AMD’s upcoming “Polaris” and “Vega” GPU architectures will appear to take advantage of performance/watt gains to keep their stream processor counts low, and sizes quite small. This is according to VideoCardz who has spotted a curious-looking CompuBench entry. If a Graphics Core Next (GCN) compute unit (CU) of the “Polaris” architecture consists of 64 stream processors, then AMD’s Polaris 11 silicon (Baffin) could feature over 1024 stream processors across 16 CUs. AMD Polaris 10 (Ellesmere) would feature 2304 stream processors spread across 36 CUs, and finally Vega 10 would feature 4096 stream processors across 64 CUs.
The “Baffin” will succeed the current generation “Curacao” silicon, which drives mid-range graphics cards. It is expected to feature a 128-bit wide GDDR5 memory interface, which will hold 4 GB of memory. “Ellesmere” will succeed the current generation “Tonga” silicon and will drive performance-segment cards. It could feature up to 8 GB of GDDR5(X) memory. We could see these two chips by mid-2016.
The third chip is Vega 10, which could succeed Fiji. This chip will take advantage of the HMB2 interface and could feature 16 GB of memory. It could launch some time in early 2017.
Keep in mind that AMD is claiming a massive 2.5x performance per watt increase for Polaris over its current GCN 1.2 architecture so these stream processor counts do look deceptively insufficient.