HNHacker News
TopNewBestAskShowJobs

coherentpony

2,363 karma · joined March 30, 2013

submissionscomments
coherentpony··on Making AMD GPUs competitive for LLM inference
Disclaimer: I have no idea how machine learning algorithms work.

I work with problems so huge they do not fit on a single device. Multiple devices each own a small piece of the global problem. They solve a local problem and they must communicate over a network in order to solve the global problem. This is almost universally true.

You would know more than I in this field, and I expect it really is better to swap a partition than it is to use a network.

There are certain methods in HPC applications that are almost universally avoided because of how terribly they scale to a large distributed memory system. Matrix multiplies are one of them. Outside of a handful of ab initio computational chemistry algorithms (which are incredibly important), basically the only reason someone does a large dense matrix-multiply on a supercomputer is usually because they're running a benchmark and they're not solving a real science problem.

Folks more knowledgeable than me here feel free to jump in.

coherentpony··on Making AMD GPUs competitive for LLM inference
What do you mean? Are you talking about capacity? Or bandwidth from RAM?

I'm in the HPC space, and pretty much everything I do on the GPU is bound by how quickly I can get data in and out of DRAM.

The point at which data motion to/from DRAM is not the bottleneck is when you do enough work per byte of data moved. How much work is that? On today's server GPUs it's in the region of 50--100 double precision floating point operations per byte. You can work out an exact number by taking the theoretical maximum floating point operations per unit time you can execute and divide by DRAM throughput (data moved per unit time).

O(50--100) double precision flops per byte is a _lot_ of work. We're talking BLAS-3 type operations. Anything level 2 or lower, or sparse operations, are typically bandwidth bound.

coherentpony··on Show HN: Van, truck or car camp for $0 a night
One of the many idiosyncrasies of the English language.

The sentence

> Don't make fire rings where they already exist

means both:

1. If there is an existing fire ring, do not re-use it. Make a new fire ring somewhere else instead; and

2. If there is an existing fire ring, please re-use it. Do not make a new fire ring somewhere else instead.

What they _meant_ was the second interpretation, and not the first one.

coherentpony··on Nvidia releases new AI chip with 480GB CPU RAM, 96GB GPU RAM
> 1 exaflop

To be clear, this is floating point quarter-precision operations when using the FP8 tensor core arithmetic unit [1].

[1] https://resources.nvidia.com/en-us-grace-cpu/grace-hopper-su...

coherentpony··on A formula for the nth digit of 𝜋 and 𝜋^n
> In "n’th digit of 𝜋 and 𝜋^n" are both n the same?

Yes.

> In other words, for say 𝜋^50 does the formula only give me the 50th digit

Yes.

> or does it give me any arbitrary digit I want?

It gives you the n-th digit of \pi^n. Here n is arbitrary.

coherentpony··on FreeBSD optimizations used by Netflix to serve video at 800Gb/s [pdf]
Not really. Ethernet and Infiniband are both perfectly capable from a bandwidth perspective. Streaming video isn't remotely close to latency-bound, which is where Infiniband would be better suited.
coherentpony··on Ask HN: Which books do you consider real gems in your field of work/study?
Of course. I don't disagree with any of that.
coherentpony··on Ask HN: Which books do you consider real gems in your field of work/study?
I'll counter. Don't tell my why you recommend it; tell me briefly what it's about instead and I'll decide for myself whether or not I want to read it.

I like to have moments to myself without feeling like I am being sold something. I feel like I am constantly being sold things on the internet; let me do the discovery myself. Let me decide for myself whether or not it is interesting to me.

coherentpony··on Ask HN: Easing wife onto Linux laptop How to?
I don't agree with "don't educate someone" being a defensible position.

I teach my wife things all the time, including technology. She teaches me things all the time, including technology.

coherentpony··on America has a drinking problem
> our Puritan values run deep.

Not at all. It's expensive and very harmful. Why would I drink alcohol when I can enjoy a meal out with friends without drinking at all?

coherentpony··on Killing TurboTax
I'll actually give a response that is not Turbotax specific.

The forms I get every year from the ~10 banks I have accounts with and the ~2 companies I earned a salary from are essentially just copies of forms that they also had to send to the IRS. So in principle the IRS already has the same information I do when it comes to how much tax I paid and on what income.

It is not clear to me why I have to share:

- My name

- My address

- My _social security number_

- My spouse's name

- The names of my children

- All their social security numbers, too

- How much money I made

with a for-profit company. This is information that the IRS already has. To me, this is totally unnecessary. I would much rather the government increase the IRS's budget so that they can implement services that provide any help I might need directly, rather than through a totally unrelated corporation. The IRS's job is to make sure you paid the tax you owed. They don't dictate how much tax you pay. I truly believe that with the right financial resources they can make it easier to make that payment (and correctly) for: 1) us; and 2) them.

For folks with very simple tax situations, this seems like a no-brainer. For folks with complicated situations, you are still more than welcome to talk to a tax professional for advice. But a company _doing your taxes_ for you seems unnecessary.

