insane RAM prices: it's not just AI

[DiTBho]

and now there is also a "Moto G86 Power"  with 24GB RAM, 8GB + 16GB boost, 24GB soldered 
The price is not that insane, ~ 400 euro

But that crazy thing now has more RAM than my computers

No, it only has 8 GB of RAM (which is already pretty beefy for a phone). The extra 16 GB is just flash memory reserved as extra virtual RAM, in other words swapping. It's just that now flash (in form of SSDs) has become so fast that swapping doesn't slow down the whole thing significantly has it used to. Mobile device manufacturers now market this feature as though it was some crazy technology.

damn, I was convinced it was RAM, not flash used as some sort of swap, virtual memory, or cache, so considering that (stock) Android can still run smoothly at 4GB of ram, I am afraid this feature doesn't make any difference in terms of performance.

Corollary: I guess, it's basically a marketing gimmick (1)

Moto G86, 8Gbyte ram -> 250 euro
Moto G86, 8Gbyte ram + 4Gbyte boost -> 300-320 euro
Moto G86, 8Gbyte ram + 16Gbyte boost -> 400 euro

[SiliconWizard]

Well, it is marketing gimmick. The only physical RAM you have is 8 GB, and it's the same for all mobile devices with such "RAM boost" or other marketing term.

To dig further, it's usually not just swapping to flash, but a combination of swapping and fast on-the-fly data compression to compress less often used data in RAM itself so that more of it can be stored in physical RAM and there's a little overhead when it needs accessed again, just like some kind of glorified cache scheme.

[abeyer]

everything old is new again

[brucehoult]

everything old is new again

It's never gone away.

MacOS has had RAM compression by default since Mavericks and iOS since version 7, both in 2013. Android 4.4 and mainline Linux kernel adopted zram/zswap in the same year, though distros have been slow to enable it — Fedora 33 did in 2020 but Debian and Ubuntu never have, you have to enable it yourself. Windows 10 has used compressed RAM since 2015.

[jonovid]

video games industry has been hit hard by this.  next gen video game consoles are now way too expensive for most of us.

[Siwastaja]

Sure, raw read speed is cool, but RAM isn't read-only. Flash is slower to write and something has to managed pages and stuff.

But it isn't that much slower to write, it's the same order of magnitude, for write-speed optimized design in the same speed ballpark. The trick is small enough erasable unit, large enough memory in itself so that operations parallelize, and a capable controller (correct shape of wear leveling and pre-erasing).

When SSDs came ~15 years ago, they tackled this problem. Memory cards were slow-write, SSDs were not - almost the same type of NAND flash inside. I believe the main fix was the sheer size - the parallelism it gave. It's the same mechanism how EVs with larger batteries have linearly faster charge speed given the same battery technology - only car journalists are idiot enough to look at the "minutes from empty to full" spec instead of "kilometers charged per minute". With disks, we are sensible enough to look at the GB/second spec, which improves linearly with larger number of flash chips behind a capable enough controller.

[PlainName]

You win. Forty-two is the number of angels on a pinhead.

[DiTBho]

Well, it is marketing gimmick. The only physical RAM you have is 8 GB, and it's the same for all mobile devices with such "RAM boost" or other marketing term.

To dig further, it's usually not just swapping to flash, but a combination of swapping and fast on-the-fly data compression to compress less often used data in RAM itself so that more of it can be stored in physical RAM and there's a little overhead when it needs accessed again, just like some kind of glorified cache scheme.

But is this "marketing gimmick" worth having? you see, it costs more money, it doesn't come for free.
My use-case is not "smartphone for gaming", and I don't run any AI-stuff on my smartphone.
Except AI-based (speech to text, and text to speech) application that runs entirely on the phone and teaches you Norwegian.

On my smartphone, I have and use only these apps:

