I wanted to switch to DS4.1F from K3 (not even K3.1). Even the 1017 snapshot takes 5x-10x the tokens that K3 uses for the task, and often misses more intricate gotchas that K3 catches.
One issue I've had with swapping fans for Noctuas is the airflow matching. The louder fans aren't louder just because they're lower quality (which they definitely are), but also because (almost always in my experience) they force more air through.
That unit from the article sounds very temperature sensitive. I'd be wary of throwing a precision unit like that out of spec and out of calibration.
FWIW I've noctua-swapped so many things including an oscilloscope.
He mentions that he's running it in an air-conditioned studio, so presumably if the unit's datasheet lists an environmental range warmer than that, it has plenty of cooling headroom in this environment.
The last rubidium oscillator I played with, had to heat the vapor cell within the physics package, but then prevent that from overheating the rest of the unit, so it specified mounting to a heatsink. There was no heatsink recommendation given, just a certain package temperature as reported by sensor value X in the diagnostic data page. Get below that and you're golden. Get too much below that and you're wasting extra energy in the heater. Which is harmless, it just drives power consumption up.
So depending on how the unit's thermals work out, a slightly less aggressive fan may end up being mildly beneficial.
One of my slightly embarassing ones was the fan on one of my Nanlite lights was always on, do I thought to swap it with a noctua one so at least it'll probably be quieter.
When I took it apart, it turns out it was already a 12cm noctua... Black edition.
Airflow is typically measured for worst-case scenarios or not measured at all but just something pulled off a datasheet and then they install the cheapest fan they can find that meets the specs. My SOP is to open the case, get thermal readings, drop in a Noctua, and keep dropping the voltage while taking thermal readings to make sure nothing is getting too hot. I've got an NVR where I've dropped the speed to the lowest at which it will still start and it's still blowing cold air through the vents, so the 40mm screamer that was there initially was total overkill.
You can also get non-Noctua hydrodynamic bearing fans with polycarbonate blades and silicone mounts which are supposed to be very quiet, but I've never compared one against a Noctua. Anyone done a comparison?
The best way is to measure temps before (ideally over at least 1-2 hours), then to measure temps the same amount of time after and compare.
In the case of this Rb oscillator, it is temperature sensitive, but also built for pretty extreme environments.
I'm not too worried about this particular application, because the clock will be in an air conditioned environment, and I'm monitoring the clock accuracy to +/- 2ns from another GPSDO / rubidium clock. The performance before and after the fan swap is within my measurement noise for 24H, though I haven't done 30+ day measurements and ADEV comparison. Still working on a consistent test setup.
The GRClok datasheet[0] gives its operating range as -10 to 55 °C (14-131 °F). The integrated Calnex unit only says it should be kept at +/- 1 °C for rated stability, but otherwise it just mentions 25 °C.[1]
Considering the new airflow rating is 9.4 m^3/h (vs 16.8[2]), yes it sounds like plenty of headroom.
> You can just look at the airflow charts and buy accordingly?
Not with Noctua and the likes, not always at least. These fans are quiet, but they are not made of magic fairy dust. They simply max out at lower flow rates than some of the the noisy, high-RPM fans people replace with them. Those 40 mm Noctuas, apart from voltage options, don't really come in many varieties.
It's a trade-off - you definitely lose the upper end of the device's original temperature range, but if it's stable under your conditions - great. How do you find out? Well, you replace the fans and check if it sizzles or not.
And funnily enough, a LOT of equipment these days isn't made with formal thermal design, they simply throw in high-cfm fans and... check if it sizzles. It usually doesn't because there's plenty of headroom. It's loud, but who cares since it's not meant to sit on your bench.
Source: replaced PSU and chassis fans in countless pieces of network and test & measurement equipment.
I wish Noctua fan fans realized they're fan equivalent of "unsafe languages such as C++" Rust fanatics. Even actual Rust fanatics would see how that kind of attempted advocacy is counterproductive.
I found that surprisingly many k-pop songs do microtonal shenanigans. Nowhere near most, but it's there. Even something as popular as Boombayah (check the 4th note, it's not an on-tune semitone).
