Peripheral vs. part of the machine
Toggle the two ways to attach a QPU to a supercomputer.
From abstract sketch to measured system
Hover the dots. The 2017 accelerator framing meets a 2025 measurement.
Who built it
NVIDIA plus nine institutions, thirty authors.
Reaction time: last measurement → correction lands
Every microsecond of waiting is idle time for the qubit. Illustrative exponential-decoherence model.
Throughput is a different failure mode
If decoding each round takes longer than the syndrome cycle, backlog snowballs until the machine stalls.
The building blocks
Hover or tap any block. Dashed flows are the real-time interconnect.
Why Ethernet instead of a PCIe card
Slide the number of PPUs. A host has a handful of slots; a switch fabric keeps scaling.
One program, many devices: device_call
From inside a __qpu__ kernel, call a function on a CPU, GPU or PPU. Swap the PPU for its emulator.
Round-trip distribution
Mock sample (n = 2000) drawn to match the reported mean, spread and max.
Against the yardsticks
The paper's own scaling example assumes 20 µs.
Anatomy of the loopback test
Step through one 32-byte packet.
Unreliable on purpose
Retransmission jitter versus a cleanly lost packet. Schematic, not measured.
How much classical compute per decoder?
Two decoder families, two growth laws. Toggle series and the AI-decoder scaling factors.
Evidence map: what is shown vs. proposed
Editorial assessment from the discussion. Cold blue = no evidence, red = demonstrated. Hover a row.