Getting 50 GB/S Back from the Apple Neural Engine(eiln.github.io) 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. |
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.
That got a laugh out of me.
For example, if I browse HN, I open the linked article in a new tab instead of in the same one where the comments are. And I close the tab as soon as I’m done looking at it. Whether that be a couple of seconds because of annoying dark patterns, or a long time thanks to being actually readable and interesting. From that tab I might navigate to another page on the same site if the site seems good. If they then pull horse crap with the browser history in that tab I will most likely close the tab and not read anything more on that site. If I repeatedly get annoyed by a domain from having gone to it and found dark patterns I will eventually remember that the whole site is bad and stop clicking any links to it either. But mostly I don’t have to remember so many sites to actively avoid. Closing a tab is cheap when all of the history is from the same site that tab was opened for in the first place.
It’s unfortunate and annoying that they get to play tricks with dark patterns still, and I wish nobody did that / could do that in the first place. In general I mean. I dont know if this particular page does it intentionally or accidentally.
I usually try to not comment on it even if I do come across that. As I know many consider it off topic, and even the HN guidelines say to not comment on it [1]. But I still appreciate when others, like your comment here, point those things out and I then don’t click on to the linked article in the first place when I see comments about that unless other comments are indicating that the content itself is particularly interesting.
[1]: https://news.ycombinator.com/newsguidelines.html
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
Sure, they're somewhat common issues, but so are ads for malware served by Google. Surely we should call those out?
In this particular instance, it works fine for me on Chrome. When I test it in Safari, I can see the same behaviour that people are complaining about, but I also find that if you double-click the back button, it takes you back to HN.
I point that out because it seems clear to me that the site author has not set out to implement a “dark pattern” to trick visitors into staying on their site against their will; they've just tried to implement some in-page navigation, and perhaps only tested it in Chrome.
That's outside the boundary of reasonable grounds for us to penalize a page or site. It puts us in the position of having to investigate a page's behaviour in different browsers and to adjudicate on the severity of the dysfunctional browser behaviour and the author's intent (i.e., their inner thoughts). We clearly can't be getting into that; it takes up too much time and we will never be able to consistently reach fair decisions.
I can understand the frustration, but as well as asking HN users not to fixate on offtopic distractions like this, we also ask people to "assume good faith", and I see no reason not to follow that guideline in this case.
A quick email to the author would have more likely led to a favourable outcome than this meta-subthread.
Sites are always going to do malicious shit. The browser should know that when I hit the back button I want to go to the place I clicked that brought me here. SPAs have severely perverted this simple idea, but I’m sure browser devs can figure it out.