You’re right, it’s not the same die, but the advanced packaging techniques that they keep improving (like the vertical stacking you mention) make for a much tighter set of specs for the raw flash storage silicon compared to what they might be putting in USB drives or NVMe sticks, in power consumption/temperature management, bus speeds/latency, form factor, etc.

So it’d be more accurate to describe it as a system on a package (SiP) rather than a system on a chip (SoC). Either way, that carries certain requirements that aren’t present for a standalone storage package separately soldered onto the PCB, or even storage through some kind of non-soldered swappable interface.

Packaging flash storage onto the actual SoC SiP costs more than manufacturing the same amount of storage into an M.2 or external USB form factor, so that price can’t be directly compared. They’re making a big chunk of profit on storage upgrades, and on cloud subscriptions, but it’s not exactly cheap to give everyone 1TB of storage at that base price.

The problem is that there are too many separate dimensions to define the tiers.

In terms of data signaling speed and latency, you have the basic generations of USB 1.x, 2.0, 3.x, and 4, with Thunderbolt 3 essentially being the same thing as USB4, and Thunderbolt 4 adding on some more minimum requirements.

On top of that, you have USB-PD, which is its own standard for power delivery, including how the devices conduct handshakes over a certified cable.

And then you have the standards for not just raw data speed, but also what other modes are supported, for information to be seamlessly tunneled through the cable and connection in a mode that carries signals other than the data signal spec for USB. Most famously, there’s the DisplayPort Alt Mode for driving display data over a USB-C connection with a DP-compatible monitor. But there’s also an analog audio mode so that the cable and port passes along analog data to or from microphones or speakers.

Each type of cable, too, carries different physical requirements, which also causes a challenge on how long the cable can be and still work properly. That’s why a lot of the cables that support the latest and greatest data and power standards tend to be short. A longer cable might be useful, but could come at the sacrifice of not supporting certain types of functions. I personally have a long cable that supports USB-PD but can’t carry thunderbolt data speeds or certain types of signals, but I like it because it’s good for plugging in a charger when I’m not that close to an outlet. But I also know it’s not a good cable for connecting my external SSD, which would be bottlenecked at USB 2.0 speeds.

So the tiers themselves aren’t going to be well defined.

The only devices that don’t have at least Thunderbolt 3 on all ports do use the Thunderbolt logo on the ones that support it, except the short-lived 12-inch MacBook (non-Pro, non-Air). Basically, for data transfer:

• If it’s a 12-inch MacBook, the single USB-C port doesn’t support Thunderbolt, and only supports USB 3.1 Gen 1.
• In all other devices, if the ports are unmarked, they all support Thunderbolt 3 or higher
• If the ports are marked with Thunderbolt symbols, those ports support Thunderbolt but the unmarked ports on the same computer don’t.

For power delivery, every USB-C port in every Apple laptop supports at least first generation USB-PD.

For display, every USB-C port in every Apple laptop (and maybe even the desktops) supports DisplayPort alt mode.

It’s annoying but not actually that hard to remember in the wild.

Everything defined in the Thunderbolt 3 spec was incorporated into the USB 4 spec, so Thunderbolt 3 and USB 4 should be basically identical. In reality the two standards are enforced by different certification bodies, so some hardware manufacturers can’t really market their compliance with one or the other standard until they get that certification. Framework’s laptops dealt with that for a while, where they represented that their ports supported certain specs that were basically identical to the USB 4 spec or even the Thunderbolt 4 spec, but couldn’t say so until after units had already been shipping.

Ok so most monitors sold today support DDC/CI controls for at least brightness, and some support controlling color profiles over the DDC/CI interface.

If you get some kind of external ambient light sensor and plug it into a USB port, you might be able to configure a script that controls the brightness of the monitor based on ambient light, without buying a new monitor.

They chiplet past 500

I don’t know if I’m using the right vocabulary, maybe “die size” is the wrong way to describe it. But the Ultra line packages two Max SoCs with a high performance interconnect, so that the whole package does use about 1000 mm^2 of silicon.

My broader point is that much of Apple’s performance comes from their willingness to actually use a lot of silicon area to achieve that performance, and it’s very expensive to do so.

Apple does two things that are very expensive:

1. They use a huge physical area of silicon for their high performance chips. The “Pro” line of M chips have a die size of around 280 square mm, the “Max” line is about 500 square mm, and the “Ultra” line is possibly more than 1000 square mm. This is incredibly expensive to manufacture and package.
2. They pay top dollar to get the exclusive rights to TSMC’s new nodes. They lock up the first year or so of TSMC’s manufacturing capacity at any given node, at which point there is enough capacity to accommodate other designs from other TSMC clients (AMD, NVIDIA, Qualcomm, etc.). That means you can just go out and buy an Apple device made from TSMC’s latest node before AMD or Qualcomm have even announced the lines that will be using those nodes.

