Hacker News: eilnhttps://news.ycombinator.com/user?id=eilnHacker News RSShttps://hnrss.org/hnrss v2.1.1Sun, 13 Sep 2026 05:07:26 +0000<![CDATA[New comment by eiln in "Getting 50 GB/S Back from the Apple Neural Engine"]]>RTL performance erratum in the Apple M3 Neural Engine throttles DRAM weight streaming throughput down to 17–19 GB/s from the nominal 45–60 GB/s. Avoiding the problematic path in the kernel DMA engine's speculative prefetch ring increased Llama 3.2 1B token throughput from 10.0 to 24.3 tokens/s.

]]>Thu, 10 Sep 2026 00:16:10 +0000https://news.ycombinator.com/item?id=49636480eilnhttps://news.ycombinator.com/item?id=49636480https://news.ycombinator.com/item?id=49636480<![CDATA[Getting 50 GB/S Back from the Apple Neural Engine]]>Article URL: https://eiln.github.io/posts/ane-dma.html

Comments URL: https://news.ycombinator.com/item?id=49636479

Points: 108

# Comments: 22

]]>Thu, 10 Sep 2026 00:16:10 +0000https://eiln.github.io/posts/ane-dma.htmleilnhttps://news.ycombinator.com/item?id=49636479https://news.ycombinator.com/item?id=49636479<![CDATA[New comment by eiln in "Getting 50 GB/S Back Out of the Neural Engine"]]>RTL performance erratum in the Apple M3 Neural Engine throttles DRAM weight streaming throughput down to 17–19 GB/s from the nominal 45–60 GB/s. Avoiding the problematic path in the kernel DMA engine's speculative prefetch ring increased Llama 3.2 1B token throughput from 10.0 to 24.3 tokens/s.

]]>Wed, 09 Sep 2026 03:43:16 +0000https://news.ycombinator.com/item?id=49620643eilnhttps://news.ycombinator.com/item?id=49620643https://news.ycombinator.com/item?id=49620643<![CDATA[Getting 50 GB/S Back Out of the Neural Engine]]>Article URL: https://eiln.github.io/posts/ane-dma.html

Comments URL: https://news.ycombinator.com/item?id=49620642

Points: 3

# Comments: 1

]]>Wed, 09 Sep 2026 03:43:16 +0000https://eiln.github.io/posts/ane-dma.htmleilnhttps://news.ycombinator.com/item?id=49620642https://news.ycombinator.com/item?id=49620642