![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
This slide deck was from the last HotChips in 2012, but it has a bit of interesting info about GT4: http://www.highperformancegraphics.org/media/Hot3D/HPG2012_Hot3D_Intel.pdf
At 1GHz, it's projected to have 2 TFLOP of shader performance. Haswell's 40EU (160 ALUs) GT3/GT3e, by comparison, has 832 GFLOP of shader performance (SP) at 1.3GHz.
Page 19, where GT4 is mentioned, includes a diagram with 96EU. IB/Haswell shared a "gen 7" ALU architecture, but something seems to have been improved in Broadwell's GPU to reach 2 TFLOP @ 1GHz. I'm not getting the 2 TFLOP number without something like: 96 EU * 5 ALU/EU * 4 FLOP/ALU * 1000MHz = 1.92 TFLOP, but 5 ALU per EU is just such an odd number. (widening the ALU to 320 bits would also work out with 4 ALU/EU.)
At 1GHz, it's projected to have 2 TFLOP of shader performance. Haswell's 40EU (160 ALUs) GT3/GT3e, by comparison, has 832 GFLOP of shader performance (SP) at 1.3GHz.
Page 19, where GT4 is mentioned, includes a diagram with 96EU. IB/Haswell shared a "gen 7" ALU architecture, but something seems to have been improved in Broadwell's GPU to reach 2 TFLOP @ 1GHz. I'm not getting the 2 TFLOP number without something like: 96 EU * 5 ALU/EU * 4 FLOP/ALU * 1000MHz = 1.92 TFLOP, but 5 ALU per EU is just such an odd number. (widening the ALU to 320 bits would also work out with 4 ALU/EU.)