Nvidia’s Maxwell GPU architecture will access system RAM
Nvidia’s GPUs have traditionally been limited to only being able to access memory on the card, though Kepler can access other GPUs through the firm’s GPU Direct technology. However the firm’s next generation Maxwell GPU architecture will allow the GPU to access main memory, reminiscent of AMD’s unified memory architecture.
Nvidia co-founder and CEO Jen-Hsun Huang said Maxwell, which will replace the present Kepler architecture, will be the firm’s first GPU architecture that is able to access main memory.
Although Nvidia’s Kepler architecture can address terabytes of memory, the cost and memory density limitations of GDDR have meant that its Tesla boards are limited to 8GB of memory. While the firm’s GPU Direct allows Kepler based Tesla boards to access memory on other Kepler boards, its bandwidth is limited by the network connection, which typically tops out at 40Gbit/s, while DDR3-1600 bandwidth is 12.8GB/s – or about 100Gbit/s – and avoids the network latency.
AMD has been pushing the ability of GPUs to address system memory as part of its unified memory addressing technology that will be a key part of its Heterogeneous System Architecture (HSA). The firm is expected to produce its first HSA enabled APU later this year and the technology has been incorporated in the APU that Sony uses for the PlayStation 4.
Nvidia will be using its unified memory access primarily to help researchers and those that use its Tesla GPU accelerators work with larger data-sets Nevertheless, that Nvidia will be incorporating unified memory addressing in its future GPU architecture justifies AMD’s decision to back the technology so aggressively.