I think your leading statement is a bit more authoritative than what the linked paper actually demonstrates. It demonstrates your claimed qualities for specific algorithms, methodologies, and models.
It explicitly says that Quantization-Aware Training is out of scope, may reveal different trends and has potential for extreme quantization. Heavily-quantized model releases today generally use QAT or something similar, so I don't think the paper can be universally applied to modern quants (e.g. [0]).
That said, I don't see any evidence that TFA does any of that so the paper's claims may be quite applicable there.
And things like GSQ-RCO optimization from ISTA labs, EXL that you linked. Preserving bits for certain (large) domains like agentic coding. Pruning others. It really works and creates functional models.
I recently did some toying around with (very) small language models (<200k). I used QAT from the beginning and was surprised that I got byte identical results from the int8 model vs the float32 one.
If that scales with parameters I don't see why QAT shouldn't be the norm.
(As I was only comparing the QAT f32 model vs the quantization its possible that a not QAT f32 would have performed even better)
France is spiraling into a debt whirlpool needing to use debt to pay off debt, basically having no good way to resolve the issue without a lot of pain (“austerity”). Any politician with enough foresight to do what needs to be done gets removed, so those that remain are selected for their thick foreheads.
What’s more fun is that while france is teetering on the brink, it’s hardly alone, and it’s collapse will tip the dominoes.
A massive portion of our lawmakers are absolute clowns too, it's not exclusive to the US.
Note that this amendment to the law was passed with 11 votes for, 10 no... With a national assembly of 577. I have to assume it was voted at night, that's when peak democracy happens.
I've been self-hosting my email for years. Something that wasn't obvious to me in the beginning is that you can separate sending and receiving: host the receiving yourself and use a service for sending (e.g. Mailchimp, Sendgrid etc.). That can address many concerns.
I did that in the beginning but later set up my own relay hosted at Hetzner. I could do it on my receiving server (a PC in my house) but I don't have a static IP and IP reputation is a big deal in email deliverability.
I do very occasionally get sent to spam but it's fairly rare and I'm not sure my free Gmail account was any better.
Does anyone know if the voices are AI-generated? I was a little surprised; I don't think I would have questioned it if I heard them in a real game or video but since this seems to basically be a joke, I'm doubtful real voice actors were used.
Can you direct me to such a frontier? Bonus points if it is cheap.
I want to use voice synthesis to read my blog articles out loud to me as a self-editing step, but the experience was nothing like a human when I last tried.
The voices sounded natural-ish alright, but intonation was all over the place.
Interesting but for photography at least, there's a massive hole: RAW support. What they've implemented is very limited and basic, nothing like parity with Lightroom or Photoshop. Just try loading a RAW photo into Lightroom/Photoshop and put the window next to Lightcraft/Photocraft - the results are completely different and not in a good way. Lightcraft's image looks like an amateur attempt at RAW processing from first principles, without any of the most basic colour profiles and adjustments that Lightroom applies. Photocraft doesn't even offer RAW processing, it just dumps the raster straight in.
I'd treat this as completely new, far from complete software heavily inspired by the Adobe Suite, rather than a "port".
But maybe folks who use workflows that don't involve RAW photos have better luck.
Or maybe you'll have better luck in a week when somebody familiar with RAW and a 20 year grudge against Adobe chooses to spend a day of tokens on the problem. Maybe me.
Rome wasn't built in a day, nor Adobe destroyed in a single git push. But they both shall fall.
And 500 levels of org tree, and grumpy customers that freak out if you change anything about their workflows, and decades of legacy code across dozens of micro-services, and backwards compatibility requirements with 20 year old formats...
That's exactly what I was thinking. It's only a matter of time. Surely the models, when well directed, can get most or all of the way there in short order.
I find it awfully funny that your response to someone creating the entire Adobe software suite in a day is “the colours aren’t rich enough for my liking”
Raw was a tough one when I tried it a few months ago -- I never thought about how individual color pixels on the sensor come together to form a color value for a point in an image file. Super interesting, and not at all standard between the different manufacturers. It's a hard problem for software to solve in a general way.
