Nice investigation.It is always surprising to me when a nice round number like 1MiB results in the “bad performance” configuration (although it happens).Are you sure erratum is the right word in this context? I usually see it used to describe the notice that a document has an error in it.

&quot;Apparently the memory controller&#x27;s throughput has a dominant harmonic with wavelength 2048 in tensor-dimension space.&quot;That got a laugh out of me.

Not all systems affected
M1 and M5MAX are OK<a href="https:&#x2F;&#x2F;x.com&#x2F;anemll&#x2F;status&#x2F;2098454204478366132?s=20" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;anemll&#x2F;status&#x2F;2098454204478366132?s=20</a>

I guess any processes affected by this will solve this in software as time goes on?

Just checking here- this systemverilog is a hypothetical telling of what you think is going on? Or do you have the actual source of the RTL?

This goes way over my head and I don&#x27;t understand most of it lol. I noticed you&#x27;re still in the middle of getting your B.S. degree and you&#x27;re already writing things like this...amazing.

This website hijacked my back button during a simple page load. You should fix that, it’s not an acceptable way to behave.

Damn bots copy and pasting a single sentence from the article as a comment.

RTL performance erratum in the Apple M3 Neural Engine throttles DRAM weight streaming throughput down to 17–19 GB&#x2F;s from the nominal 45–60 GB&#x2F;s. Avoiding the problematic path in the kernel DMA engine&#x27;s speculative prefetch ring increased Llama 3.2 1B token throughput from 10.0 to 24.3 tokens&#x2F;s.

HN

Getting 50 GB/S Back from the Apple Neural Engine

Comments (29)