There's also more niche stuff like Flying Microtonal Banana.
https://www.youtube.com/@MoritzKlein0 has revolutionized my understanding of audio signal processing in software. Understanding what the voltages do is absolutely amazing.
This video seems to be about LM on untreated copper (which I thought wasn't really advised).
Now, I personally have had two builds with LM. One with LM as the material for a CPU re-lid, and another for a direct-die application. Neither touched aluminum or untreated copper.
The re-lidded CPU is still going strong, no issues. But with the direct die application, I found the performance to significantly degrade after two years. Apparently there's a pump-out effect.
I ended up switching to PTM. Same temps as LM, but zero degradation so far.
> LM on untreated copper (which I thought wasn't really advised).
I think the supposed issue is nickel (the most common waterblock coating) has a much lower thermal conductivity so you might lose a degree or two under load vs a bare copper block.
Which kind of makes sense: people who know what they're doing use liquid metal to squeeze out that last couple degrees vs PTM/graphene so going bare copper and dealing the the buildup problems maybe makes sense for them too.
He mentions in this video that liquid metal really should just be for the extreme overclockers who don't mind changing it out and resurfacing their waterblocks regularly which is good advice. The fact liquid metal is even semi-mainstream and comes standard on any consumer hardware is kind of stupid.
I never really understood why you would risk using liquid metal versus something like Honeywell's PTM with the effect size in temperature difference indistinguishable from statistical noise. The only drawback I've seen for the PTMs is that the paste is extremely sticky to the point that I've ripped pins off an old 3600 when I tried to remount the cooler without warming it up first.
From what I've seen the whole PC review industry operates on emotional narratives rather than scientific measurements. When is the last time you saw a cooler spec'd by its thermal conductance (Watts per degree C) as opposed to a whole rigamarole of we used so and so RGB motherboard with so and so PC case and this version of Windows with so and so benchmarking software. And every cooler in the limited set we compared it to was worse, but of course we left a whole bunch out. Like sure integration testing can be useful, but it doesn't beat characterizing units in isolation with hard numbers. And they never reduce the number of variables enough to do that.
Part of the problem is that any given cooler doesn’t have a single W/K value. It depends on the temperature gradient across the heatsink which depends on airflow over the heatsink (and both the heat and heat density of the CPU being cooled), and the airflow depends on the case (and fans). And airflow temperature matters too (are you cooling the CPU’s heatsink with air that’s already been warmed by a GPU?)
Note as well that I said heat density - different benchmarks (or applications) will load the CPU differently, and the same amount of power consumption will result in very different heat outputs (heavy single core use will result in part of the die and heat spreader getting really hot, whereas heavy multi core use will result in the whole die and heatspreader getting kinda hot). If the heatsink has a very high surface area and high airflow it might have the ability to dissipate quite a lot of heat, but if the contact surface sucks (or is aluminum) it might not actually be able to transfer heat away very effectively without a massive temperature delta - resulting in high CPU temps for single core use even with a high cooling capacity for multi core use, for example.
Your average person can’t understand all of these variables, and the average reviewer can’t control for all of them. So, “this CPU with these benchmarks in this case with this fan configuration” is the best you can really expect. Because otherwise, what reporting are you actually going to do? “This cooler with 3.2CFM of 26C airflow on a CPU dissipating 80W on one core and 10W on the other 7 cores results in a peak core temperature of 87°C and a package temperature of 81°C” is not useful reporting. Especially because, when the average person doesn’t understand it, they’re gonna buy a case fan that says 3.2CFM on the box and then wonder why their CPU’s throttling.
Sure, but the goal of science is to distill these variables down into a model of what matters. Define a reference contact patch that's supplying the heat, and you've suspended all of the concerns about single core vs multi core. Then for more in depth articles, there could be a few to represent single core (smaller area), multi core (consumer processor), and multi core (larger server packages). Same with the airflow over the secondary parts of a cooler.
Another in depth article could then even analyze CPUs and how they themselves differ from that reference contact patch, and their own thermal resistance from each core to their surface.
