Rumor has it AMD is planning a dual-GPU graphics card based on its forthcoming Vega 10 architecture that's due out next year. The card will be aimed at the professional market (initially, at least) where it will serve up an impressive amount of power for professional level content creation and things such as deep neural learning and artificial intelligence.
In a previous leak that laid out AMD's GPU architecture roadmap, Vega was pegged for a 2017 launch. It's said to feature 64 Compute Units, 24 TFLOPs of 16-bit computing power, and a 225W TDP. Assuming AMD sticks with 64 compute cores per compute unit, as is the case with GCN, we can extrapolate from that 4,096 cores for a fully functioning Vega 10 GPU.
A dual-GPU variant would effectively double those specs. If that's the case, then a dual-GPU Vega card would likely have 128 compute units and 8,192 stream processors. It will also have 16GB of HBM2 (second generation High Bandwidth Memory) per GPU, so look for it to be marketed as a 32GB HBM2 card with 1TB/s of total memory bandwidth (again, we're just doubling up things here).
Shall we take things a step further? Vega 10 is rumored to wield 24 TFLOPs of 16-bit half-precision computing power, or floating point performance. Cutting that in half gives the dual-GPU Vega 10 card 12 TFLOPs of single-precision floating point performance.
Dual-GPU cards typically sport slower clockspeeds than their single-GPU brethren. Previous rumors pegged Vega 10 running at 1,465MHz so it's reasonable to assume a dual-GPU version will run a few hundred megahertz less. If you estimate around 1,100MHz, you're looking 18 TFLOPs of single-precision performance. To put that into perspective, NVIDIA's new Titan X based on Pascal delivers 11 TFLOPs of single-precision floating-point performance.
Buckle up folks, it looks like 2017 is going to be another wild year in graphics!