Kevin Milczewski

@jhleath >ebs drives are like the disks that are attached to your laptop. they don't have any logic inside of them to allow multiple machines to coordinate on the same set of data. docs.aws.amazon.com/ebs/latest/use…

@Geenimetsuri @straceX Yeah. They compile to the same thing. Clang generates a branchless vectorized loop for both. godbolt.org/z/cdvE4Ecxr

@straceX Doesn't the compiler optimize those anyways?

The humble if statement is secretly destroying your CPU's performance. Look at these two snippets. They do the exact same thing, but Snippet B can execute up to 6x faster on unsorted data. How does removing an if statement speed up the code?

@Sirupsen The 1-2us "SSD" speeds look like block cache times. I'm pretty darn sure you're not getting through PCI + interrupt stack in 2 us. Hell, if your machine has deep sleep power modes on, the interrupt wake latency is over 100us.

updating napkin math benchmarks, it's been too long github.com/sirupsen/napki… using c4-standard-48-lssd now as a representative machine (turbopuffer sponsored) previously a dinky 6 core 2019 VM

@CalebChamberla6 Near norwalk? :P

Wheels turning on OSH Cut East. Serving you faster.

@max0x7ba @ChShersh I feel like the talk probably about the lock free queue itself where you already have data in L1d and want to shove a pointer through a queue. The cache coherency there is usually high if there is contention with traditional CAS based approaches. I want to find the talk!

The sub-10 ns hot paths cannot possibly afford executing software in the CPU, though. A load from CPU L3d cache latency is ~46 CPU cycles, 9.2 ns with a 5GHz CPU. A 5GHz CPU ticks 50 CPU cycles in 10 ns, executing multiple instructions simultaneously and out-of-order. Low-latency NICs with kernel bypass communicate with CPU and RAM over PCIe. Incoming network frames are streamed in 64-byte blocks as they arrive from the wire straight into CPU L3d cache lines. One-way PCIe transfer latency is ~500 ns. Within a 10-ns budget, a modern AVX-512 CPU running at 5GHz could load an entire 64-byte L3d cache line with one 64-byte ymm load instruction in 46 CPU cycles, and have 4 CPU cycles remaining to compute with these 64 bytes of data, but its stores won't become visible to anything else outside the CPU until many nanoseconds later.

An SWE came to me and asked a very basic question. I assumed he’s a Junior. Later, I learned he gives a talk, so I attended. Turns out, the guy is cracked as hell. He described a highly complex low-latency lock-free multithreaded system (sub-10 ns hot path processing) he built.

@ChShersh @thedcoffman What CPU is that running on? Sub 10ns implies Shared L2 unless it's some 5ghz monster? And maybe doing it without locked instructions for the write? So no CAS and opportunistic reads?

@thedcoffman ikr

@chatgpt21 So it recreated something that already existed and was highly text based? Not really surprising. Still cool.

Anthropic had 16 AI agents build a C compiler from scratch. 100k lines, compiles the Linux kernel, $20k, 2 weeks. To put that in perspective GCC took thousands of engineers over 37 years to build. (Granted from 1987 - however) One researcher and 16 AI agents just built a compiler that passes 99% of GCC's own torture test suite, compiles FFmpeg, Redis, PostgreSQL, QEMU and runs Doom. They say they "(mostly) walked away." But that "mostly" is doing heavy lifting. No human wrote code but the researcher constantly redesigned tests, built CI pipelines when agents broke each other's work, and created workarounds when all 16 agents got stuck on the same bug. The human role didn't disappear. It shifted from writing code to engineering the environment that lets AI write code. I don’t know how you could make the point AI is hitting a wall.

@Adriksh "no memcpy" -> writes a bytewise memory copy manually and stores the whole data in the ring. The way to do this right is to pair it with an object pool and pass the pointers through the ring. That is zero copy. Of course your object pool is likely to use the same structure.

@CalebChamberla6 Xometry went to crap when it became more B2B focused (among other things). These dinosaur companies are the heart of the problem you've fixed with innovation. Don't let them drag you down into their game of waste an inefficiency.

