HNHacker News
TopNewBestAskShowJobs

ladberg

3,485 karma · joined August 10, 2017

Contact me at hackernews@leoadberg.com
submissionscomments
ladberg··on AI's Dial-Up Era
Do you know a single person who'd buy an iPhone without a camera? I don't
ladberg··on Smartphone Cameras Go Hyperspectral
They're not claiming to get that many values per pixel, they're getting that many values overall for the medium through which light passes between the card and the phone. The idea light comes from a source (e.g. sun), bounces off the various colors of the card and thus produces hundreds of different spectra, those all pass through a medium, and land on the phone camera. So you're getting one measurement consisting of hundreds of RGB values that each represent intensity of different spectra, and you combine it all together to get a single spectrogram.
ladberg··on Gemma 3 270M re-implemented in pure PyTorch for local tinkering
Ah yep read the labels backwards and meant that - ty for catching and for the explanation
ladberg··on Gemma 3 270M re-implemented in pure PyTorch for local tinkering
Given that the compiled version is slower than then eager version on A100, there's definitely something suboptimal happening there
ladberg··on HRT's Python fork: Leveraging PEP 690 for faster imports
I think you're misunderstanding what that page is: it's not an advertisement to invest with the company, it's an advertisement to trade via/with the company in the same way you might otherwise go manually trade from a Bloomberg terminal (or any other method).

There is no way to invest in the company, and the only way of becoming a "customer" is to engage in trading.

ladberg··on HRT's Python fork: Leveraging PEP 690 for faster imports
> Its a finance firm - i.e scam firm. "We have a fancy trading algorithm that statistically is never going to outperform just buying VOO and holding it, but the thing is if you get lucky, it could".

HRT trades their own money so if it didn't beat VOO then they'd just buy VOO. There are no external investors to scam.

ladberg··on Ollama's new app
I can't comment on whether these particular pieces were generated, but models are certainly good enough now to handle these cases and more
ladberg··on Jane Street barred from Indian markets as regulator freezes $566M
This strategy by definition wouldn't work if it was "background noise" because it relies on being able to move the market
ladberg··on Nvidia adds native Python support to CUDA
The main release highlighted by the article is cuTile which is certainly about jitting kernels from Python code
ladberg··on AI 2027
It doesn't claim it's possible now, it's a fictional short story claiming "AIs can do everything taught by a CS degree" by the end of 2026.
ladberg··on Is the world becoming uninsurable?
Health insurance does have profit caps, so like the sibling commenter said their margins are small (6%) but also decently under the cap (20%) in the first place.
ladberg··on Is the world becoming uninsurable?
Insurance companies have pretty thing profit margins regardless, even in areas where profits are not capped. It's a competitive marketplace!
ladberg··on Getting Faster for You
Fun anecdote, ridewithgps.com was originally shown on HN 15 years ago by the founder: https://news.ycombinator.com/item?id=1030015
ladberg··on Rosetta 2 creator leaves Apple to work on Lean full-time
For the record I was interning on Cameron's team while he worked on Rosetta 2 and didn't even know myself what he worked on (the rest of the team and I were working on something else). I only found out later after it was released!
ladberg··on URAvatar: Universal Relightable Gaussian Codec Avatars
Gaussian splats can pretty much be rendered in any off the shelf 3D engine with reasonable performance, and the focus of the paper is generating the splats so there's no real reason for them to mention runtime details
ladberg··on Create block devices on your computer backed by Redstone Memory
The linux boot sequence in a redstone ARM CPU might take 3 semesters by itself...
ladberg··on Taichi: Productive, portable, and performant GPU programming in Python
Idk about this particular library but the no-jit restriction wouldn't apply to GPU code where pretty much every platform is always jitting the code, and shipping AOT-compiled code is often the exception.
ladberg··on [dead]
You've determined this about a 3.5 hour long podcast within the 14 minutes it was out before commenting?
ladberg··on Fast Multidimensional Matrix Multiplication on CPU from Scratch (2022)
Glad it was a learning experience for you and I apologize if I came off argumentative at all! I was mainly so incredulous because this is my day job haha, so I have a bit more experience than most in the area.

