[HN Gopher] MIMD Interpretation on a GPU [pdf]
___________________________________________________________________
MIMD Interpretation on a GPU [pdf]
Author : ColinWright
Score : 23 points
Date : 2022-10-07 17:03 UTC (5 hours ago)
(HTM) web link (aggregate.ee.engr.uky.edu)
(TXT) w3m dump (aggregate.ee.engr.uky.edu)
| HappySweeney wrote:
| > The toolchain is expected to become a public domain source code
| release as soon as it has reached an appropriate level of
| usability and stability.
|
| So, never?
| dragontamer wrote:
| So first off, GPUs are MIMD, in that each compute-unit /
| symmetric multiprocessor is independent of one another.
|
| NVidia's 3xxx line of GPUs are further 2-instructions per clock
| tick / superscalar, which is more "traditional" MIMD, in that the
| cores are now doing 2 different things per clock tick on top of
| all the SIMD-goodness they're doing. AMD is mostly SIMD, but
| arguably MIMD from the perspective that it could be executing a
| "scalar" instruction (branch / loop instructions) in parallel
| with a "vector" instruction (add/subtract/multiply)
|
| ----------
|
| So we can say that the assembly-language model of NVidia PTX /
| AMD RDNA/CDNA is SIMD, but the execution is now MIMD.
|
| This pdf would then add a virtual machine on top of that, giving
| us an MIMD virtual-machine that compiles into SIMD instructions
| that's executed on MIMD hardware.
|
| I dunno where I'm going with this, I'm just amused by it, that's
| all. Each step makes sense but its kinda funny when you look at
| it all from the top down.
___________________________________________________________________
(page generated 2022-10-07 23:01 UTC)