the upcoming Neoverse cores.
For what it's worth, when I looked into it more, I felt like the fact that you could get the count directly and use it as the loop increment alleviates some of my fears. So I am less skeptical of the V extension now than I was, although I still have concerns.
-
-
I'm still skeptical for a couple reasons. I need to see performance numbers. There's an alarming tendency to regard SIMD as a 'scientific computing throughput game' where a few extra cycles of latency are NBD because we're all doing massive matrix multiplies. ...
-
My taxonomy of SIMD is: 1) traditional SIMD where you're doing one thing over and over, 2) SIMD where the size of the thing you're operating on is big but not unbounded (i.e. "simulate a 256 state NFA") and 3) Uses of SIMD to get at weird functionality not on GPRs (e.g. PSHUFB).
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.

Actually, throughput/latency numbers for V1/N2 sooner rather than later would be great. To me this is more significant than the (laudable) simulator/software support. A lot of stuff I've designed lives and dies by these numbers.