When an LED bulb dies, it's typically not the LED, it's the electronics driving it. This is worsened substantially by heat. If you have recessed downlights or something, that's close to worst-case for an LED bulb.
Lights with non-replaceable bulbs can actually be far better, because they can be designed for their environment. Recessed downlights with built-in LEDs for example can thermally couple the LEDs directly to the metal housing, so the heat can be dissipated into the surrounding air, instead of into the recess.
Maybe in theory. In practice, the Phoebus Cartel was a real thing, and they limited lightbulb lifespans, worldwide. The same thing exists today, albeit more because companies that sell lightbulbs want to keep selling them, instead of putting themselves out of business. So, many LED lightbulbs are intentionally designed to fail early.
The only place where this doesn’t happen is in the country of Dubai. Because of national law, lightbulbs sold there are required to have very long lifespans. That’s why there is a world-wide grey market business for taking existing LED lightbulbs apart, snipping off a resistor so that they pull about half the amperage and then putting them back together, and suddenly their lifespan is increased by 10-100x. This is called a “doobie”, after the name of the country. Check out “bigclivedotcom” on YouTube if you want to see this in action.
You will notice this if you buy Philips Ultra Definition bulbs in Dubai, versus the same bulbs sold at the same amount of lumen output from the UK or elsewhere in the world. The “doobie” version has twice as many LED filaments inside, compared to the UK version. Same price, way different life span. Philips is one of the better manufacturers of lightbulbs, and nowhere remotely close to the only one doing this.
With regards to LED light fixtures (ones that don’t allow you to screw in your own bulb), most manufacturers do the same thing. They’re designed to fail early. Who wants light fixture styling that is fifty years old, just because the LED lights inside haven’t failed yet? They want to make you buy new fixtures every couple of years, just like they do for bulbs.
Yes, there are exceptions in both cases. But they are the rare exceptions, not the rule. And if you haven’t studied up on the Phoebus Cartel, you can start with the article on Wikipedia.
I'm well aware of the Phoebus Cartel. It's not as straightforward as often reported. Have a look at Technology Connections: https://youtu.be/zb7Bs98KmnY
Dubai isn't a country. It's a city in the United Arab Emirates.
The Dubai bulb isn't some trick Big Lightbulb doesn't want you to know about, it's an engineering tradeoff: lower current on more filaments gives you higher efficiency and lower temperature but at higher cost and with substantially lower light quality (CRI).
You'll find that after EU energy labeling, there are products available to European consumers that well exceed the specifications of the "Dubai" bulbs. Look for EU energy class A. They often come with a 5-year warranty and rated lifespan of 50000 hours. The "Dubai" bulbs were rated for 25000 hours. You'll also find though that they're substantially more expensive and usually don't have a CRI above 90.
Regarding planned obsolescence, in many cases I can believe that a product will fail earlier than it would if designed for longevity. Rather than planned obsolescence, I believe this is simply a result of being cheap, a consequence of the race to the bottom. If you're buying cheap Chinese fixtures for $10, or cheap Chinese bulbs for $2, you're not going to get a high-quality, long-lasting product.
If you want products that last a long time, buy them and use them properly. If you want bright lights, make sure they're adequately cooled. Consider buying lower-voltage fixtures and bulbs so that you can have a separate power supply and don't put the full ~230V to ~3V conversion in a tiny space.
> Dubai isn't a country. It's a city in the United Arab Emirates.
I learned recently that Dubai is also the name of the emirate. I don’t have a good understanding of the UAE governmental structure, but the Emirates do seem to have their own governments. Maybe it can be thought of as a country in the same way Scotland can be thought of as a country.
The United Arab Emirates is a federal state, like the US or Canada (and notably not like the UK--Scotland is a devolved component of a unitary state, not a member of a federal state). The Emirate of Dubai is thus most closely analogized to a US state or a Canadian province, though I'm not entirely sure off-hand how the power of a UAE Emirate compares to federal subjects in the countries whose politics I'm more familiar with.
