Back in early December 2015, AMD invited their preferred media partners to brief them about the new company structure, where Radeon Technology Group (RTG, also known as ATI 2.0) will continue to bring products to market under the AMD banner until the (almost inevitable) spin-off occurs.
In a run-up to CES 2016, which takes place in Las Vegas, NV, AMD made a formal announcement introducing the Polaris GPU architecture. In a video bewlo, Raja Koduri (head of RTG) spoke of innovation that was enabled by a logical transition from planar transistors (single-gate, i.e. 2-D) to multi-gate (i.e. 3-D) technology, with a variant of 3-D transistors being called FinFET. If you recall, Intel was the first to introduce 3-D transistors with its 22nm Haswell processors, but the research on multi-gate transistors was lead by many, including AMD themselves (at the time when they owned foundries).
For years, AMD was saying that SOI (Silicon-On-Insulator) and FinFET is the future of their silicon, and with Zen x86-64 CPU, K10 64-bit ARM CPU and Polaris GPU architectures the company is pulling no stops to bring their technology to market and beat NVIDIA in semiconductor space for the first time ‘since forever’. Nvidia will manufacture Pascal and Tegra P-class silicon using TSMC’s 16nm FinFET, while AMD will utilize now-defunct Common Platform (GlobalFoundries and Samsung). Future AMD parts should come out of New York (GlobalFoundries) and Texas (Samsung), meaning we will probably have an odd combo of ‘Diffused in the United States, Assembled in Malaysia / Taiwan / South Korea).
Arctic Island family of 14nm GPUs is divided into Baffin (mainstream), Ellesmere (performance) and Greenland (enthusiast). All utilize 14nm FinFET process, and the company will use Baffin and Ellesmere to enable GlobalFoundries and Samsung to ‘optimize’ the manufacturing process, sort out as much as potential ramp issues as possible before Greenland arrives to fight head to head with GP100, i.e. high end Pascal part.
On the question what will the first product be, AMD was quite cautious, as they don’t want to “Osborne” the delayed/upcoming Gemini dual-GPU card. The focus of enthusiast community needs to be on the upcoming R9 ‘Gemini’ (Radeon R9 Fury X2) part, which was delayed to align with Oculus Rift retail/e-tail launch. Thus, too much talk about 14nm FinFET future parts might have a detrimental effect on sales of Gemini cards. After all, Gemini uses GCN 1.3, while Baffin / Ellesmere will use Polaris GPU architecture (GCN ‘2.0’).
Thus, it is no surprise that AMD’s/RTG’s first 14nm FinFET part is targeting mainstream, high-volume market. Internally known as Baffin, this graphics chip is expected to launch in just two months (March 2016 timeframe – think GDC San Francisco). Still, the main ‘harvest’ for the company should be Computex 2016, since mobile 14nm parts will make its debut there, and given the much touted ‘2x increase in performance/watt’, AMD will be expecting to win a big number of OEM designs.
Baffin is quite a small GPU targeting the $149.99 price point range, and will square off against Nvidia’s GeForce GTX 950, as compared to in an AMD-sourced video you can watch above. Ellesmere will replace the 300 series, and go against GeForce GTX 960/970, while Greenland will go and hit 980 Ti / Titan X and their Pascal-based successors.
This also means AMD / RTG will bode farewell to its current, somewhat confusing product lineup that features different iterations of Graphics Core Next (GCN) architecture. At the moment, AMD sells APU and GPU products featuring GCN 1.1, 1.2 and 1.3. With the completion of 14nm switch we will see every part, both APU and discrete GPU using the same Polaris architecture, a.k.a. Graphics Core Next 2.0.
This year, the battle will be held in 1080p, 4K, 5K and VR. RTG is coming into the fight better equipped than ever – Socket AM4 will bring DDR4 memory, Zen APU, Arctic Islands GPUs… probably the strongest line-up ever. Will it be enough to capture the momentum and marketshare? Only hiccups in execution, product availability and sales strategy might endanger that. And of course, this…