twtw

48 ms·

421.
by twtw 8y ago
It's worth noting that modern GPUs have several megabytes of register file. That translates to a ~million registers, if I did the math correctly. Manageable parallelism is an answer to "how to use a bajillion registers."
422.
by twtw 8y ago
> If you would like to offer a solution to the fact that the vast majority of people in this country expect software to be free This is critical. I think it even goes beyond this: most developers expect software they use to be open sourc
423.
by twtw 8y ago
I can't answer the question, but I'm curious how you think NVIDIA should have introduced real time raytracing to the industry? They can make it proprietary at the API level, or proprietary at the ISA level, but at some point you&#
424.
by twtw 8y ago
Radeon Rays is more comparable to OptiX (2009), in that both are GPU accelerated ray tracing libraries. RTX (2018) differs in that it uses special purpose ray tracing units on the GPU, rather than running only on the SMs/CUs.
425.
by twtw 8y ago
> Google could in principle make its results insanely left-leaning and the first amendment should protect them This is a far too simplistic and superficial analysis. Yes, Google could decide to position themselves as a publisher and clai
426.
by twtw 8y ago
> digitized metadata What are you referring to here? It doesn't recognize you singing because it doesn't have the same spectral content as the full song. Shaman doesn't really identify tunes, it identifies recordings. (And
427.
by twtw 8y ago
Was it better when it first launched than DuckDuckGo is now?
428.
by twtw 8y ago
... And it's a better search because it has two decades worth of mindshare, and the revenue and data that came with it.
429.
by twtw 8y ago
> and it worked by creating new technologies that nobody expected. Where "new technologies nobody expected" means "built new user-friendly products on top of a technology that had been in development for 30 years, funded b
430.
by twtw 8y ago
HN title seems misleading. I read it as ITER has achieved first fusion, when in fact the news is a revised plan.
431.
by twtw 8y ago
Thanks for responding. I would however argue that the existence of hierarchy in physical memory is orthogonal to the existence of a linear address space. Caches work really hard to maintain the ability of programs to use a single linear add
432.
by twtw 8y ago
> we've always talked about Gilbert Strang as the best math professor who never actually worked at the university we went to. I very much relate to this. I've never attended the same school of any of my favorite professors.
433.
by twtw 8y ago
> But memory isn't a linear address space Would you mind expanding on this point a bit?
434.
by twtw 8y ago
Awesome, that makes a lot of sense. Thanks for helping me understand these things better.
435.
by twtw 8y ago
I first read about this when I was a teenager, and frequently think about it when making decisions, but have never once found that being aware of it has been helpful. In the face of uncertainty, you never know what is going to happen. I hav
436.
by twtw 8y ago
For people who are interested in these things, I'd highly recommend Chris Okasaki's thesis/book on the topic. I was not familiar at all with this stuff when I read through it the first time, so it was a tad mind-bending and I
437.
by twtw 8y ago
Thanks for your response. > When not all call sites can be found (for example in compilation of separate files), it is still possible to create two versions of (ADDER x), one that uses lambda lifting and one that does not. I'm start
438.
by twtw 8y ago
It looks like it's fast because it makes use of x86 and arm SIMD extensions, but you could probably get the scalar version to run on an arduino with some effort. It would likely be slow, though. If you are looking at arduino audio appl
439.
by twtw 8y ago
It is if you drop the resolution.
440.
by twtw 8y ago
DGX 1V has $7500 worth of CPU alone to feed the gpus. Throw in 8 TB of nvme ssd for training data and you're looking at something more like 1/5th the cost. V100 has ~50% higher memory bandwidth than 2080Ti, so you probably wouldn&
441.
by twtw 8y ago
Stealing JS example from another comment... function adder (x) { return function (y) { return x + y; } } How would your implementation handle this? Something like lambda lift + curry lifted function + parti
442.
by twtw 8y ago
Which is why it's nice that the pre-trained models are available. If you want to train on your own dataset, that price does not seem unreasonable to me.
443.
by twtw 8y ago
> actually a fairly complicated subject This is heavily dependent on your background. Deep learning is essentially the simplest thing that could possibly work (parameterized nonlinear function + SGD). If you have experience with calculus
444.
by twtw 8y ago
the reusability is the major differentiator versus Apollo missions.
445.
by twtw 8y ago
The closest I can think of is shared memory in CUDA / local memory in OpenCL.
446.
by twtw 8y ago
It's not ironic at all. C doesn't not do anything promoting a syscall per byte. Even if you think that's what you are doing, libc is probably buffering for you behind the scenes.
447.
by twtw 8y ago
A) that's a marketing image. The green box is placed over a random portion of the die, not just the raytracing units. (They are distributed throughout the die, not shoved on one side). B) that said, looks more like 30% to me.
448.
by twtw 8y ago
> Something like that may be at least 10MB of code - just for a silly animation. For reference, the Linux kernel can easily be made to fit in 10MB. I don't understand how it can possibly be code, it has to be assets.
449.
by twtw 8y ago
> It's annoying that 50% of the die is allocated to hardware Source on 50%?
450.
by twtw 8y ago
It's useful to have some background on where the existing regulations came from when thinking about these things. There is quite a pile of inconclusive research on the medical effects of microwave exposure, and as such none of it could

More ›