B3D收集：
Navi brings CU/SIMD changes, judging by the recent patents.
The "features" from latest patents are literally fundamental CU-level changes.Operand reuse cache and the addition of side ALU is not a minor change.It's not even iterative.
It's going back to designs so ALU-heavy they can bruteforce through everything.Though meaningful geometry setup changes are coming in Navi iirc.

毕竟穷。没钱挖人。

The first patent is like a combo of Fermi (dual-dispatch, except not really, it's bundled into VLIW2 instruction by compiler) and Maxwell (operand reuse cache to save a LOT of power and some register pressure).Also NV saved even more power by dumbing down the register file.Maxwell was the best of both worlds, truly marvellous stuff,VLIW5 wasn't smart enough, GCN is too smart. Maxwell balanced it really well.But if AMD straps HWS+ACE combo to their not-Maxwell......that would be funny.Like even better of both worlds.

基本靠抄

别的不晓得 反正ROCm amd就是在抄作业 抄的还挺快 性能很强

对着nv的架构抄呗，显存带宽压缩，硬件调度，大幅度削双精度，反正只要把游戏GPU和计算GPU分开了还是很好办的（分开的过程也极度简单，计算GPU只需要不惜工本设计一个大核心，类似vega，除了这个东西以外，往下全系芯片全部往死里砍，凡事游戏用不到的功能全部砍光做超算的，总不会去买580这档次的做计算卡吧？全部便宜了矿卡）

那文字想当然！说不好听就是胡说八道，不过很合 JS口味

也就是读了下gcn3的引言对比了一下哪些新加的nv早就有罢了，有什么价值？ scalar alu啥的 nv到现在也没有。当年gk110的grid特性，出来的时候大吹一顿，然而gcn一出来就支持。对桌面卡都是些无关紧要的东西

瞎死我了，我还以为CSGO的Navi

GGCN的程度。