It definitely sucks to be led astray and have time wasted by a bug inherited from the original repo though, sorry to hear that :/

ladberg··on Fast Multidimensional Matrix Multiplication on CPU from Scratch (2022)
I hate to say it but that simply doesn't work: you can't write out of bounds to trick the compiler, it'll just ignore your out of bounds work.

You can look at the generated sass on godbolt: https://cuda.godbolt.org/z/19excTxM3

Note that there are 1024 FFMA instructions in the loop but you would expect 16*8*BK = 2048. This would suggest half the operations are skipped, which lines up with the half of writes that are out of bounds being omitted.

After the compute loop when you're calculating the final result and storing it, you can see that the FFMAs referencing out of bounds indices write QNAN instead of any real results.

Is it possible that the NANs are what are messing with your tests? Those are notoriously hard to deal with correctly, but you should assert that the result doesn't have any NANs whatsoever.

ladberg··on Fast Multidimensional Matrix Multiplication on CPU from Scratch (2022)
Sorry I'm a bit lost here, could you explain the reasoning behind this and why it works?
ladberg··on Fast Multidimensional Matrix Multiplication on CPU from Scratch (2022)
Your `sum` array is only 64 elements but you're indexing with indices out of bounds, which is UB and the compiler knows it at compile time so it's skipping a bunch of work.

E.g. consider the line:

  sum[y2*8 + x2] += ...
In the final loop iteration when y2=15 and x2=7, the index is 127.
ladberg··on An experiment in UI density created with Svelte
I love plotly for all my graphics needs (mainly 2D but supports 3D too)! Can export to a standalone interactive html file, can be used as a pandas plotting backend, can be easily extended with some client-side JS if you want to add more interactivity to the final result
ladberg··on How to optimize a CUDA matmul kernel for cuBLAS-like performance (2022)
If this were true (and I highly doubt it) it's obvious how to make money from it: collect a 7 figure paycheck from Nvidia, AMD, or any FAANG.
ladberg··on How to optimize a CUDA matmul kernel for cuBLAS-like performance (2022)
cuBLAS is not really Nvidia's moat: every competitor has a workable BLAS implementation and some are even drop-in replacements for cuBLAS.

In fact cuBLAS and CUDA are kinda orthogonal in that you're either calling a pre-built cuBLAS kernel or writing your own CUDA kernel but not really combining the two.

I'd say CUDA shines more because of stability, documentation, community support + examples, and ability to use modern C++ features in GPU code.

ladberg··on Apple Maps on the web launches in beta
Your last line is right, it's definitely a location thing. I've been using Apple Maps for years in the various big cities I've lived and I strongly prefer it over Google Maps, but whenever I'm on vacation I'll switch to Google.
ladberg··on Crash course in deep learning for computer graphics
I think the graphics association comes from having written the implementation in HLSL. HLSL compute shaders are general-purpose (like CUDA) but often used in the graphics context, so graphics programmers might feel more comfortable with it?

Tbh I don't really know, I've done extensive ML and graphics work and I'd just pick the right tool for the job (PyTorch and just drop down to Triton/CUDA/whatever when you need custom GPU code).

ladberg··on Kawaii – A Keychain-Sized Nintendo Wii
The sensor bar isn't actually a sensor, just two IR blasters that the cameras on the wiimotes use for positioning.

You can use any two sources of infrared light instead!

ladberg··on Run CUDA, unmodified, on AMD GPUs
Just to check here, if you're given something like the following PTX:

  wgmma.mma_async.sync.aligned.m64n256k16.f32.bf16.bf16
Do you reverse it back into C++ that does the corresponding FMAs manually instead of using tensor hardware? Or are you able to convert it into a series of __builtin_amdgcn_mfma_CDFmt_MxNxKABFmt instructions that emulate the same behavior?
ladberg··on Run CUDA, unmodified, on AMD GPUs
I don't really see how any code that depends heavily on the underlying hardware can "just work" on AMD. Most serious CUDA code is aware of register file and shared memory sizes, wgmma instructions, optimal tensor core memory & register layouts, tensor memory accelerator instructions, etc...

Presumably that stuff doesn't "just work" but they don't want to mention it?

← PreviousPage 2 of 24Next →