markhahn

17 ms·

241.
by markhahn 2y ago
little annoying to see the one-core-compared-to-whole-gpu comparisons - now decades past when this was an innocent wrong. compare a 500W GPU to all the cores of a 500W CPU, please. I'm not expecting the CPU (say, a 192-core AMD that
242.
by markhahn 2y ago
Oh, thank you for the validation! For three years I've been measuring, weighing, logging all my nails and hair and poop, and I thought I was alone!
243.
by markhahn 2y ago
I think you should compare synapse and weight counts if you want a measure of the network's state/capacity. If you want something closer to its compute power, compare neurons and hidden units.
244.
by markhahn 2y ago
the problem is that when you care (unpredictable but hot branches), there are probably too many of them to execute in parallel (use up all your speculation depth). 2^n, you know! not to mention that you burn a lot more power.
245.
by markhahn 2y ago
I think of it that way too. Entanglement is about information, and since reality is quantized, we can normally only talk about the state (wavefunction) of smallish sets of particles. "Measurement" means that a simple-state thing
246.
by markhahn 2y ago
Pretending that money is a social construct is not quite meaningless, but certainly non-constructive. It doesn't help understand or decide wisely. It's like saying that consciousness is a retrospective illusion. The point is tha
247.
by markhahn 2y ago
in the olden days of gp-cpu this was certainly true. is it still true? do you just mean "latency overhead for setting up a single PCIe transaction is much larger than flinging a cache line across QPI/etc"?
248.
by markhahn 2y ago
I think this is a really unfortunate way to explain it. The issue is not that CPUs have a lot of different instructions - hardly anyone uses decimal math instructions, for instance, and no one cares about baroque complex addressing modes.
249.
by markhahn 2y ago
The moving-data-around cost is conventional wisdom in GP-GPU circles. Is it changing though? Not only do PCIe interfaces keep doubling in performance, but CPU-GPU memory coherence is a thing. I guess it depends on your target: 8x H100s acr
250.
by markhahn 2y ago
to be clear, you're saying that ROCm works on a much larger range of GPUs than AMD's official support list? that's pretty exciting!
251.
by markhahn 2y ago
I always wonder about premises in cases like this. a pair of compression-power cycles is a nice way to question the usual. has anyone gone through all the features of a conventional engine and asked: what if there's an alternative? f
252.
by markhahn 2y ago
that project deserves a permanent web presence, if only a blog post!
253.
by markhahn 2y ago
Not completely, but it wasn't a huge place, and the communities overlapped, went to some of the same seminars, etc. It was all in Wean, pretty much clustered by floor. Andrew was very much about distributed computing ala Athena; SPIC
254.
by markhahn 2y ago
reminds me of instructions for using curl|bash. do people really want to do things they don't understand at all?
255.
by markhahn 2y ago
just a suggestion: let the curator's activity be called "preening" rather than "grooming"...
256.
by markhahn 2y ago
I'm imagining a reading curriculum that regards the irregularities as funny little foibles, oddities to point out and wonder about. Like if you go to a park and there's that one place it's too steep to walk, or the other tha
257.
by markhahn 2y ago
and automatic internet-supplied dependencies are a sign of supplychain vulnerability
258.
by markhahn 2y ago
I'm curious about the implications of "deeper relationship".
259.
by markhahn 2y ago
I always wonder: why do some people like to do this (spreading)? I wonder the same about people who are "distro tourists". The latter tend to spend a lot of time on what seems like unproductive diddling (desktop skins, etc).
260.
by markhahn 2y ago
https://en.wikipedia.org/wiki/Ultradian_rhythm in short: cycles faster than a day. ultra- (more, greater), -dian (daily). analagous to circadian (cycle per day).
261.
by markhahn 2y ago
Always first look at the raw data. Look at the scatter of the dots in raw expression plots (figure 1e dots). Maybe that's normal for this field, but adjacent scores are all over the place, usually >50% different. That is: very noi
262.
by markhahn 2y ago
how big is the market of people who have money to burn? or is it that some people mistakenly believe that bigger devices are more cost-effective? this part of the release is interesting: https://documents.westerndigital.com&#x2F
263.
by markhahn 2y ago
I'm curious whether this helps anyone - why and how. I suppose there's some value in minimizing the "cultural impedence mismatch" - using genz terms/style/jokes. But the quality is not as good as "normal&#
264.
by markhahn 2y ago
it's a very interesting product - sort of the lovechild of a 1980s bitsliced microcode system optimized for application-specific pipelines (systolic array, etc). do you know whehter it's had any serious design wins? I can easily
265.
by markhahn 2y ago
those weren't "real" cores. you know what current chip has FUs that it falsely calls "cores"? that's right, Nvidia GPUs. I think that's the answer to your question (pushing 20k).
266.
by markhahn 2y ago
I'm not really sure why you say "one FPU per core". are you talking about the programmer-visible model? all current/mainstream processors have multiple and parallel FP FUs, with multiple FP/SIMD instructions in fl
267.
by markhahn 2y ago
IB didn't invent RDMA, and it's not even the only way to do it today. it's also not amazingly great, since it only solves a small fraction of the cluster-communication problem. (that is, almost no program can rely on magic R
268.
by markhahn 2y ago
bad thing to joke about.
269.
by markhahn 2y ago
why would initramfs updates be slow? do you mean that most initramfses are large? how much time are we talking about?
270.
by markhahn 2y ago
cost and power are legitimate motives. I'd probably have started with a rpi-02 though: storage and ethernet over USB is not going to win any races, though it'll compete with what he ended up with...

More ›