The hills, or at least the internets are alive with the sound of rumours, this time indicating that AMD will be using 14nm FinFET due to 20nm problems.
The hills, or at least the internets are alive with the sound of rumours, this time indicating that AMD will be using 14nm FinFET due to 20nm problems.
AMD’s Radeon R9 400-series “Arctic Islands” GPUs – the flagship of which is codenamed Greenland – is still some way away with an expected 2016 launch date, but that hasn’t stopped rampant speculation even as we await the R9 300-series. The latest, from the worthies over at Chinese tech site EXPreview (Google Translate Link) and echoing similar reports from Fudzilla, claim that AMD will be skipping the 20nm process in favour of a 14nm FinFET process with an as-yet unnamed fab.
Moving straight to 14nm is certainly a bold move, and one which Nvidia is also expected to take when their Pascal architecture enters production. At this stage it’s widely accepted that yields on TSMC’s 20nm process have been poor, and as a consequence both Nvidia and AMD moved ahead with 28nm variants of Maxwell and Fiji respectively. Now 20nm has been deemed unsuitable for a chip of Greenland’s projected size, and when compounded by Apple’s A8 SoC carving off the lion’s share of TSMC’s 20nm production makes AMD’s use of the process unviable.
Utilising 14nm should, other factors notwithstanding, make the GPUs more energy efficient and hence improve the all-important performance per watt metric which NVIDIA recently claimed thanks to the release of the GTX 900-series. A derivative of the architecture seen on Greenland GPUs is also expected to be introduced into AMD’s Zen line of APUs, seen as critically important not only for AMD’s HSA initiative but also their continued presence in the desktop and server CPU market as a whole.
There is no word of a foundry in the frame to manufacture AMD chips and options are limited. Intel has a proven 14nm process, whilst Samsung/GloFlo are also manufacturing the Exynos 7 SOC using a 14nm FinFET process; however Intel at least have been reticent in leasing their manufacturing to competitor products in the past. Despite the poor yields both AMD and Nvidia may end up sticking to TSMC (on a new 16nm process), hoping that the reduced die area and a different process will yield better results than the now tarnished 20nm production method.
—
Supplementing the above is the news that AMD’s low power Arctic Islands GPUs will also make use of HBM2 memory, in the process doubling both bandwidth and capacity over HBM1 whilst also halving power consumption compared to GDDR5.
SOURCE: EXPreview, FUDZILLA
