The eDRAM cache offers a performance improvement bigger than a new arch, or alternatively power savings equal to a new node, and we consider it expensive?
Yea, because only parts with eDRAM will be faster. Costs for new architecture will be spent once and its something they do anyway(R&D). With eDRAM its persistent cost with every part they sell. Its not like eDRAM speeds up many CPU applications anyway. The fact is there is also packaging complexity.
I think that's Broadwell 2C+GT3 die size, maybe CPU-Z is reading Skylake wrong.
That's
GPU-Z. Also GPU-Z is consistently wrong with specs of Intel GPUs. GT2 Broadwell has 3 Samplers, with 4 TMUs each, making it 12. That screenshot is also assuming it runs at 350MHz with 16 TMUs, but that's for Haswell. Broadwell has 24. It also says Broadwell GT3 in one part and Skylake GT2 for the name.
Sysoft is right were Skylake should be, that Cinebench score is worrisome instead.
I hope it's just wrong and there is "some" improvement clock/clock over Haswell because better graphics (and power consumption?) aren't exactly my priority now...
Sandra isn't better either. If you ignore memory BW and cryptography benchmark, also FP, and look at ALU benchmark the gain is ~7%. The one which gets 21.5% improvement is likely using the GPU.
PCMark 8 is boosted because it includes a graphics portion, and if you look at the original link, here's the shocker:
Skylake is barely 5% faster, and is sometimes slower. Firestrike Physics score is showing 7% improvement meaning maybe we get 5-7% improvement. Not really better than Broadwell gains. The sadder part is that 5-7% improvement might be over Haswell.