Every plug-in I ship floats on an iceberg of arithmetic you never see. Impulse-response deconvolution. Schroeder backward integration. Swept-sine analysis. Cross-correlation for time-of-flight. Filter design by brute optimisation. None of it glamorous, all of it FFT-heavy, and lately it all ends the same way. Me, staring at a progress bar, losing my train of thought.
So I started pricing a local NVIDIA AI machine for the desk instead of renting somebody else's by the hour. This is where that research stands, and why the box under the desk keeps winning.
It's not the model. It's the loop.
DSP isn't one big calculation. It's ten thousand small ones, nose to tail, each one asking the same nervous question. Did that just get better, or worse? Design the crossover. Sweep the filter order. Re-measure the decay. Compare to the reference. Nudge. Run it again. Nudge. Run it again.
The magic was never in any single run. It's in how many runs you can afford before you forget what you were chasing.
And that's exactly where cloud services put the friction. Every experiment has a price tag, a network trip, a rate limit, and, for someone whose entire business is unreleased audio code, a little voice asking whether I really want to upload this. When the meter's running, you run fewer experiments. And running fewer experiments is the one thing guaranteed to slow the whole thing down.
What the local box changes
The machine I'm eyeing is an NVIDIA Grace-Blackwell desktop with a fat pool of unified memory, or a workstation card in a tower I already own. It flips three things at once.
The math goes wide. A GPU is thousands of little calculators all shouting the answer at the same time. And the maths that runs DSP, the Fourier transforms, convolution, matrix work, correlation, is precisely the maths GPUs were born to chew. NVIDIA's CUDA FFT library turns a big batched transform from "go make coffee" into "done before you let go of the mouse."
The meter stops. Once the box is paid for, an experiment costs electricity, not an invoice. Suddenly the parameter sweep runs ten thousand candidates instead of ten. Suddenly "measure it four times to be sure" is free instead of a splurge.
The work never leaves the room. Impulse responses, measurement sets, half-baked algorithms, all of it stays on a machine I own. No upload. No terms of service. No wondering who else got a look.
Where it actually bites
Straight out of the tools already on my bench.
Convolution and spectral work
Reverb, deconvolution, spectral processing, all forward-FFT, multiply, inverse-FFT. On a CPU, a long convolution is a coffee break. Batched on a GPU, you run a whole shelf of impulse responses in the time one used to take. That means designing a reverb by auditioning a hundred at once, not one, wait, next.
Measurement and verification
The time-of-flight and speaker-alignment tools live on cross-correlation and stubborn double-checks. Measure, measure again, take the median, refuse to trust a number until the spread goes quiet. That paranoia is dirt cheap on parallel hardware and painfully expensive everywhere else. More checks, more certainty, same clock on the wall.
The thing I couldn't do at all, DSP that learns
The frontier in audio right now is DSP that's part learned. Differentiable filter design, neural models of analogue circuits, denoising and separation trained on real recordings. Training that isn't a quick cloud errand. It's a loop you live inside for weeks. Doing it locally is the difference between a direction I can actually chase and one I just get to read about.
Let's be honest about the trade
It's not free and it's not magic. It's money up front instead of money as you go. The truly enormous frontier models still live in the cloud, and for those, renting is the right call. I'll keep doing it. There's setup, there's power, there's the maintenance a hosted service quietly eats for you. If you only need big compute now and then, the cloud is cheaper. Use it. No guilt.
But that's not the shape of this work. This is a tight loop of heavy, repetitive maths I want to run constantly, privately and overnight, and without watching a meter tick. For that shape, the machine under the desk isn't just cheaper over time. It changes what I'm brave enough to try.
The goal was never a faster computer. It was permission to run the experiment I'd normally talk myself out of.
The research continues. If it lands where it's heading, the next round of MEY tools, the measurement corner especially, gets to be more thorough, more certain, and a whole lot bolder about the maths it's willing to attempt.