory in the future as part of a memory hierarchy. Given that AMD, Intel and Nvidia all offer somewhere between 6GB and 8GB of GDDR5 on their high-end GPGPU accelerator cards, it might not be too long before a subset of that capacity trickles down to consumers. µ
Sign up for INQbot – a weekly roundup of the best from the INQ