And yeah, I know this is more work. The point is that doing the work allows the results to be used and compared across articles, even when some things are subtly different (eg contact patch). The problem is we're missing a strong incentive for these things to be objectively comparable, rather than having to read article after article and getting tripped up in the details.
The incentives aren't there (in my opinion) because the average person doesn't actually want to see that level of info. They're reading (or more likely, skimming) these sorts of reviews to have someone tell them what they should buy. The reviewer could go into a crazy amount of detail to figure out all of the thermal characteristics of the thing they're reviewing, but in the end it won't make a lick of difference compared to doing "just enough" and recommending something "good enough." Especially when, in most cases, coolers on the market these days are either "good enough or better" or "junk" and it's not hard to tell them apart.
There are exceptions to the above, like GamersNexus, who do take a slightly more scientific approach to things, but at the end of the day, most people don't care.
It also doesn't matter except for the most extreme overclockers. Even the worst tests are granular enough to tell you which ones are terrible and which are competent. The rapidly diminishing returns when you aim for the top 10% of performance in CPU building seems to hold here. Plus Steam's hardware survey puts the vast majority of users within factory overclock ranges. With so much of tech 'journalism' funded by ad views, you'll rarely get anything like extremely rigid methodologies outside of a few rare uber-nerds with dedicated followings (eg Igor's Lab, Snark's Domain) simply because it doesn't matter if, for eg, your 5700X3D is 70C or 73C under load.
Indeed, also, the variation between individual applications is enough to invalidate the results of the number of tests isn't large. It is super easy to mess this up. Too much paste, too little paste, a little bit of grease left behind or maybe a small piece of dirt... A nice way to practice application of paste is to use a piece of glass instead of an aluminum cooler. That will show you exactly what happens between the piece of aluminum and the CPU and it can be quite revealing how hard it is to get a perfect layer without any air trapped. Once you can do it perfectly with the piece of glass you'll be in a much better position to lay down a nice layer with the real thing.
> PTM thermal paste—most famously known as Honeywell PTM7950—is a Phase Change Material (PCM) used as a high-performance alternative to traditional CPU and GPU thermal paste
I ended up going with a Synthetic subscription (they host some open-weight models including K3 and GLM 5.3). They're the only subscription that I found that's simultaneously cheaper than token rates, and is extremely transparent about usage and pricing.
How does that make economical sense from the side of synthetic though? (please feel free to correct me if I am wrong, I usually am)
If they're simultaneously cheaper than token rates but couldn't they have sold these tokens in the first place on a platform like openrouter?
And the platforms that are hosting these tokens on openrouter aren't really earning any money at all in the first place as that itself is a race to bottom (and its more so just neutral, neither positive nor loss or barely profitable) so unless synthetic is actively making loss, and if that's the case, then it might not be sustainable, I don't quite understand the economics behind it.
Is the economics behind this the fact that since its a single subscription instead of token based pricing. Some people might use less tokens than what their subscription costs and so it offsets/subsidizes the other set of people who need more tokens.
But then, this means that as you had suggested in the first place, it would be a good deal for the people who want more tokens, so they would flock even more to the platform and it should eventually balance out unless external money is being burnt as I had said earlier.
I am genuinely interested in how they make this work so I would love if someone can shine some more light to it as I would like to read more about its economics that makes it work, thanks in advance! :-D
It appears, from the outside, to be a collection of small optimizations:
- The way they pace and limit usage seems to echo some strategies I've seen to reduce the impact of burst usage
- They cut down the maximum context length of most models
- They claim to have some custom sauce that improves how they handle API requests and reduces error rates
- Their explanations appear focused around optimizing concurrency in the same way that the 'flex' mode of some APIs does, though I've not seen specific docs of them claiming that this is part of how they save on costs
There may be more that I haven't seen or noticed. I don't work for or with them.
I tried searching for "Synthetic AI Subscription" and "Synthetic Data" is such a massive keyword in AI pre-training circles it's impossible to find the company.
The libraries, yes, but the code you write is portable (at least until you get into squeezing the last few percent and switch to Triton / Helion in case of GPU, and even those are decently portable).
There's also Halide, where you write the algo but the framework gets you the scheduling and SIMD.
reply