Those are business decisions that others simply can’t afford to follow.

The biggest problem they are having is platform maturity

Maybe that’s an explanation for desktop/laptop performance, but I look at the mobile SoC space where Apple holds a commanding lead over ARM chips from Qualcomm, and where Qualcomm has better performance and efficiency than Samsung’s Exynos line, and I’m thinking a huge chunk of the difference between manufacturers can’t simply be explained by ISA or platform maturity. Apple has clearly been prioritizing battery life and efficiency for 10+ generations of Apple Silicon in the mobile market, and has a lead independent of its ISA, even as it trickled over to the laptop and desktop market.

Well, specifically, they’re promising battery life that beats Qualcomm’s implementation of an ARM laptop SoC.

Qualcomm is significantly behind Apple. I’m not convinced that the ISA matters all that much for battery life. AMD’s x86_64 performance per watt blew Intel’s out of the water in recent generations, and Qualcomm/Samsung’s ARM chips can’t compete with Apple’s ARM chips in the mobile, tablet, or laptop space.

To be honest, no. I mainly know about JPEG XL only because I’m acutely aware of the limitations of standard JPEG for both photography and high resolution scanned documents, where noise and real world messiness cause all sorts of problems. Something like QOI seems ideal for synthetic images, which I don’t work with a lot, and wouldn’t know the limitations of PNG as well.

You say that it is sorted in the order of most significants, so for a date it is more significant if it happend 1024, 2024 or 9024?

Most significant to least significant digit has a strict mathematical definition, that you don’t seem to be following, and applies to all numbers, not just numerical representations of dates.

And most importantly, the YYYY-MM-DD format is extensible into hh:mm:as too, within the same schema, out to the level of precision appropriate for the context. I can identify a specific year when the month doesn’t matter, a specific month when the day doesn’t matter, a specific day when the hour doesn’t matter, and on down to minutes, seconds, and decimal portions of seconds to whatever precision I’d like.

Sometimes the identity of the messenger is important.

Twitter was super easy to set up with the API to periodically tweet the output of some automated script: a weather forecast, a public safety alert, an air quality alert, a traffic advisory, a sports score, a news headline, etc.

These are the types of messages that you’d want to subscribe to the actual identity, and maybe even be able to forward to others (aka retweeting) without compromising the identity verification inherent in the system.

Twitter was an important service, and that’s why there are so many contenders trying to replace at least part of the experience.

This isn’t exactly what you asked, but our URI/URL schema is basically a bunch of missed opportunities, and I wish it was better designed.

Ok so it starts off with the scheme name, which makes sense. http: or ftp: or even tel:

But then it goes into the domain name system, which suffers from the problem that the root, then top level domain, then domain, then progressively smaller subdomains, go right to left. www.example.com requires the system look up the root domain, to see who manages the .com tld, then who owns example.com, then a lookup of the www subdomain. Then, if there needs to be a port number specified, that goes after the domain name, right next to the implied root domain. Then the rest of the URL, by default, goes left to right in decreasing order of significance. It’s just a weird mismatch, and would make a ton more sense if it were all left to right, including the domain name.

Then don’t get me started about how the www subdomain itself no longer makes sense. I get that the system was designed long before HTTP and the WWW took over the internet as basically the default, but if we had known that in advance it would’ve made sense to not try to push www in front of all website domains throughout the 90"s and early 2000’s.

Your day to day use isn’t everyone else’s. We use times for a lot more than “I wonder what day it is today.” When it comes to recording events, or planning future events, pretty much everyone needs to include the year. Getting things wrong by a single digit is presented exactly in order of significance in YYYY-MM-DD.

And no matter what, the first digit of a two-digit day or two-digit month is still more significant in a mathematical sense, even if you think that you’re more likely to need the day or the month. The 15th of May is only one digit off of the 5th of May, but that first digit in a DD/MM format is more significant in a mathematical sense and less likely to change on a day to day basis.

Functionally speaking, I don’t see this as a significant issue.

JPEG quality settings can run a pretty wide gamut, and obviously wouldn’t be immediately apparent without viewing the file and analyzing the metadata. But if we’re looking at metadata, JPEG XL reports that stuff, too.

Of course, the metadata might only report the most recent conversion, but that’s still a problem with all image formats, where conversion between GIF/PNG/JPG, or even edits to JPGs, would likely create lots of artifacts even if the last step happens to be lossless.

You’re right that we should ensure that the metadata does accurately describe whether an image has ever been encoded in a lossy manner, though. It’s especially important for things like medical scans where every pixel matters, and needs to be trusted as coming from the sensor rather than an artifact of the encoding process, to eliminate some types of error. That’s why I’m hopeful that a full JXL based workflow for those images will preserve the details when necessary, and give fewer opportunities for that type of silent/unknown loss of data to occur.

