afaiu morsel-driven means the workload gets turned into 'smallish' chunks (morsels)

instead of having to pre-allocate upfront (e.g. 4 nodes get 1/4 each) it is more granular and dynamic

a worker that's "done" can request another morsel

pragmatic approach because nodes might not all be equally fast (cache, cpu frequency, throttling, …) and also some morsel workloads take longer than others depending on the values they contain and what kind of work needs to get done

so this approach tends to balance out nicely

I'm sure someone else can explain it better / correct me (please do!)

When I read up, it sounded like the same idea as work-stealing to me. Not surprising that different fields come up with the same idea under different terminology.

DuckDB and LadybugDB use the same terminology to describe internals.

Exactly!

finally a snow leopard like release focused on performance and fixing user experience

It's just being pitched this way by marketing and the C suite. If it were really a snow leopard release, someone should have informed the engineers they were supposed to be improving resiliency and fixing bugs, because this is news to them. cough

My impression was that they're doubling down on that horrid Liquid glass.

Apparently there's a new fancy slider for making it more (but not completely) opaque? Did I miss an option for turning it off?

[accessibility settings -> display -> reduce transparency] is the main option afaik. while you're in there, try "reduce motion" too, it's pretty nice imo.

Keep in mind Apple would never admit mistakes on Liquid Glass. But: Looks to me they're fixing some of the worst aspects. I'm on the fence.

The iOS 7 flat redesign was a UX disaster. But they got back up to speed in subsequent releases.

There IS something to be said for design resets with follow-up refits to accomodate for actual human beings. Most companies just add crap on top of crap.

Not saying what everything Apple does is perfect, even as a user/fanboy since '86.

What I most enjoyed about todays's annoucement that they're doing a Snow Leopard performance/bug reset, because that was expected and needed. And they started out with it, so they know their WWDC audience.

So: Both a technical and UX debt effort, with some privacy-focused AI on top.

I can't complain.

Have you tried the Accessibility setting "Reduce Transparency"? Apple tends to hide too many things there.

that's been available from the beginning

> It seems entirely possible that, in due course, electronic or chemical "machines" will outdo the human brain in most of the functions we now consider exclusively within its province.

nb: poster is Daniel Lemire (https://lemire.me), who is very skilled in getting performance out of compute hardware (e.g. via simd, cache usage etc)

As he likes to share often, "He ranks among the top 2% of scientists globally (Stanford/Elsevier 2025) and is one of GitHub's top 1000 most followed developers. "

based on citations and github stars? or what's the context there?

I was adding further citation based on his own claims. Not sure what context is missing.

Still, Microslop has repeatedly proven their ability to slow everything down to a crawl no matter how powerful the hardware. If you want it to be fast, don’t use Windows.

latency and bandwidth

Is there any data on whether Google, Amazon, Microsoft, Anthropic, OpenAI etc are most cost efficient in getting datacenter compute online and operating it?

I'd be interested in how large the range is here across company and region and specific data center and how it relates to companies like Hetzner if at all.

Well, Elon seems to take the fastest path possible to these DCs. One can envision a future where these get shut down for the severity of the pollution, not to mention being built and operated illegally [0].

[0] https://www.selc.org/news/xai-built-an-illegal-power-plant-t...

Which is precisely why there has been a push to weaken the EPA and other regulating agencies

> Is there any data on whether Google, Amazon, Microsoft, Anthropic, OpenAI etc are most cost efficient in getting datacenter compute online and operating it?

Well considering that ~80% of the price is hardware deprecation, I don't know why they'd be considerably worse than anyone else at negotiating hardware deals.

Typically when you buy in bulk, you have more sway.

Companies like Google also have in-house chips like TPUs that are substantially cheaper for inference for them to make than anyone else can get through Nvidia.

Amazon / AWS is very efficient at it. They are the largest cloud by a significant margin. Much larger than Hetzner.

Google and MS may be close behind.

Not sure about Meta but they are also renting from Amazon so...

Anthropic mostly rents from Amazon, Goog & MS.

I’ve seen some numbers related to datacenters in Ireland and they would stress price per MW as a way to see where to build them. But then you have depreciation of equipment as well. Depreciation can be played with when filing taxes though.

I don't think they are most efficient for small GPUs. I think they might only be the one which have capex and certainty required for multimillion dollar purchase of GB200 NVL72 or something of that scale.

that's asking the cart before the horse; is there any data on what compute actually results in real GDP improvements?

Nothing other than vendor promises and white papers.

ty, I always forget about the codeberg migration

ty for this writeup Andrew, all the best to you

very interesting, do you have a pointer with more info on what kind of SIMD support it has?

Hopefully comparable or better than ESP32S3.

But with the weird alignment thing fixed

The same as P4 rev 3.x, which is still undocumented. Assembler source and esp-nn/esp-dsp are your friend. Some people have also tried some stuff. And HP core 0 is scalar-only now.

Why on earth SIMD instead of the risc-v vector extensions that are supposed to be better?

For compatibility and simplicity. Most SIMD instructions in P4 and S31 (compatible with P4 rev 3.x) are an direct evolution from S3. Espressif just doesn't want to rewrite their optimized assembly libraries.