• browser, with no more than 12 tabs open
• Google maps
• Google mail, Gmail
• bike reservation (useful, when I travel)
• train ticket reservation
• hotel reservation
• pet food reservation
• book reservation (public library service)
• Weather (one-time paid application, usable forever, which works well, except in Sweden... where it convinced me to go out without an umbrella, and then I found myself showering outside...)
• Amazon (I buy a lot of eBooks and physicl books here)
• eBay
• Learn Norwegian (it's ~ 900Mbyte + dictionary, the biggest one)
• Booking a postal service, whether to send packages or have them delivered to a locker. In both cases, it saves you from having to wait in line at the post office
• Turn on the front light. A simple app that does just that, and it's damn useful when the power goes out, and damn frustrating when you buy a new phone and find that Motorola hasn't included this app in its firmware, so the need to buy it from the store for 2 euros
• BuyCycle (buying/selling bicycle parts). They have a shipping service in partnership with DHL. The app guides you through the preparation of your package, or even sends you all the packaging materials if you're selling a complete bike or wheels. It's fantastic because it handles everything, from tracking to payment, from instructions to support, without you having to go out and use other apps. The problem is that it consumes several hundred MB of data when running and is most likely not optimized.)
• Garmin Connect (edit or plan maps for my Garmin Edge 830, useful for both training and visiting new places by bike)
• Plastic bottle recycling. There's a machine near every train station and supermarket. You bring your smartphone and bottles with you. Suitable bottles are water, tea, and milk bottles. They must not be broken and must have a label. Once you open the app on the phone, the machine recognizes your "user ID" and allows you to insert them one bottle at a time. Then it reads the barcode on the bottle label and, if it's acceptable, it awards you "points," which you can use to enter a draw for prizes like electric bikes (1000 UKP), backpacks (90UKP) with hydration bladders—great for jogging but also for more challenging activities—and sport watches

Considering the 8GB of RAM on the Moto G86 phone, I don't think any of these apps will benefit from the "RAM boost" (flash-cache) mechanism.

I mean, I can save money and buy the cheapest G86 Power in my previous list 

[tszaboo]

Well, it is marketing gimmick. The only physical RAM you have is 8 GB, and it's the same for all mobile devices with such "RAM boost" or other marketing term.

To dig further, it's usually not just swapping to flash, but a combination of swapping and fast on-the-fly data compression to compress less often used data in RAM itself so that more of it can be stored in physical RAM and there's a little overhead when it needs accessed again, just like some kind of glorified cache scheme.

But is this "marketing gimmick" worth having? you see, it costs more money, it doesn't come for free.
My use-case is not "smartphone for gaming", and I don't run any AI-stuff on my smartphone.
Except AI-based (speech to text, and text to speech) application that runs entirely on the phone and teaches you Norwegian.

On my smartphone, I have and use only these apps:

• browser, with no more than 12 tabs open
• Google maps
• Google mail, Gmail
• bike reservation (useful, when I travel)
• train ticket reservation
• hotel reservation
• pet food reservation
• book reservation (public library service)
• Weather (one-time paid application, usable forever, which works well, except in Sweden... where it convinced me to go out without an umbrella, and then I found myself showering outside...)
• Amazon (I buy a lot of eBooks and physicl books here)
• eBay
• Learn Norwegian (it's ~ 900Mbyte + dictionary, the biggest one)
• Booking a postal service, whether to send packages or have them delivered to a locker. In both cases, it saves you from having to wait in line at the post office
• Turn on the front light. A simple app that does just that, and it's damn useful when the power goes out, and damn frustrating when you buy a new phone and find that Motorola hasn't included this app in its firmware, so the need to buy it from the store for 2 euros
• BuyCycle (buying/selling bicycle parts). They have a shipping service in partnership with DHL. The app guides you through the preparation of your package, or even sends you all the packaging materials if you're selling a complete bike or wheels. It's fantastic because it handles everything, from tracking to payment, from instructions to support, without you having to go out and use other apps. The problem is that it consumes several hundred MB of data when running and is most likely not optimized.)
• Garmin Connect (edit or plan maps for my Garmin Edge 830, useful for both training and visiting new places by bike)
• Plastic bottle recycling. There's a machine near every train station and supermarket. You bring your smartphone and bottles with you. Suitable bottles are water, tea, and milk bottles. They must not be broken and must have a label. Once you open the app on the phone, the machine recognizes your "user ID" and allows you to insert them one bottle at a time. Then it reads the barcode on the bottle label and, if it's acceptable, it awards you "points," which you can use to enter a draw for prizes like electric bikes (1000 UKP), backpacks (90UKP) with hydration bladders—great for jogging but also for more challenging activities—and sport watches

Considering the 8GB of RAM on the Moto G86 phone, I don't think any of these apps will benefit from the "RAM boost" (flash-cache) mechanism.

I mean, I can save money and buy the cheapest G86 Power in my previous list 

12 tabs, 500MB each.
A programmer went and actually said "OK looks good, let's deploy it"
And then we are in a situation where the Microsoft Windows weather widget uses 1.2 GB of RAM, has a splash screen because it was loading so slow, shows you ads and probably asks for your Login on windows where you are already logged in. And has 2 updates a day.
If it doesn't show, I'm a bit fatigued by the absolute shit quality bloated software we get everywhere.