Also, Dubai is effectively a city-state, as the vast majority of the population of the Emirate lives in the city itself.
Biggest problem of leds is we have no good standards. Classic screw in bulb makes no sense with LEDs. Other formats usually mean you are screwed when you have to replace. So far LED strips seem the best way. A bit of soldering but you can replace them by parts.
Screw in bulbs make plenty of sense. You are not going to get people to convert to LEDs by forcing them to replace every single fixture in their home. That would be ridiculousmaxxing.
It would be nice to have an updated standard for newer buildings, though. In practice fewer and fewer fittings are using screw-in lightbulbs at all but there's not another standard.
What happens when something does go wrong though? Now, you have to replace the entire thing instead of just simply replacing a bulb. If only we had some other previous existing examples of fancy tech that used non-replaceable parts requiring a full replacement...
The lightbulb that you're throwing away also contains all the same electronics that a larger light fixture does, just in a smaller and less thermally efficient space.
Mounting options that allow larger LED lights allow you to use lights that are more thermally efficient without sacrificing any replaceability. JIS C 8310 mounts for example.
I think they're suggesting a lightbulb equivalent of going from USB-A connectors to USB-C. Replaceable bulbs in sockets that are just as easy to use, but with a smaller profile.
The problem is going from mains voltage to 3.3v means some sort of waste heat. Rewiring the whole house to 12v is one idea I've seen, but again, total house rewire and fixture replacement. In the mean time, it's A9 bulbs with LED electronics going from 120v to 3.3v generating heat. Making cheap electronics means that heat will kill the electronics. Making good electronics at a higher retail price is the solution, but consumers are not educated enough so all they see is sticker shock.
Your comparison of USB-A to USB-C is nonsensical and does not actually acknowledge the issue of why the existing screw in bulb direct replacement have issues
Screw-in sockets are more durable and long-lasting than any replacement so-far, most of which have been replaced again (remember those weird fluorescent mini-bulb sockets?).
12V is too high for your LEDs, you still have the power supply. 12 volts is not near enough voltage for a house. You can get around it with very big wires but copper is extremely expensive. Even in cars, 12 volts is borderline. Trucks are fairly often go to 24 or higher voltage because even at the size of a large truck (semi), higher voltage is an advantage.
Well, we can certainly argue with 110 or 220 is a better voltage for a house. It is quite clear that 12 volts is too low.
Yeah 24VDC or 48VDC would make more sense but tbh you only need to drive the LEDs with it.
At 48VDC you could run 16AWG like 100ft and still be fine as long as the lights pull around 60W or less all together (which works out to ~12 lights on a circuit). And that circuit would only add another ~2W or so in power losses when energized.
That circuit is essentially what power over ethernet does today so you could replace romex with PoE Type 3/Type 4 plenum rated ethernet and it'd be like half the cost.
And if you really want to optimise for losses you could put the switches on the ethernet data line and use that for sensing rather than running the PoE circuit out to each light switch (and then de-energise the PoE circuit when the switches report off). Bonus points in that this would allow you to "reprogram" which light switch controls which circuit without needing to do any re-wiring.
You still have to run Romex everywhere, though. People want to run other things in various rooms. They want the option of turning a bedroom into an office should that make sense. That means they're going to see enough power to run a decent sized computer. TV's and bedrooms are common. Stereo systems. Just me and my wife charging our beds by the phone overnight is going to take a significant amount of the power that you allocated to the room.
And again, your whatever voltage, 48 volts, is way too much for an LED, so we still have a power supply. What's the difference between a power supply running on 48 volts DC versus 110 or 220 volts AC? Ultimately, there's not much of a difference, well, you could do a very cheap linear regulator and be horribly wasteful of power, but more likely you're going to be doing a social power supply, which is going to be more complex and have similar efficiency losses. I'm not enough of an electrical engineer to know this for sure but I would suspect the power supply in your Ethernet switch and then the second one to get to that your voltage is less efficient than just having a single power supply that goes from regular line voltage to whatever the LEDs need.
