On Wed, Feb 21, 2024 at 2:13 PM Richard Fateman <fate...@gmail.com> wrote:
> One rationale for NaNs is based on the assumption that you might have > pipelined/ vector/ etc > computers where an "interrupt" in impossible -- it may happen at a time > and place that no longer exists. > So the NaN is carried along and as Stavros says, it colors all future > results as NaNs. > Since the NaN-ness of the data is encoded in the exponent, the fraction > part can be a payload of > some sort, explaining what happened. If that is possible. e.g. the > program counter at the time... > I don't think anyone does that. And AFAIK, all cpus except x87 and alpha had precise traps so you knew exactly what instruction caused the interrrupt. IIRC, for x87, the traps happened at the NEXT FPU instruction. But for x86 sse and friends, the traps happen at the instruction that caused it. > > On another topic, it was my understanding that stuff like tensorflow cpu > float precision was way way > lower than IEEE. Like 10 bit mantissa. I don't know about the various > GPU instruction sets, but > it seems like way overkill for graphics to have double (or even full > single) precision. But who > Probably half precision with 5 bits for the exponent and 11 bits for the mantissas (with a hidden bit). See https://en.wikipedia.org/wiki/Half-precision_floating-point_format uses GPUs for graphics. > My coworkers across the hall worked on WebGPU, so yeah, they spent a lot of time using the GPU to do graphics. :-) -- Ray
