[HN Gopher] MeMPA: A Memory Mapped M-SIMD Co-Processor to Cope w...
___________________________________________________________________
MeMPA: A Memory Mapped M-SIMD Co-Processor to Cope with the Memory
Wall Issue
Author : PaulHoule
Score : 29 points
Date : 2024-02-26 22:40 UTC (2 days ago)
(HTM) web link (www.mdpi.com)
(TXT) w3m dump (www.mdpi.com)
| mpweiher wrote:
| Ahhh...memories of PixelPlanes[1][2] and SLAM[3].
|
| This has always seemed like such a good idea, I wonder what the
| factors are that keep it from being successful, and when/if its
| benefits ever overcome those factors.
|
| [1] https://dl.acm.org/doi/10.1145/74333.74341
|
| [2] http://www.cs.unc.edu/~fuchs/publications/PixelPlanes4.pdf
|
| [3] https://dl.acm.org/doi/book/10.5555/13358
| mikewarot wrote:
| Inhomogeneity is always the thing that keeps these
| architectures from being successful, in the end. It's a
| premature optimization that _seems_ like the right thing to do
| at any given moment. Then conditions change, you find out that
| FP32 isn 't needed, and FP8 is ok... then 3 bits... then
| Trinary... all require new hardware.
|
| Memory that computes directly, in as simple a model as possible
| is required. Imagine a huge grid of LUTs that just talk to
| their neighbors, and I believe, you've got an approach that can
| work for almost any compute task.
___________________________________________________________________
(page generated 2024-02-28 23:01 UTC)