You still have to run romex for the other circuits. But it's not hard to run those on a separate breaker. Rooms often have like one and a half 15A circuits each because lighting used to draw so much power but now you can drop it down to one 15A circuit per room and run a single circuit that switches out to the entire house off of one 15A breaker.
> what's the difference between a power supply running on 48 volts DC versus 110 or 220 volts AC?
It's already DC. A properly sized buck converter based LED driver is going to have like 80-95% efficiency when the LED is on (and it's trivial to draw no power when off) and they are like 0.25USD per IC.
They are going to have equal or better performance than AC->DC switching mode power supplies because buck converters are already the simplest and most efficient type of SMPS.
It doesn't need to be for the whole house, it just needs to be for the lighting fixtures. Obviously nobody is going to use a microwave or blender on 12V. It's not like our bulbs plug into NEMA 5-15 outlets. Having good DC standard voltages and plugs would be incredibly useful. Don't bother running 12V everywhere, but like after the 120V light switch go to a single controller that then drives the lights for that room. Or have that be in the box with the switch, although that might still have the same kind of cramped heat issue.
And it's not like we need to just stop producing Edison socket styles tomorrow. But let's make sane standards today.
I don't know how it is where you live, but most apartments where I've lived in Paris didn't have dedicated wiring for the lights in the "living" spaces. They had them in the bathroom / hallway / kitchen, but not in the living rooms, for example.
My current apartment has dedicated circuits in the living room and bedroom. But in the living room they're in awkward places.
I always had to resort to multiple lamps that would plug into an outlet so I could place them where I needed them. Running 12 V for the bulbs would require rewiring most of the apartment.
Now, I'm not against having a new socket standard. But it would have to be connected to 220 V.
That is fairly typically in the US as well. Most often there is a switch in the living spaces, which controls have the outlet half of a duplex outlet. Plug something into the top (sometimes bottom - there is no standard) and the switch controls it so you can get the lights on/off when you enter/leave the room, plug into the bottom (again sometimes top) and it is always on, perfect for whatever else you have in the room.
Its not like your bulbs are plugging straight into your wall outlets. You've got some kind of lamp fixture that the bulbs screw in to, correct?
Your lamps would still plug into mains, same as they do today. The lamp fixture overall would contain the driver components, hopefully in packaging which lets it manage heat well and have good build quality. But whatever parts do fail routinely should be able to be easily replaced with some kind of standardized socket. It just seems like we're trying to squeeze too much too cheaply into a form factor that's ill suited for the requirements by trying to shove all the driver circuitry into the same envelope of an older Edison A-shape socket style bulb.
This has already kind of been a thing for a while with low voltage halogen lamps. The bulbs themselves aren't made for mains voltage, the lamp structures themselves give the right voltage for the halogen mini bulbs. I've replaced a few bi-pin halogen lights for LEDs with drop-in replacements (especially for outdoor lights), but this same kind of bulb packaging hasn't really been very common for interior lighting from what I've seen.
> Now, I'm not against having a new socket standard. But it would have to be connected to 220 V.
Why would it need to be 220 V in the "socket" for the bulb? Why do we need lamps themselves to just be a frame to hold a wire, why not put the driver circuitry in there to step it down to whatever voltage we decide to drive the DC LED bulbs at?
> Why would it need to be 220 V in the "socket" for the bulb? Why do we need lamps themselves to just be a frame to hold a wire, why not put the driver circuitry in there to step it down to whatever voltage we decide to drive the DC LED bulbs at?
I was talking about the lamp itself. Maybe I used the wrong term? I meant the fixture which holds the wire and expects a bulb. That fixture needs to go into mains, because that outlet may have to also service something else entirely.