[DiTBho]

12 tabs, 500MB each.
A programmer went and actually said "OK looks good, let's deploy it"
And then we are in a situation where the Microsoft Windows weather widget uses 1.2 GB of RAM, has a splash screen because it was loading so slow, shows you ads and probably asks for your Login on windows where you are already logged in. And has 2 updates a day.
If it doesn't show, I'm a bit fatigued by the absolute shit quality bloated software we get everywhere.

me too, but it's an Andoid Phone, not a Windows Phone.

I don't think the browser tabs consume more than 30 MB because I don't browse strange sites, and a few tabs are for Gemini. One of the sites I visit is EEVblog, which doesn't have much JavaScript or other heavy stuff.

The two apps that take up the most space are the Norwegian course, because it preloads a lot of audio files, and the BuyCycle app, which I believe has NOT been optimized.

The weather app is paid, and it's not bad. It doesn't bother anyone; only when you ask for an update by clicking on its button, it connects to the server, downloads data to display elevation maps, and provides forecasts.

Garmin Connect consumes up to 900MB of data when preloading an activity map for planning, which is smart considering the size of the preloaded maps on my Garmin Edge 830. I mean, regional map files, such as Garmin Cycle Map Europe, typically take up 7-12GB depending on active map updates.

There's a lot of poorly written software for Android, too, but I tried to narrow it down and discarded anything that was annoying with ads, bugs, or that needed to use up too much RAM.

-> purge games!
-> purge AI-garbage!

[Siwastaja]

"Let's put some extra Tensor Processing Unit and 8GB more RAM on your phone so that it can do AI things".

The AI runs completely in the cloud. No AI runs on the AI phone.

What a joke. Just a made-up excuse for the usual and predictable bloat curve going upward.

[SiliconWizard]

Local "AI" (what they call "at the edge") is in practice rarely used (because there just isn't any money in that), so yeah, that's BS. It can be used for some tasks such as face recognition or voice recognition, but these require much lighter models than any kind of LLM and certainly do not require gigabytes of RAM to run.

[DiTBho]

Alien, in its simplified version, is available for PlayStation 1 as well as for the Nintendo GBA; the game is based on artificial intelligence, meaning a predictive engine tries to predict your next move to provide the aliens with the information they need to hunt you down.

Modern versions use improved predictive engines.

PSX (MIPS/R3000): takes 500Kbyte
GBA (ARM7TDMI): takes 300Kbyte

[DiTBho]

but the real problem is not that the modern predictive algorithm consumes more RAM and CPU: sure, they certainly consume more than the consoles of 2000, the problem is that modern stuff puts a lot of advertising players and useless crap.

[DiTBho]

voice recognition

The Norwegian course has a voice recognizer that recognizes what you say and corrects you.
It works in aeromode, meaning with Wi-Fi completely turned off, so it does NOT require a cloud service.

Love it 

[abeyer]

Local "AI" (what they call "at the edge") is in practice rarely used (because there just isn't any money in that)

Google's fairly aggressive push to expose the ability in chrome & web apis seems to imply that's changing, or at least they want it to. I'm not sure how much adoption it has yet, but if it's available to web pages I suspect they'll start using(/abusing) it.

[Someone]

Local "AI" (what they call "at the edge") is in practice rarely used (because there just isn't any money in that), so yeah, that's BS. It can be used for some tasks such as face recognition or voice recognition, but these require much lighter models than any kind of LLM and certainly do not require gigabytes of RAM to run.

Between universal do anything frontier models that need a cluster of high end GPUs, and small edge inference such as wake on word/phrase, there are still useful things in the middle. Just want some spelling/grammar/tone suggestions? perfect.

[kite31]

More than that, their effectiveness at pattern matching at a token level can be very useful. I had two related files where I needed to know which lines corresponded, which were unrelated, and which should have a companion but did not. There was no useable key on the files and column headings which were supposed to correspond had differences of expression (shortened words, different separators). Rather than cleaning up the file (Easy Data Transform is useful) I handed the two files to a local model with instructions as described in my second sentence above. It quickly worked out what what effectively matched in both column and row terms, discarded irrelevant column data and gave me a condensed side-by-side view from which I could make the necessary fixes in the source data.