Raspberry Pi RAM Restrictions No Big Deal, Frankly

September 22, 2026 by Arya Voronova 33 Comments

Hacking on Raspberry Pi board internals is one of my favourite topics. I know a bunch of obscure things about these cute little boards. Three years ago, I covered a Raspberry Pi 4 RAM upgrade story. Getting a BGA RAM chip and swapping it in seemed like a no-brainer to me – apart from all the numerous uncertain parts about it, you know. It was a joy to see hackers pull it off, and for it to function as well as it did!

Things changed. You can’t really get RAM chips anymore. You also can’t get RAM sticks. You can’t get even SSDs with RAM chips on them. Even getting Raspberry Pi boards can be hard unless you know where to look. This is where a recent three-minute video by [Jeff Geerling] finds us.

Turns out, Raspberry Pi Foundation pushed binary-blob bootloader changes that limit your ability to upgrade RAM. I’ve known about it since last year through the grapevine, and somehow, as I read about it, this didn’t bother me at all. Not enough to write a Hackaday article about it, even, much less talk about it more widely. Why didn’t it bother me? Today, I sat down and pondered this for a bit.

Here’s my conclusion: I don’t think it’s a big deal at all, even if it seems that many people would disagree. Come in, as you are, and I hope you find my thoughts on the situation entertaining.

Continue reading “Raspberry Pi RAM Restrictions No Big Deal, Frankly” →

LLM Moats Quickly Evaporating

August 30, 2026 by Bryan Cockfield 55 Comments

In the business world, a moat is a quality of a business that makes it difficult for competitors to take that company’s profits. With how hard it is to train models for large language models (LLMs) and generative AI, it might seem like Anthropic, Open AI, and other LLM companies would have huge moats given the amount of compute it takes to build models. But open source models are quickly draining that moat, and now the only thing standing in the way of a customer using one of these models on their own hardware instead one from the larger companies is physical computing resources. [TerminalBytes] demonstrates a few of these models on personally owned computers to show the current state of the art.

[TerminalBytes] started off running the 27B version of the Qwen3.8 on a Mac Studio with 256 GB of unified RAM, which is plenty for this task. But it’s also enough to benchmark a few different models. Qwen3.6 is compared to 3.8, and then the different quants of each model are also compared. Quants are compressed versions of models that need fewer bits to store weights, meaning that the same models can run in less memory with smaller losses in fidelity. Many of these quants run on machines with 32 GB of RAM or less, encompassing many average gaming PCs. There’s even a 1-bit quant that [TerminalBytes] tested which can easily run on a machine with 16 GB, although with mixed results.

Keep in mind that this is just the current state of affairs with open LLMs. Future versions of these models are likely to optimize the number of tokens produced per unit time, or otherwise increase quality of responses while requiring less computer resources. We don’t really think that the ease of running local models will be the sole reason that the AI bubble pops, though. The fact that not every computer user is running Linux is proof enough of that.

RAM And EMMC Prices Are So High, Pine64 Has Stopped Linux Device Production

August 28, 2026 by Jenny List 35 Comments

We all know that the price of RAM and storage has shot up due to demand from the AI industry and other factors. In most cases that means we grumble about the price, but if we really need the part we can fork out for it. [CNX Software] are reporting that rather than continue to push up their prices, Pine64 are responding to the crisis by halting production of their Linux boards for the time being.

We’ve seen online comment over the now-exorbitant cost of other boards such as a fully-loaded Raspberry Pi, and this follows in that vein. If we had to guess we’d speculate that the high prices have resulted in too little in the way of sales, which considering the knock-on impact on our community if other vendors follow suit, could be concerning.

If there’s one bright spot in the current situation, it’s that for many applications where a single-board computer might be used, a microcontroller might now be a better choice for the job than something running Linux. We’re in a very different situation from that we were in when cheap Linux boards appeared, the current generation of high-power microcontrollers have significantly closed the gap between the two. Given that microcontrollers have onboard memory and storage, their immunity from the price hikes makes them even more attractive. As to Pine64, we hope that sales of their other products make up for it.

Header image: BasilicumTree, CC0.

Hackaday Links: March 15, 2026

March 15, 2026 by Tom Nardi 4 Comments

Some days, it feels like we’re getting all the bad parts of cyberpunk and none of the cool stuff. Megacorps and cyber warfare? Check. Flying cars and holograms? Not quite yet. This week, things took a further turn for the dystopian with the news that a woman was hospitalized after an altercation with a humanoid robot in Macau. Police arrived on scene, took the bot into custody, and later told the media they believed this was the first time Chinese authorities had been called to intervene between a robot and a human.

The woman, reportedly in her seventies, was apparently shocked when she realized the robot was standing behind her. After the dust settled, the police determined it was being operated remotely as part of a promotion for a local business. We’ve heard there’s no such thing as bad publicity, but we’re not sure the maxim holds true when you manage to put an old lady into the hospital with your ad campaign.