Now, sure, usually those fixtures tend to be larger than the bulbs, so they could hold the conversion circuitry. I'd be fine with that. But the commenter I was replying to talked about having an alternative DC circuit for the lights.
Sure, and we can have that too, in the places where the lights are built-in to the ceilings or cabinets or whatever. When we have lights in the ceilings, the fixtures again can handle this specialized LED bulb style. And one might even link several lights into a single DC circuit for all the lights in the room versus having each individual light do its own AC->DC conversions in compromised envelopes.
My vote would be 24v or 48v. You cannot power substantial LED with 3.3v. You’d spend a fortune rolling out ultra-thick gauge wire to avoid pissing away non-trivial amounts of power into heat. Many, many LED strips come as 24v. It’s the perfect balance of wire gauge for the amount of power you’ll be drawing.
I've seen on Amazon packs of 4 bulbs for $7 which is significantly less than the price of a single Hue bulb for $25. While the Hue bulb is "tuneable" white adjustable from cool and warm, the price is 3x for 1/4 of the number of bulbs. Not really sure that counts as marginal. Unless you're marginal is describing the cost to manufacture, but what's that got to do with anything?
I would love to buy an LED bulb that lasts a long time. I thought the reason for their premium price was supposed long lifetime. Mine all last as long as the older bulbs but cost 5X as much. Where can I buy such bulbs in the US?
I have yet to replace any of the LED bulbs I bought at Costco in 2013. I dunno who manufactured them, or what kind of unobtainium they're made from, or whether I will jinx them by talking about it, but they just... Keep on going. Basic Edison screw-ins, from the first generation of bulbs that emitted warm light, that I've used exactly as I would any light-bulb. They weren't really expensive - maybe $2-3 apiece, though I really don't remember - and, I dunno. I hear everyone constantly complaining about their LED bulbs, and maybe they've been enshittified since, but I have absolutely no complaints.
I've been using LED lights for over a decade. Can't say I have replaced any of them.
And nowadays you can get E27 LED bulbs for 75 cents at IKEA, which is cheaper than incandescents ever were. Not sure how they managed that, because it sounds insane.
Similar experience. I'm perplexed at the people saying LED bulbs fail frequently. I have probably 100 of them in service for 10 years or so (basically since Costco began selling them) and possibly one has failed. Memory is hazy but definitely no more than one.
There can be a lot of variation. From what I've seen, I'd say 2-5 years is a typical lifespan for an LED bulb. We have some recessed lights in the kitchen that require brighter bulbs, and those typically last only 1-2 years. But then we also have some other (non-replaceable!) recessed lights that are 14 years old and still work fine.
There was one time I bought some cheap ones from Amazon, and several of them failed within a few months. There was also an LED bulb in the front porch light (so, outside) that lasted around 10 years, despite being outside and used daily. I had several of that same type and that particular one lasted longer than all of the others, where the others were all inside and not installed upside down. So yeah, it varies.
My Philips Hue bulbs are going strong for just over 7 years now. They are the premium bulb in my experience. I've bought other LED strips using controllers compatible with Hue, but they died where there are obvious burned points all along the strip. We all have different experiences that will counter each other, so it's frustrating to no end on finding trust worthy devices.
Regardless of all of that frustration, thank the gods we no longer have to deal with CFL
I couldn't decide on how to respond to someone saying they like CFL. Literally, I was at a loss of words. My first reaction was coughing into my hand while saying bullshit.
My Philips Hue bulbs are of a similar vintage, and I love them. Literally the first thing I did after moving into my home was removing every single CFL bulb and replacing them with either a Hue bulb or another relatively high quality LED bulb with a high CRI and 3000 kelvin temp.
I also find it hard to believe someone really does like CFLs as much as buttocks describes, and it makes me wonder if they are in fact just trolling.
My kitchen had 8 4' fluorescent fixtures in the kitchen when I moved in, and a couple of them flickered. I didn't bother check ballasts or bulbs and instead I've replaced them with the Hue strips + extenders. I've seen ballast replacements for LED tubes, but at the time did not see Hue compatible so I went with the strips.
