The workhorse of Soviet science made concrete. The BESM-6 (1968), last of Sergei Lebedev’s BESM line, ran at roughly one million operations per second on a 48-bit word with normalized floating point, and it pioneered the conveyor — instruction pipelining — in the USSR. This panel is runnable: normalize a decimal into a (sign, exponent, mantissa) triple and round-trip it back, and watch pipeline speedup climb toward the stage count.
source S. A. Lebedev, the BESM series; BESM-6, 1968, Institute of Precision Mechanics and Computer Engineering, Moscow. Room: THE MACHINE. Lebedev is a SONIA keeper — this is his machine. Rendered, not quoted.
Sergei Alekseevich Lebedev built the USSR’s first stored-program machine, the MESM (Kiev, 1950), then the BESM series in Moscow. The line culminated in the BESM-6 (1968) — the workhorse of Soviet science for two decades.
MESM = Malaya Elektronnaya Schetnaya Mashina, “small electronic calculating machine.” BESM = Bolshaya…, “large.” The BESM-6 was the machine that put Soviet floating-point science on a fast, pipelined footing.
A BESM-6 word is 48 bits. A number is stored as a sign, an exponent (characteristic), and a mantissa — the fraction bits. The mantissa is kept normalized: its leading digit is significant, so the fraction lands in [0.5, 1) and no precision is wasted on leading zeros.
value = sign × mantissa × 2exp, mantissa ∈ [0.5, 1). The exact bit-field split (mantissa ~40 bits, characteristic sharing the rest) is a history detail; the math — that normalization pins the mantissa into [0.5, 1) and round-trips — is what the engine proves.
One unbroken line of machines, each Lebedev’s or his institute’s:
MESM (1950) → BESM-1 (1953) → … → BESM-6 (1968) → Elbrus (from 1978). The BESM-6 stayed in production and in service for >20 years; roughly 350 were built.
The pipeline idea (window 0) and the register/associative-memory ideas of the BESM-6 fed directly into the Elbrus superscalar machines that followed.
Feed one decimal number. The engine normalizes it into a (sign, exponent, mantissa) triple with the mantissa in [0.5, 1), then reconstructs the decimal. Set the pipeline k (stages) and N (instructions) to watch the conveyor speedup.
The mantissa is quantized to 40 bits to model the BESM-6 word; round-trip holds to a relative 1e-9, far inside that budget.
THE LIT REPRESENTATION — decimal → sign / exp / mantissa → back
| field | value |
|---|
reconstruct = sign × mantissa × 2exp = —
relative round-trip error = —
integer part in binary: —
PIPELINING — the conveyor: N instructions in k + N − 1 cycles
speedup(k, N) = N·k / (k + N − 1) → k as N grows, never above it.
| N | serial = N·k | pipe = k+N−1 | speedup |
|---|
| x | sign·m·2^e | rel err | ok |
|---|
And the conveyor bound: speedup(6, 1000) = 6000/1005 = —, speedup(6, 1e6) → 6 but stays under it. Every witness re-checks live.
What is LIT is the arithmetic in the centre: the normalized mantissa really lands in [0.5, 1), the round-trip really holds, and the pipeline speedup really is bounded by k. The performance numbers are the machine’s reputation; the math is the machine.
“The USSR had no serious computers.” Cut. The BESM-6 ran the numerical science of a superpower for over two decades — hardly a toy.
“Pipelining came from the West.” Cut — parallel invention. The BESM-6’s conveyor overlapped instruction stages in 1968; the idea was not one nation’s.
“A k-stage pipeline runs k times faster.” Cut. Only in the limit. Fill and drain cost k−1 cycles, so speedup = Nk/(k+N−1) < k for every finite N. The centre proves it; window 8 breaks it.
The red move: replace the pipeline law with the naive linear claim — speedup = k × N — pretending fill and drain are free. The number balloons past k, which is physically impossible.
Assert linear speedup and the witness (5) re-checks the bound speedup ≤ k: it fails and the badge turns red. A real pipeline can never beat one result per cycle times its stage count. The attack is caught.