Manufacturing X, I have strategy question. OSH Cut started with one goal: make it easy to buy custom sheet metal and plate. We did that with custom self-serve software, providing instant prices and lead-times, instant manufacturability checks, and deeply integrated production systems. Our target customers - people like us - absolutely love it. Our Net Promoter Score is through the roof, in the mid-eighties. As we grow, we spend increasing time serving customers who value our systems less. Engineers and builders love us. Purchasing departments? Not so much. We have full-time people now whose only job is to service the endless stream of onboarding packets from customers. Custom NDAs, custom terms of service, net terms applications, quality flowdowns, audits, 25+ page onboarding forms, supplier portals, legal departments. I'm certain that this could unlock a lot of revenue from big enterprises. On the other hand, targeting big businesses erodes our core value proposition. Professional buyers don't care that they can upload parts and get instant prices, lead-times, and DFM feedback; they don't want to login to our system to get production status for their orders; they don't value our self-serve, instant model at all. The engineers at these companies adore us. But the professional buyers create a lot of friction. We can help overcome that by streamlining the onboarding process. We've been working on that. On the other hand, this creates a lot of overhead that enables revenue but that simultaneously dilutes our value proposition. Eg. we have a great enterprise business funnel, but to them, we are "just another shop." Quality is the table stakes, but after that, net terms and price often take precedence over speed and simplicity. Professional buying eliminates those kinds of quality of life factors by design. Even so, these enterprise requests are burying our team. So we face a choice - grow the enterprise support team, even though it's not a big revenue driver today, and even though it erodes our core value proposition; or start saying "no," understanding that this may have a huge long-term impact on revenue, and on the kind of business we are in the future. On its face, this seems a minor choice. Hiring another couple people is super easy and low risk. On the other hand, the kinds of customers we focus on can have a large impact on future OSH Cut. So this really boils down to strategy, not immediate cash flow and revenue growth. I bounce back and forth on this question all the time.

@CalebChamberla6 Just Play Factorio!

Minecraft needs conveyor belts.

@CalebChamberla6 Heck yessss.

What do you think? Wire EDM from OSH Cut?

@redbullracing That man is a monster, and I love it.

Late on the breaks, first out the corner 👏 Simply lovely from Max 💙

@CalebChamberla6 @fabworks_guy You should see some of the horrible things we've done with pulleys and gears made from stacked laser cut plates! When doing it with aluminum, the laser edge finish does wear belts quite a bit.

@fabworks_guy I'm a little worried about rapid wear and tear on the timing belt if the pulley layers aren't perfectly aligned

First-pass assembly built! Learned some things, spinning some changes.

@parthingle_x Small, high RPM gears yeah it's probably not the best. I've seen it done but takes a lot of care and you have to run them in a lot, and you'll never have the right tolerances. Large high torque gears work great. Stacked laser cut is the best way to get 7075 gears I've found.

Laser cutting gears is, indeed, not a good idea

@lauriewired And this is why you apply strong error detection codes to your data.

What happens if a cosmic ray hits a voting machine? In Belgium’s 2003 elections, a relatively unknown Communist Party candidate received 4096 extra votes…from a spontaneous bit inversion. It was more votes than was mathematically possible at that polling station.

@AgingWheels It almost looks like you slid into the curb!

Anybody need rescuing? Quad motors and incredible traction control is giving me a superiority complex

@DJSnM Is the RF environment up there easier or harder than terrestrial? One of the problems with wifi data rates is how clogged the channels are and needing application protocols that can deal with lossy networks.

The ISS has WiFi antenna on the exterior which are used for things like video feeds from space suits, but a few years ago JASA demonstrated HD video over WiFi from 600meters away from the station. youtube.com/watch?v=yv_eqn…

@AgingWheels Which car are they in now?

Turns out, my mower was 90% batteries

@AgingWheels EZCal Calipers have no business being as good as they are for the price.

This is going to look so cool when it's done, but drilling and tapping all of these holes suuuucks

@davidbrevik Marvel Heroes was well after data backup practices should have been well understood. Surprising!

Incredibly this was not the only time it happened! It happened twice in my career on two different projects (Diablo II, Marvel Heroes)

True story. Just weeks before gold master, we lost everything. We had to cobble the depot and source code back together from random machines. Some stuff was lost forever.