Wouldn't it be lovely if instead I went to www.irs.gov to file my taxes, and I was presented with a very similar interface to what for-profit companies are providing, but with the information already filled out because they already have it? Essentially, all I'd be doing is sanity-checking the inputs.

coherentpony··on Marketing Follies of 2020
The GP didn't claim something bad was going to happen to them if they followed the link. They simply stated it didn't load under uMatrix.

uMatrix doesn't stop bad things from happening to you. It's simply browser add-on software that can block specific types of content. In this case, probably trackers or ads. Perhaps the GP is attempting to maintain their privacy. We can't know their motivations with certainty because they never made them clear.

The person that put the page together might consider testing whether or not it loads under uMatrix before exposing it to the world.

coherentpony··on The State of Linux Debuggers
And you can make gdb send program output to a different terminal with the --tty option if you don't want program output to mess up your gdb windows.
coherentpony··on Ask HN: How do I get a word in edgeways?
On con calls? Or in person?
coherentpony··on CAPTCHAs don’t prove you’re human – they prove you’re American (2017)
Do you have a source for this statement?
coherentpony··on On Learning Chess as an Adult – From 650 to 1750 in Two Years
I have found that lichess has a more intuitive interface but the lessons and puzzles on chess.com are far superior. The explanations are clearer in the lessons, and I think the puzzles are less confusing on chess.com.

To each their own, though.

coherentpony··on “Overpaid Executive Tax” in SF hits firms that pay CEOs 100X more than workers
It reads totally differently if the first word is simply removed. It's remarkable how influential first impressions are.
coherentpony··on Apple’s A14 Packs 134M Transistors/mm²
Ultimately it depends on the design but you're basically right, yes.
coherentpony··on Apple’s A14 Packs 134M Transistors/mm²
These things aren't _really_ two-dimensional. They're not really three-dimensional either, but they are objects built out of layers of two-dimensional things. When you measure number of transistors per unit area you will inevitably see something more dense than the "number of 5nm square in a millimeter of area".

It's the silicon equivalent of measuring one's BMI.

coherentpony··on How Discord Won
Thanks!
coherentpony··on How Discord Won
Oh my god that sounds wonderful. Can I ask what the product is called? I'd like to check it out.
coherentpony··on The software engineering lifecycle: How we built the new Dropbox Plus
> I would like to add that one of the reasons I’m a paying subscriber of Dropbox is because it’s independent of all the major other providers.

If I remember correctly, it's using Amazon in the backend. So how exactly is it that it is independent of all the other major providers?

coherentpony··on Ask HN: Software for producing terminal-based tutorials?
Do you need audio?
coherentpony··on Tesla Q1 2020 Update
And for tech?
coherentpony··on Tesla Q1 2020 Update
Is 25% high? I have no clue.
coherentpony··on 0.999...= 1
> There is no proof that will ever satisfy a person dead-set against this.

Yes there is. There is a proof that uses only fundamentals of first year university analysis. When you see

0.99999....

this can be written as an infinite sum

\sum_{i=0}^\infty 0.9 x 10^{-i}.

Truncate the sum and set

S_n = \sum_{i=0}^n 0.9 x 10^{-i}

and now simply use the rules of arithmetic progressions to get the limit out:

0.1 S_n = \sum_{i=0}^n 0.9 x 10^{-i-1}

S_n - 0.1 S_n = 0.9 - 0.9 x 10^{-n-1}

0.9 S_n = 0.9 ( 1 - 0.1^{n+1} )

S_n = 1 - 0.1^{n+1}

Now let n tend to infinity to find the limit, which is 1.

You don't need to imagine infinity to do any of this.

coherentpony··on Coronavirus clue? Most cases aboard U.S. aircraft carrier are symptom-free
Are you of the opinion that this is simply just a slightly worse flu?

> All the data points to the fact that coronavirus is more widespread and less deadly than most outlets are reporting

You couldn't be more wrong. There are nurses and doctors working on the front lines dealing with patients, and they're completely and utterly overloaded. Lack of resources. Lack of help from local and federal governments. They're the ones telling us this is way more deadly than it seems.

coherentpony··on Ask HN: What are you learning?
I'm very much a software person, so I decided to learn more about hardware and the design of circuits. Pretty cool stuff.
coherentpony··on New MacBook Air
> The Pro is still a much better choice for programmers.

I'm not totally sure I agree. I typically do a lot of programming on my MacBook Air. It's simply a box I use to ssh to something else more powerful to do the actual runs.

Yes, if you're running computationally intensive stuff on your MacBook Air you're going to be disappointed, but if what you're mostly doing is typing into a terminal screen then it's probably perfect. And probably even over-powered? :)

coherentpony··on FreeNAS and TrueNAS Are Unifying
> Right, but both were being managed by a single group of developers - that's just inefficient and prone to cause bugs.

I totally agree with you, but that's a different statement than the one in the comment I originally replied to.

← PreviousPage 3 of 29Next →