twtw

Thus my second paragraph. You have to opt in to use other formats.

Casting to bigint doesn't work because the problem occurs when converting the decimal constant in the source to floating point. You would have to convince the parser to parse the constant as something besides a float.

Can't comment on the situation with other GPU languages, but CUDA on GPUs since fermi are 754 compliant, with the exception that certain status flags are unavailable.

Ah, thanks for the correction.

I don't currently have a license, so out of curiosity is 0.3 == 3/10 in wolfram?

Binary-coded decimal formats have more or less already lived and died (both fixed and floating point). They still have areas of applicability, but this idea is very much not a new one - x86 used to have native BCD support, but it was taken out in amd64 IIRC.

That failure had nothing to do with floating point specifically. The same failure could have occurred when converting any data type with larger range (including rationals or arbitrary precision numbers) to 16 bit integer.

I would say the fact that floats can have data larger than INT16_MAX is hardly a surprise. That was just a bug, not some great and surprising drawback of floating point.

Julia has built in rationals (as do a few other languages).

I'm not aware of any language (other than Wolfram) that defaults to storing something like 0.1 as 1/10 - i.e. uses the decimal constant notation for rationals, rather than having some secondary syntax or library.

It doesn't even illustrate that particularly well. As is, the page just seems to be pointing at floating point and yelling "wrong", with no information on what's actually happening.

By all means embrace the surprise and educate today's 10,000, by why not actually explain why these are reasonable answers and the mechanics here behind the scenes?

I don't understand all the crap that IEEE 754 gets. I appreciate that it may be surprising that 0.1 + 0.2 != 0.3 at first, or that many people are not educated about floating point, but I don't understand the people who "understand" floating point and continue to criticize it for the 0.1 + 0.2 "problem."

The fact is that IEEE 754 is an exceptionally good way to approximate the reals in computers with a minimum number of problems or surprises. People who don't appreciate this should try to do math in fixed point to gain some insight into how little you have to think about doing math in floating point.

This isn't to say there aren't issues with IEEE 754 - of course there are. Catastrophic cancellation and friends are not fun, and there are some criticisms to be made with how FP exceptions are usually exposed, but these are pretty small problems considering the problem is to fit the reals into 64/32/16 bits and have fast math.

The nearest doubles to each of these two decimal constants end up being roughly 2 apart. Whereas for fp32 both decimal constants are stored as the same float.

The late 1980s (after the Mead&Conway revolution) seems like such a magical time for VLSI to me, as someone who wasn't around at that time.

Does this spirit of getting a few engineers together and taping out a chip in a few months still exist anywhere?

Right, and the paper itself is very upfront about that. I don't have any objection to the research, which is likely very valuable, but I do find it strange that the article claims they solved two problems without mentioning that only one of the solutions was correct.

It would be great if the article said they solved one and made good progress on techniques for the second that will likely work on next generation hardware, but the article didn't say that. I don't see how it being a matter of scale makes this acceptable. It's like the U.S. national labs unveiling the world's first exaflop supercomputer, with a footnote indicating that in fact the computer is only 100 petaflops at the moment.

... and couldn't get the correct answers for LiH. Interesting that the article didn't mention this.

From the paper:

> For lithium hydride, LiH, we were not able to reproduce closely the ground state energy with the currently available hardware. When accounting for 3 orbitals and using a scaling factor of r = 4, we already had to use 1558 qubits, which is a large fraction of available qubits. To summarize: the investigated method in general works, but it might be difficult to apply it to larger systems.

Nice try. I've seen how this goes over with those types.

The parents escalate to the adminstrators, who either involve the student counsellors or talk to the teacher directly. One of these two tell the teacher how they need to better support the student because they need a "little extra help" or can't handle the stress required to turn in homework assignments.

I wish this stuff worked, but these parents typically don't want an answer that requires their student to work harder (or differently). They want the curriculum, assignments, or grading scheme to change until their child gets an A - and they don't care about mastery.

This sounds a lot like my mom's situation. In the last decade, parents and admins have made her job terrible.

The parents make similar demands regarding grades - if their kid does poorly on an exam, it doesn't even enter into their mind that they didn't study or are super lazy, it's automatically the teacher's fault for making the test too hard or teaching the material poorly. Same goes for kids failing a class because they never turn in assignments - the teacher is failing to sufficiently motivate their student.

This would be manageable but for the fact that the administration always takes the side of the parents. If a kid cheats on a test, it's a "coping mechanism" or a "useful real world skill - there's no black and white" - this for a kid who literally broke into a filing cabinet to get the answer key. When kids fail tests, the administration or counsellors pressure the teachers to make easier versions of the test, extend the time limits, or give credit for test corrections. Several time teachers have resisted, because this totally undermines the information content of grades and takes a massive amount of extra teacher time (grading, writing, proctoring), and it's gone over poorly for the teacher.

