4 ms·
Why does this have (1997) in the title? Looks like a 2018 paper. Although I wonder what’s the earliest you could have tortured GPUs to do machine learning effec
by joefourier 4y ago
Why does this have (1997) in the title? Looks like a 2018 paper. Although I wonder what’s the earliest you could have tortured GPUs to do machine learning effectively, it’s certainly possible with fragment shaders only…
- dang 4y agoWe've s/1997/2018/'d the title. I was struck by that year too!
- ahefner 4y agoSometime around 1998/1999 (Q3A-era) I recall John Carmack commenting, probably either in a .plan entry or at Quakecon, that you could calculate arbitrary RenderMan shaders using ye old 1.1-ish fixed-function OpenGL given enough rendering passes, long before GLSL or even ARB_fragment_program was a thing. So the answer is probably longer ago than you'd think.
- PAPPPmAc 4y agoEarliest record I'm aware of was two limited-scope papers about rigging linear algebra systems onto GPUs in 2003: Krüger and Westermann out of TU Munich and Bolz et. al. out of Caltech. Both were published in the same 22(3) issue of IEEE Transactions on Graphics. Followed by the BrookGPU stuff out of Stanford that kind of cobbled a stream processor API out of OpenGL1.3 or DX9 calls in 2004. Earliest practical would presumably be the first round of GPUs with programmable shaders in ~2005 with the ATI R520 (X1000) and the nVidia G70 (7800GTX). Coming up on a decade later, ATI/AMD will surely commit to a cogent programming solution soon...