- 22 months ago
It isn't entirely clear that what advantages they could produce by introducing a new Volta chip. Simply checking naive division: [chip] [process] [#Gtransistors] [area in mm**2] [Mtransitors/area] 1030 12nm 1.8 74 24.3 1050 12nm 3.3 132 25.0 [same core used for 1050ti] 1060 14nm 4.4 200 22.0 1080 14nm 7.2 314 22.3 [including 1070, 1070ti] 1080ti 14nm 12.0 471 25.4 [including Titan X, Titan XP] volta 12nm 21.1 815 25.9 [numbers from Titan V]
Don't expect to get any more transistors from Volta and TSMC's 12nm process (Volta needs more communication lines, but also the FP64 logic should be more dense than consumer FP32 logic leaving a wash [at least for numbers derived from wiki]).
Nvidia made volta for their supercomputing and AI customers, presumably for a reason. So maybe the 12nm gives sufficiently high performance to give it a run. But don't expect miracles, and don't expect a GPU any time soon. Note that TSMC just claimed to have started producing "mass volume" on its 7nm process: nvidia may be targeting that and possibly waiting to see how it comes out before putting in massive manhours on a chip [you know Moore's law is dead if a company can wait for the process to be available. This last bit is entirely speculation on my part and reading quite a bit into "a long time from now"].
Note - Nvidia stuck with 28nm before dropping to 12-14nm. Perhaps they want to hold pat until they can cut the pitch in half at 7nm (of course, they shipped a few generations at 28nm, but there has been little reason to spend the money since).
On the Vega front, rumors have it that GloFo will produce Vega20 in 7nm as a "pipecleaning" run. Presumably this means a professional Polaris job [don't expect any volume from GloFo, and I doubt they plan on wild success from HBM[2-3?] either] with a consumer [i.e. remotely sanely priced] Radeon for another year once you see the RadeonPro (don't hold your breath for GloFo). AMD is likely pleased as punch with Raven Ridge, and expect further GPUs to be built primarily for ALU integration (I'm assuming this has been true since they started making ALUs, coinciding with total concession of the high end GPU market to nvidia).