The end result of all this is that the most entitled children or those with the most "engaged" parents get As, even if they earn a C,D, or F on the baseline system, while students who earn Bs but don't squeak get As.

I've been trying to convince her to just give the whole class As for years.

It's a pretty awful experience - I'm surprised anyone still teaches and would have quit years ago.

It's probably a few kilobytes of symbols, a few kilobytes of crt0, some ELF overhead, and a few hundred bytes of user code.

Meh. Aras is comparing apples to apples, so I don't see how it matters.

I'm not sure what the takeaway here is. If desktop users don't need (or can't use) more cores, then don't get them - and call an end to the periodic performance improvements, because CPUs won't get faster without more cores. Maybe that's fine.

On the other side, high performance computing demands more computational power, and the only promising path forward (at this time) is parallelism. I'm not motivated by optimism, but by pragmatism.

As a sidenote, I don't understand this section:

> We look at computing for graphics, neural nets, signal processing, and Bitcoin mining, and we see a pattern: parallelizing algorithms work best on hardware that is (a) specifically designed to execute them, and (b) can’t do anything else!

The same hardware is used for all of these things, and it was only (arguably) designed specifically for graphics (and even there "specific" is not certain).

> wouldn’t creating OR and AND gates directly reduce the number of transistors

No. Gates require varying numbers of transistors to implement in CMOS.

This idea is almost exactly backwards - logic is reduced to nand because nand requires the fewest transistors, not because CMOS can't (or couldn't) make anything else.

Absorbance characteristics of sub 100 MHz are substantially different from those of multiple GHz.

How can HN just choose to be noncompliant? Aren't there penalties?

I don't understand the jurisdiction of GDPR very well, but I thought it applied to all EU users.

Obviously, I have a bias coming into this discussion (and so do you).

Mine comes from experience in a world of hardware and physical products that gives me a greater affinity for the business model of hardware companies than "open source friendly" advertising giants. I have a fascination with the war of words free software adherents wage against nvidia - the company could give out candy to children and it would get criticized because Linus gave them the finger.

I'll also admit that I am entertained by the no true scotsman squirming that occurs when I point at things that nvidia has done that don't support the story that it is the evil empire.

Cue the argument that these "don't count" because they use CUDA.

There are far more than two data points on all of the charts in the article. What chart are you looking at?

I wish the economics of the present were somewhat different, and that money didn't exist in the 21st century.

And yet, in the world we live in, I have a hard time faulting a corporation for not giving away their core products for free.

Were the CPU cores of Cell (PPEs) hard to program? I had the impression that the difficulty of the Cell was in the need to manage the 8 SPEs, not in writing software for the Power4 core.

The 8 core Carmel CPU in Xavier is not like the SPEs in Cell.

Homogeneous coordinates allow for affine transformations to be represented and performed with a matrix multiply. This includes perspective projection and translation.

One of the use cases for homogeneous coordinates is certainly to be able to achieve translation.

I don't follow. What do we gain by moving the homogeneous coordinate from last position to first?

I don't understand this:

> then it's just a sparse vector! Set the variables you want to!

Or this:

> [1] is the origin point in any number of dimensions.

Could you clarify?

Also, I don't think this book even discusses homogeneous coordinates. It would be sort of unusual for this type of general text and the only mention of "homogeneous" in the index is "homogeneous equation."

> There is no employee shortage, ever. Just a shortage of businesses owners who are willing to pay market rate for labor.

Is the situation the same for housing? Is there never a housing shortage, just a shortage of people who are willing to pay market rate for housing?

If not, what are your thoughts on what the difference is?

> So what these hoarders are doing is stealing the scooters -- rather than checking them out through the charger app, they're just grabbing them and tossing them in their van

This is not my understanding. The scooters I've seen throw an absolute fit if you move them more than ~1 meter without unlocking them via the app.

I don't know much about this, but I didn't think the scooters had to be returned to any particular place. So hoarders pick them up to charge them and when charged "return" them to their garage/backyard. In which case it's not as clearly theft, they are just returning them to an inconvenient place - but that's fine, they are dockless.

While interesting, I don't see how they make circuit theory more complete.

> The ideal memristor was initially and abstractly formulated as the flux–charge relationship: dΦ = M dq, (2) which, in order to be invertible, must have M defined either in terms of Φ, or q, or it must be a constant. The latter case coincides with the resistor, while if M is defined in terms of Φ ... [emphasis mine]

From my reading, this means if M is constant (like R,L,C are for "standard" components) it's just a resistor - not the miraculous "missing fourth."

Curious to get other perspectives on this.