NVIDIA 28nm Kepler performance falls short of expectations

NVIDIA has confirmed that the next generation of Kepler will be postponed until 2012. The immaturity of TSMC's new 28nm process has always been considered the culprit, but not all of them seem to blame the foundry.

According to Taiwan media reports, Kepler’s performance has failed to meet NVIDIA’s expectations. NVIDIA plans to make Kepler’s double-precision floating-point model with a 2.5-fold-per-watt double-precision floating-point model, and continue to strengthen general-purpose computing capabilities. It seems that it still needs further improvement. It is unclear what level Kepler has achieved, but Kepler had previously reported that he had not encountered any major problems.

TSMC plans to begin batch production of 28nm process in the Fab 15 wafer fab in the fourth quarter of this year and begin trial production of the 20nm process in the third quarter of 2012.

Another major customer of TSMC is still planning to release three 28nm chip 8960, 8270 and 8260A in the fourth quarter of this year, all based on the dual-core Krait architecture. AMD plans to release the first 28nm GPU in the second half of this year, opening the "Southern Islands" era. Next year's new APU Krishna and Withita will also use 28nm.