And this shows it is 2.5x as fast. And this is with Iris Pro 128MB eDRAM advantage.
Even on your 72 EU to 128EU scaling, Still a difference between 1.8x and 2.5x.
All I have to say for now is iGPU capabilities != dGPU capabilities.
- More space in chiplet, lower cost.
- More space => Binned primitive shaders, JIT in drivers; Implicit. Wider binned RBE, etc // accelerated by discrete units on die, which take space. ((Bigger caches, more registers, etc can be there as well))
- Lower cost => means HBM2 can be deploy; HBM2 G3 => 8 gigabytes @ 341 gigabytes per second(2666 MHz), HBM2 G2 => 4/8 gigabytes @ 307 gigabytes per second(2400 MHz).
(HBM G3 costs less than HBM2 G2 and HBM2 G2 is cheaper than HBM2 G1, etc)
128 megabyte eDRAM @ 102 gigabytes per second... games use more than that so... 35 gigabytes per second DDR4, etc.
There is also the clock rate increase, etc. Gen 9(14nm) != Gen 9++.x++ or whatever (14nm++ to 14nm# or 14nm-rust;; 14nm# is 14nm++ with Gen 1 ML design/14nm-rust is 14nm++ with Gen 2 ML design). ((PnR(Place N Route) design, not an AI doing everything, etc.))