Anyone that says they like CFL has to be trolling. Any other answer is just inconceivable to me.
I didn’t realize it was so controversial. Indeed, I have a set of CFL from around 2010 that have about 30 second warm up to full brightness; they look just right behind their lamp shades. I tried replacing these once with LED and the light kind of pierced through the shades and looked harsh.
Reliability engineering is a thing. I had a friend whose research was targeted around getting consumer white goods to work for a specified number of months.
They aren’t designed to fail, they are designed to work around some predictable model during the warranty and extended warranty period.
There's definitely no cartel today, there are wayyy too many manufacturers for that to be a practical reality.
The reason most consumer LED bulbs fail early today is because consumers are low-information buyers who make purchase decisions based overwhelmingly on price and secondarily on aesthetics. The typical buyer has no way to accurately judge the actual design longevity of a bulb on the shelf. Yes, bulbs have lifetime ratings, but even if people bother to read them, those ratings in many places are not well regulated and reflect an ideal scenario that oftentimes is not reflective of real-world use.
The sales success of these race-to-the-bottom bulbs has driven higher quality bulbs off of the shelf at most retailers. You can get well designed long-life LED lighting, but you'd need to step up to commercial/industrial options. There are long life bulbs available, but they generally are designed for uses where they are difficult to replace... because in those use cases a high product price can be justified by lower labor costs to maintain their installation... e.g., bulbs for use high up on a structure, factory lighting, continuous use, etc.
The reason Dubai has better bulbs is simply because they have decided to require it. The reason the US has shit bulbs is because Americans love buying $0.80-1.50 light bulbs. The Dubai bulb is an order of magnitude more expensive. This is a hard sell at Walmart. And the TCO of the $0.87 bulbs on Amazon is still less even accounting for early failures if you are a homeowner who is not factoring in labor costs.
> The only place where this doesn’t happen is in the country of Dubai.
As far as I'm aware, we've got comparable bulbs available in Canada. We've got the Philips 95 CRI ultra definition bulbs and the ultra efficiency bulbs.
The UltraEfficient bulbs seem better-specced than the Dubai bulbs - the standard A19 of the latter is specced for 6W/600 lumen/25k hours at 80 CRI, while the bulbs I can buy at the drug store here are specced for 5W/810 lumen/50k hours at 80 CRI.
Interesting. I was just thinking I have never had to replace a single bulb in my apartment. I have lived here for five years and the bulbs weren’t new when I moved in.
I also have Nanoleafs that I have had for about six years and so far no issues. They run for about 12 hours a day on most days.
I am in BC,CA. I gotta take apart my lights and see what they are.
Just to add some context to the Phoebus Cartel and incandescent bulb lifetime in general, there is an inherent trade off between efficiency/brightness vs lifetime in incandescent bulbs. Longer life bulbs are proportionally less efficient, which can end up costing the user more money in the long run than replacing efficient shorter life bulbs more often. There's no doubt that the cartel was limiting lifetimes for the wrong reasons (profit), but the result for consumers wasn't necessarily the wrong one.
Modern LED bulb lifetime is the result of similar factors: wanting more profit. But I think this is being achieved more through aggressively reducing manufacturing cost rather than planned obsolescence.
Just an anecdote but I had a huge number of LED MR16 bulbs in my house. Bunch of different brands. They were all crap so we replaced every downlight in my house with an LED disk instead. I’ve stopped thinking about light bulbs altogether.
Maybe they can be designed for thermal management, but very few of them really are. I'm also finding the LED is the most likely failure point, so I prefer to find fixtures with replaceable LED bulbs so I can install the ones I know to be reliable. That way the entire fan isn't junk when the LED dies.
Integrated LED fixtures are usually quite easy to fix when an LED actually fails. Check them all with a multimeter in diode mode and you know which one to replace.
In theory, yes. In practice I suspect a lot of these fixtures are not using integration to achieve better thermal design, they're using it because it makes the product cheaper and lets the manufacturer control the whole assembly.