• Existing JPEG files (which are the vast, vast majority of images currently on the web and in people’s own libraries/catalogs) can be losslessly compressed even further with zero loss of quality. This alone means that there’s benefits to adoption, if nothing else for archival and serving old stuff.
• JPEG XL encoding and decoding is much, much faster than pretty much any other format.
• The format works for both lossy and lossless compression, depending on the use case and need. Photographs can be encoded in a lossy way much more efficiently than JPEG and things like screenshots can be losslessly encoded more efficiently than PNG.
• The format anticipates being useful for both screen and prints. Webp, HEIF, and AVIF are all optimized for screen resolutions, and fail at truly high resolution uses appropriate for prints. The JPEG XL format isn’t ready to replace camera RAW files, but there’s room in the spec to accommodate that use case, too.

It’s great and should be adopted everywhere, to replace every raster format from JPEG photographs to animated GIFs (or the more modern live photos format with full color depth in moving pictures) to PNGs to scanned TIFFs with zero compression/loss.

Adobe is backing the format, Apple support is coming along, and there are rumors that Apple is switching from HEIC to JPEG XL as a capture format as early as the iPhone 16 coming out in a few weeks. As soon as we have a full blown workflow that can take images from camera to post processing to publishing in JXL, we might see a pretty strong push for adoption at the user side (browsers, websites, chat programs, social media apps and sites, etc.).

Semiconductor manufacturing has gotten better over time, with exponential improvement to transistor density, which translates pretty directly to performance. This observation traces back to the 60’s and is commonly known as Moore’s Law.

Fitting more transistors into the same size space required quite a few technical advancements and paradigm shifts. But for the first few decades of Moore’s law, every time they started to approach some kind of physical limit, they’d develop a completely new technique to get things smaller: photolithography moved from off the shelf chemicals purchased from photography companies like Eastman Kodak to specialized manufacturing processes, while the light used went higher and higher wavelength, with the use of new technology like lasers to get even more precisely etched masks.

Most recently, the main areas of physical improvement has been in using extreme ultraviolet (aka EUV) wavelengths to get really small features, and certain three dimensional structures that break out from the old paradigm of stacking a bunch of planar materials on each other. Each of these breakthroughs was 20 years in the making, so the R&D and the implementation details had to be hammered out with partners in a tightly orchestrated process, to see if it would even work at scale.

Some manufacturers recognized the huge cost and the uncertainty of success in taking stuff from academic papers in the 2000s and actually mass producing chips in 2025, so they abandoned the leading edge. Global Foundries, Micron, and a bunch of others basically decided it wasn’t worth the investment to try to compete, and now manufacture in those older nodes, without necessarily trying to compete on the newest nodes, leaving things to Intel, Samsung, and TSMC.

TSMC managed to get EUV working at scale before Intel did. And even though Intel beat TSMC to market with certain three dimensional structures known as “FinFETs,” the next 2 generations after that, TSMC managed to really shove them in there at higher density, by using combining those FinFETs with lithography techniques that Intel couldn’t figure out fast enough. And every time Intel seemed to get close, a new engineering challenge would stifle them. And after a few years of stagnation, they went from being consistently 3 years ahead of TSMC to seeming like they’re about 2 years behind TSMC.

On the design side of things, AMD pioneered chiplet-based design, where different pieces of silicon could be packaged together, which allowed them to have higher yields (an error in a big slab of silicon might make the whole thing worthless) and to mix and match things in a more modular way. Intel was slow to adopt that, so AMD started taking the lead in CPU performance per watt.

It’s difficult engineering challenges, traceable back to decisions made in the past decades. Not all of the decisions were obviously wrong at the time, but nobody could’ve predicted at the time that TSMC and AMD would be able to leapfrog Intel based on these specific engineering challenges.

Intel has a few things on the roadmap that might allow it to leapfrog the competition again (especially if the competition runs into their own setbacks). Intel is ramping up use of EUV in its current processes, are ramping up a competing three dimensional structures they call RibbonFET to compete with TSMC’s Gate All Around (both of which are supposed to replace FinFETs) and they’re hoping to beat TSMC to backside power delivery, which is going to represent a significant paradigm shift in how chips are designed.

It’s true that in business, success begets success, but it’s also true that each new generation presents its own novel challenges, and it’s not easy to see where any given competitor might get stuck. Semiconductor manufacturing is basically wizardry, and the history of the industry shows that today’s leaders might get left behind, really quickly.

AMD doesn’t have fabs, and contracts that work out to TSMC. Intel has fabs worldwide. That alone makes Intel a much more capital-intensive, labor-intensive business.