I'd guess they use both GDDR5 and DDR4 controllers in the design. AMD is really big on reusing designs (at 80-130 million per mask set for modern manufacture no

I'd guess they use both GDDR5 and DDR4 controllers in the design. AMD is really big on reusing designs (at 80-130 million per mask set for modern manufacture nodes, it's little wonder). Making a profit requires selling at least 10M chips (assuming around $200M masks/R&D and a $20 charge over base cost). The 2400G is 209mm^2. Doubling the GPU will move that to somewhere around 275-300mm^2 which should place raw chip cost between $35-40 and final chip cost over $60 (probably closer to $80 minimum) with R&D, masks, and profit. But then there's binning. They will get 3 binnings with the target middle being the desired chip. The extra 15mm^2 for a DDR4 controller would mean they can laser out the GDDR controller and sell it as a 2500/2600G chip or maybe (less likely) make a package that allows both DDR4 and GDDR5. That's a great use for those over/under performing chips. A huge question here is L3 cache. Will they be adding extra L3 to try to offset the much larger GDDR5 latencies?