Doesn't the LED going out after 2-3 years make it not only way more expensive (LEDs are easily 10x or more than a regular incandescent light), but also way less energy efficient and much worse for the environment?
These lightbulbs, while using less electricity, use much more energy to produce. And I wonder what with their substrates, if that isn't more difficult to recycle too. I'm not sure, and don't want to judge, but if both every efficiency and recycling are harder, then these bulbs aren't really energy and environmentally efficient.
Not sure if I buy special LEDs, but I have outfit my house with LEDs basically everywhere between 10 and 13 years ago. In that time I had maybe one or two of them die.
Traditional light bulbs have taught us that lights naturally have to be "consumables". That has been far less true for energy saving lamps (of which I have maybe 2 or 3 still left which I planned to replace by LEDs once they die, but they just won't). As far as I can tell, this isn't at all the case for LEDs.
It's much more likely that lamps and ceiling fans made out of plastic break from accidental physical impact or turn ugly due to natural UV discolorization or similar. And probably most will live on until replaced due to changing tastes.
Unless you only need the AI available some of the time, $5/hr is pretty expensive. That's an RTX Pro twice a year.
If you're using it for discrete sessions of coding or something, that might make sense for you but if you're using it for an always-on assistant, that pricing kinda sucks.
For a lot of people involved in local LLM, ownership and control of their data matters to them a lot. It's what pulls a lot of people toward local models. Renting a 5090 off of some random infrastructure provider that is reselling some random dude's 5090 does not fit the bill.
I think maybe your view is too narrow? It's not that people don't want to make a machine, it's that now they don't have to.
When I started hobby CNC a few years ago, the only compact or desktop CNC machines were flimsy Chinese things made entirely of extrusion, expensive machines made for education/prototyping environments and the Carbide3D Nomad (which I ended up buying).
I ended up building a machine from scratch because the original Nomad was too flimsy to really cut metals, even Aluminium.
But in the last few years companies like Makera came along and sell desktop CNC machines that can cut metal for under $6k USD and there's even a toolchanger.
For just a little more, there's a Kickstarter for a 5-axis machine with mist cooling, ATC a wireless probe and it can cut steel and titanium.
And for those that want a DIY project, there seem to be plenty of folks doing CNC conversions of second-hand manual mills.
According to a Feb press release, they're fabbed by TSMC:
> For FUJITSU-MONAKA, the 2nm
semiconductors will tape out next year. Due to this, Rapidus would be unable to meet the
deadline, so TSMC is handling the manufacturing.
Their CPUs are faster at inference than many GPUs. These are server CPUs with 144 cores and wide memory interface.
Fujitsu has added some ISA extensions for AI inference to the standard Arm ISA, e.g. instructions for inference with FP8.
Even their previous CPU generation was for some time the champion in energy efficiency, with better performance per watt than the NVIDIA GPUs, until a newer generation of NVIDIA GPUs has leapfrogged them.
There are good chances that this new Fujitsu CPU might be again for some time the CPU with the best energy efficiency, but it remains to be seen how it compares with the recent GPUs.
At the very top of the article:
"Achieving world-class AI inference performance through Japan-developed 2nm 3D-stacked CPU and server integrated, developed, and manufactured in Japan"
"Japan-developed CPU..." plus the server itself (parts, power supplies, case, etc.) integrated, developed and manufactured in Japan. It sounds like they're weasel-wording it a bit, but I don't think the CPU itself is manufactured in Japan, only "developed."
It explicitly says that Quantization-Aware Training is out of scope, may reveal different trends and has potential for extreme quantization. Heavily-quantized model releases today generally use QAT or something similar, so I don't think the paper can be universally applied to modern quants (e.g. [0]).
That said, I don't see any evidence that TFA does any of that so the paper's claims may be quite applicable there.
[0]: https://huggingface.co/vcruz305/DSV4.1-Flash-1017-SAGE-REXL3...
reply