Continue reading “Hackaday Links: March 15, 2026” →

Capacitor Memory Makes Homebrew Relay Computer Historically Plausible

March 5, 2026 by Tyler August 21 Comments

It’s one thing to create your own relay-based computer; that’s already impressive enough, but what really makes [DiPDoT]’s design special– at least after this latest video— is swapping the SRAM he had been using for historically-plausible capacitor-based memory.

A relay-based computer is really a 1940s type of design. There are various memory types that would have been available in those days, but suitable CRTs for Williams Tues are hard to come by these days, mercury delay lines have the obvious toxicity issue, and core rope memory requires granny-level threading skills. That leaves mechanical or electromechanical memory like [Konrad Zuse] used in the 30s, or capacitors. he chose to make his memory with capacitors.

It’s pretty obvious when you think about it that you can use a capacitor as memory: charged/discharged lets each capacitor store one bit. Charge is 1, discharged is 0. Of course to read the capacitor it must be discharged (if charged) but most early memory has that same read-means-erase pattern. More annoying is that you can’t overwrite a 1 with a 0– a separate ‘clear’ circuit is needed to empty the capacitor. Since his relay computer was using SRAM, it wasn’t set up to do this clear operation.

He demonstrates an auto-clearing memory circuit on breadboard, using 3 relays and a capacitor, so the existing relay computer architecture doesn’t need to change. Addressing is a bit of a cheat, in terms of 1940s tech, as he’s using modern diodes– though of course, tube diodes or point-contact diodes could conceivably pressed into service if one was playing purist. He’s also using LEDs to avoid the voltage draw and power requirements of incandescent indicator lamps. Call it a hack.

He demonstrates his circuit on breadboard– first with a 4-bit word, and then scaled up to 16-bit, before going all way to a massive 8-bytes hooked into the backplane of his Altair-esque relay computer. If you watch nothing else, jump fifteen minutes in to have the rare pleasure of watching a program being input via front panel with a complete explanation. If you have a few extra seconds, stay for the satisfyingly clicky run of the loop. The bonus 8-byte program [DiPDoT] runs at the end of the video is pure AMSR, too.

Yeah, it’s not going to solve the rampocalypse, any more than the initial build of this computer helped with GPU prices. That’s not the point. The point is clack clack clack clack clack, and if that doesn’t appeal, we don’t know what to tell you.

Continue reading “Capacitor Memory Makes Homebrew Relay Computer Historically Plausible” →

Surviving The RAM Price Squeeze With Linux In-Kernel Memory Compression

January 31, 2026 by John Elliot V 42 Comments

You’ve probably heard — we’re currently experiencing very high RAM prices due mostly to increased demand from AI data centers.

If you’ve been priced out of new RAM you are going to want to get as much value out of the RAM you already have as possible, and that’s where today’s hack comes in: if you’re on a Debian system read about ZRam for how to install and configure zram-tools to enable and manage the Linux kernel facilities that enable compressed RAM by integrating with the swap-enabled virtual memory system. We’ve seen it done with the Raspberry Pi, and the concept is the same.

Ubuntu users should check out systemd-zram-generator instead, and be aware that zram might already be installed and configured by default on your Ubuntu Desktop system.

If you’re interested in the history of in-kernel memory compression LWN.net has an old article covering the technology as it was gestating back in 2013: In-kernel Memory Compression. For those trying to get a grip on what has happened with RAM prices in recent history, a good place to track memory prices is memory.net and if you swing by you can see that a lot of RAM has gone up as much as four times in the last three or four months.

If you have any tips or hacks for memory compression on other platforms we would love to hear from you in the comments section!

Post-rampocalyptic Chip-Swap Provides Desktop Memory At Laptop Prices

January 27, 2026 by Tyler August 21 Comments

When you can buy something at a low price in one location, and sell it at a higher price somewhere else, you’re engaged in what economists call “arbitrage”. We’re not sure if desoldering DDR5 chips from laptop SO-DIMMs to populate a custom PCB to create much-more-expensive desktop memory counts as arbitrage, but it certainly counts as a hack. [VIK-on], who built the cards, claims he’s getting DDR5 performance at almost DDR3 prices. Nice!

Installed, the RAM apparently works well, though [VIK-on] has not shared benchmarks.

Specifically, he’s put together a 32 GB UDIMM from donor chips from two 16 GB SO-DIMMs. The memory chips themselves aren’t enough to make a stick of RAM, however: the part where we wish we had more details was in the firmware. The firmware identifies this DIY DIMM as an ADATA AX5U6500C3232G-DCLARWH, specifically. [VIK-on] is still performing stability tests, if those go well, we’re told to expect a how-to guide.

[VIK-on] is in Russia, so SO-DIMM rates may differ in your local market, but he claims walkaway costs of 17,015 ₽ — about $218 or €188, an astounding price for DDR5 in these dark days.

Some say soldering SIMMs seems severe, but hardly strange to Hackaday, and desperate times call for desperate measures. It’s ether that or optimize software, and who wants go to that effort?