The hardware answer to the problem eBPF and WASM solve in software: a pointer that carries its own bounds and permissions, enforced in silicon. Still an unratified RISC-V extension, whatever the coverage implies.
Qubit counts make the headlines. The physical-to-logical ratio is the engineering problem, and nearly everything else in the field sits downstream of it.
The Server SoC, Boot and Server Platform specifications are ratified. What makes it interesting is the specification being nobody's property, rather than the instruction set itself.
Half the infrastructure in this notebook assumes it can ask the sky what time it is. Jamming and spoofing of satellite navigation are now routine over large parts of Europe and the Middle East, which turns a free, assumed-correct input into something an adversary writes. The engineering response is holdover, and almost nobody has specified it.
Quantised models running on a workstation, a Pi or a phone. The interesting constraint is memory bandwidth rather than FLOPS, which changes which hardware is worth your money, and a capable model on a machine you own means no prompt ever leaves the building.
Since October 2022 the specification of an accelerator you are permitted to buy has been written partly in Washington, in units of processing power and interconnect bandwidth. Export control used to be a legal department's problem. It is now a line in the architecture, and it is the first time in this notebook's span that a chip's spec sheet has been set by policy.
Performance per watt stopped being a mobile concern and became the constraint that shapes datacentre architecture. Most software has no idea how much power it draws, and the first team to put a wattmeter in CI will find savings nobody else can see.
Every attestation chain ends in something you trust for reasons that have nothing to do with technology. Knowing where yours ends is more use to you than the chain itself.
What attestation proves, what the trusted computing base holds once you have counted the firmware and what becomes possible when the machine you rent cannot read the work you run on it.
The legislation deals with parts and manuals. The practical barrier has moved on to signed firmware, parts pairing and who gets to say a component is genuine.
Performance and efficiency cores, unified memory, accelerators on-package. The uniform-core model that every piece of scheduling and benchmarking folklore assumes has stopped being true, and the machine that replaces it does more per watt than anything symmetric ever will.
A first-person-view quadcopter with a warhead taped to it costs a few hundred pounds and destroys a vehicle worth several million. The airframe is a toy, the radio is a toy, and the only component that was ever hard is now a model that fits on a board you can buy. That is a cost inversion, and cost inversions do not reverse.
A pacemaker, a cochlear implant and an insulin pump are embedded computers with firmware, update paths and signing keys. Every assumption this trade makes about patching, rollback and end-of-life was written for machines you can switch off, and none of them survives contact with a device that is inside somebody.
The architecture monoculture is ending, and that sits upstream of almost everything interesting in systems work. The hard part was never the instruction set; it is fifteen years of software that has only ever been built on one.
A neural network won CASP13 outright, and a sequencer that runs off a laptop's USB port costs less than a server. Biology is becoming a field where the expensive instrument is a model and the cheap instrument fits in a coat pocket, which changes who is able to do the work and who ends up holding the data.
Speculative execution was the free lunch of CPU design for twenty years, and it leaks straight across the privilege boundary everything else was built on. This is the week that will explain the next decade of isolation work.
A great deal of software is still designed around a disk that takes ten milliseconds to seek. That disk hasn't been in the critical path of a serious system for years, and the clever architectures built to work around it are now the bottleneck.
Five Raspberry Pis, a switch and a great deal of green cable. The compute was never the point; a real network between real machines fails in ways a single-host simulation will never show you.
