2026-05-04 SFF Coolers

As much as I like small computing, I haven’t gone in much for traditional SFF builds, in SFF cases. I had one such build, in late 2017, that worked out very poorly due to me being bad at cable management and really trying to cram too much stuff in the case.

I have another one now, though, and it’s doing much better. It’s still using the HSF from that old build though: a Noctua NH-L12S. And that’s really limiting it.

I have become aware that there is a new generation of higher-performance, low-profile HSFs, and I’ve ordered one in hopes of getting some extra performance for relatively little money. The new cooler is a Thermalright AXP90-X53, full copper version. Based on reviews it seemed like a good way to try to get more performance, for not much more money.

We’re trading finstack and fan size for fan speed. Also, in its “compact” config (70mm height, as installed) the Noctua’s fan blows up/out, through the stack from the CPU side. The Thermalright’s fan blows down/in through the stack from the case side.

How does performance shake out? Here’s the numbers with a CPU setpoint of 68C, and an RTX 5050 running at 120W in the same case:

Where GHz is the average of loaded core speeds after a minimum of 30 minutes of operation; PPD is the average of the last 10 WUs crunched (for the Noctua), and 3 (for Thermalright, because the results were very unambiguous).

Conclusion: it’s not pretty for the AXP90-X53. The extra fan speed and fully copper construction just isn’t enough to keep up with the NH-L12S. The Noctua’s unconventional construction lets it have much more finstack area. The size difference doesn’t look like much as a linear measurement, but it works out to 1.7X more area.

Also, Noctua beats Thermalright in categories that aren’t quantizable. They often cost more, but you’re getting excellent human factors engineering as well as thermal and fan engineering. Installing the Thermalright was a finicky exercise in installing nuts from the backside of the mobo, which required removing the GPU in my SFF build. The Noctua is a very tight fit, but screws in from the front only, making for a much smoother experience – and it’s quieter as well.

If you need good performance in the smallest space possible, at a very low cost, the AXP90-X53 might be for you. But I’ll be putting the NH-L12S back in place later today.

Raw WU data follows:

noctua = [697322, 685012, 720909, 448021, 931078,
          946208, 670635, 842266, 685734, 452973]

thermalright = [320338, 491237, 517453]

2026-01-12 Hardware thoughts (some more)

I was thinking over the weekend about what I’d want my hypothetical “bigger” single machine to look like when Zen6 arrives. It’s mostly about core counts, so we can start there to get a more concrete idea of things.

Currently the fleet is composed of two 9900Xs and one 9950X3D. that’s 12 + 12 + 16 = 40 cores (or 80 threads) total. But only the 9950 is fully dedicated to folding: both 9900s have other, minor tasks – and all of the machines have a GPU, which needs a reserved thread. So the 9900s have reservations of 2C/4T (one thread for GPU, three for other tasks) and the 9950 has a single thread reserved, for a total of 20 + 20 + 31 = 71 threads actively doing CPU folding.

This means that I can’t replace all three machines with a single server unless it has more than 64 threads, which AMD seems unlikely to deliver in the consumer space, at least this year.

But I’m not planning to consolidate all three; I only want to consolidate the two which are on one of the desks in my office. Additionally, when that happens, I’d be moving the 9950 into the remaining existing machine. So my baseline pre-upgrade capacity will be 32 threads, minus the 2C/4T reservation that machine needs, for 28 threads doing CPU folding. That leaves a deficit of 43 threads versus the present.

The most persistent rumors are that Zen6 will move from 8 cores per complute chiplet to 12. If we assume that’s true, and further assume that consumer CPUs will retain the current 2 CCD + IOD configuration, then the Zen6 [3579]950 equivalent will have 12 * 2 = 24 cores, which is 48 threads. Reserving 2C/4T would leave me with 44 threads doing CPU folding, for a total of 72… a net gain of one thread.

So my overall folding effort would stay near-identical, modulo whatever IPC gains we see with Zen6 (10-13% seems safe, by taking the average of the last 3 generational uplifts). The real winner is my hardware and power budget.

2026-01-06 Hardware thoughts

No real changes or news since the last update, other than hitting 20k WUs for FAH – points rack up way faster for FAH than for BOINC projects, but WUs run for much, much longer.

I did have some thoughts on a couple things I said last time though, specifically:

• Keeping it at three machines for the forseeable future
• All on Zen 5/AM5 for now. Upgrades are unlikely until Zen 6 is released, possibly late next year

What I meant by the first point was really “I’m not adding any more machines”, and that has held. In fact, in conjunction with the second point, if Zen6 finally ups the core count on consumer CPUs, then I will likely consolidate two of the existing machines into one.

This line of thinking has been influenced by recent experiences with a couple of GPUs, namely a 9600 XT and a RTX 5050.

I read in reviews that the 5050 was capable of performance equivalent to a 4060 – that may be true in gaming, but it’s absolutely not true in terms of compute. It hovers right at 2M PPD when turned down to its lowest allowable wattage (110W). Meanwhile a 4070 non-Super, non-Ti cranks out 6M ro 8M at 120W. That’s a bad return.

The 9600 XT makes comparable numbers to the 5050 at 120W. I know it’s being affected by HIP not being available, but for now it is also effectively wasting electricity. Given the right WU, CPUs are sometimes within spitting distance of its performance while SMU-throttled to 75C.

Taken together, this makes me think that my two current desktop machines would be most efficiently combined into one – again with the proviso that a non-negligible increase in cores is possible. At the same time a step up to a single GPU with more grunt than the 4070 would probably be the most efficient way to do work. I guess we’ll see how things pan out in eleven months or so…

2025-12-04 We’re back

Yeah, it’s not 2026 yet, but this is likely to be 2025’s only entry. I’ll keep it brief:

• We left WGC and all other BOINC projects in early 2024
• There was burnout around chasing the projects themselves, as they did and did not have work available
• We left FAH as well for roughly 18 months
• There was also burnout with maintaining the fleet and how much space it took up
• Returned to FAH in late October
• Only one fully dedicated machine now, and only three total. The other two spend the vast bulk of their cycles on FAH, but they do have other purposes
• Keeping it at three machines for the forseeable future
• All on Zen 5/AM5 for now. Upgrades are unlikely until Zen 6 is released, possibly late next year
• So until then, updates are likely to be sparse… but who knows?

Older updates

For older news, check out the

• 2022-2024 updates
• 2021 updates
• 2020 updates
• 2019 updates
• 2018 updates