GEMMs are dense O(N^3) work operations that have roughly the same access pattern and data reuse properties across all matrices. Of course, I’m simplifying things a lot here; tall-skinny and short-fat patterns are much harder to get performance out of but the spirit of the approach is the same as big square matrices.
Sparse LU solves have a different character. There is nowhere near O(N^3) work. You typically expect something closer to O(N^2) but getting performance out of these operations is notoriously difficult because it depends a lot on the sparsity pattern of the linear system. Making matters worse is that you may commonly have a sparse A that factorises to dense L and/or U matrices.
I'm working with small matrices (e.g. 10x10 to 100x100), where I believe the effect of caches/pipelines/registers/etc will kick in before the O(N^2)-vs-O(N^3) discussion. Then dispatching to the hardware accelerators (SME2 FMLA or AMX FMA) and doing a _dense_ solve with 512-bit vectors could still be faster than a sparse solve at small matrix sizes or NEON.
Though as mentioned elsewhere in the thread, these accelerators only offer throughput, and latency suffers...
They lost me when they advocate for global dependencies instead of bundling. Are you supposed to have one `python` in your machine? One copy of LLVM (shared across languages!) ? One `cuda-runtime`?
Cool! Does PLECS do any symbolic simplification for its equations? Without that step, we found it impossible to handle even the smallest examples for realtime audio circuit modeling.
We could call the state-space approach of PLECS quasi-symbolic, because it reduces circuits with linear components to state-space equations. As a result, it needs to solve for much fewer variables. The price to pay is that the obtained system is dense. But for power electronics, this is great, because a tailored implementation of RADAU5 can be used (one of the gold standard for very stiff ODEs). For real-time though, the system has to be built differently with companion models for the swtiches, especially to be able to run it on the FPGA.
On the SPICE side of PLECS Spice, we also perform this kind of reduction for some components if we can, using graph algorithms. For instance, a basic SPICE solver would use a zero voltage source as an ammeter, which adds not only a current variable, but also a nodal voltage because the element has two terminals. Currents can be bad for Newton's method, because they have a very different scale. Often, we can completely avoid this by computing the current from surrounding components. But SPICE is very far from anything real-time.
Cubasis and Blackmagic Camera are cross platform, not that "most people" would use these over whatever was preinstalled or the camera interface in their social app.
The Android audio latency issues were solved long ago with Pro Audio. Whether Android audio apps chose to use it is on them and the significance of latency on their audio app.
Unless you were using hardware from the 00's, or had a lot of packages in your startup.jl or something, I suspect your memory may be exaggerating quite a bit.
That said, julia startup times definitely were a painpoint, but have improved quite a bit over the last few years.
The way he approaches the problem, which essentially uses voxels, it shouldn't be too hard: for each voxel, compute the distance to the closest triangle and you have your SDF.
The thing is, you have a SDF and now what? What about textures and materials, animation, optimization, integration within the engine,... None of it seem impossible, but I won't call it easy.
For context, I'm working on a proper SPICE component-level Diode Ladder.
I tried this for laughs with Gemini 3 Pro. It spit out the same ZDF implementation that is on countless GitHub repos, originating from the 2nd Pirkle FX book (2019).
Since there is a Ursa Major project on github, made by an owner, who reimplemented this also based on observation, made into a plugin, I wonder how much was regurgitated by the AI agent.
> I've managed people for decades and this has been a common pattern. I'd have people come to me with their plans to be in the C-suite in five or ten years, based precisely on self-help advice like that. None of them ended up there.
According to the Gervais Principle, these people are Clueless. Of course they end up in middle management. The top spots are reserved for Sociopaths.
Yeah, this should be properly communicated.