1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
//! The device energy ledger — first-class, because on this hardware class the story IS the I/O.
//!
//! Prices are per-node-operation costs of a device model. For the Z1-class costs (SPICE-derived,
//! pre-silicon; arXiv:2608.01615 Table IV) a WRITE costs ~21,700 Gibbs cycles and a READ ~239.
//! The vendor's own conclusion follows from these three numbers: the architecture wins where
//! "many local updates are performed between infrequent I/O operations" — and the ledger makes
//! that arithmetic executable instead of promotional.
/// Per-operation energy prices, in joules. These describe a DEVICE MODEL, not measured silicon,
/// unless the source says otherwise; keep the provenance in the name.
#[derive(Clone, Copy, Debug)]
pub struct Prices {
/// One Gibbs update of one node.
pub e_sample: f64,
/// One node value read out to the chip edge.
pub e_read: f64,
/// One node's couplings/bias/clamp state flashed.
pub e_write: f64,
/// Maximum sustainable full-graph reflash rate, Hz (None = unstated).
///
/// Read by [`Ledger::reflash_seconds`]: a workload that reflashes the whole graph faster than
/// the device can sustain is not a fast workload, it is an unphysical one, and a joules figure
/// computed for it prices a run that could not happen.
pub reflash_hz_cap: Option<f64>,
/// WHAT these numbers describe, and where they came from.
///
/// Not documentation. A joules figure is a claim about a machine, and a `Prices` without a
/// subject can be applied to any machine at all — which is exactly what happened here: every
/// fabric in the tree declared `Z1_SPICE`, so a Hitachi CMOS annealer and an FPGA both reported
/// Extropic's pre-silicon SPICE estimates, and the HTTP surface reported them for a plain CPU
/// run on a laptop. Nothing was lying; nothing had been asked to say whose numbers these were.
pub source: &'static str,
}
impl Prices {
/// No published or measured per-operation energy for this machine.
///
/// [`Ledger::joules`] returns `None` for these rather than a number, because the alternative —
/// borrowing another device's prices — produces a figure that looks exactly like a real one.
/// An unstated cost is a fact about the world, and reporting it is more useful than a guess
/// wearing a decimal point.
pub const UNSTATED: Prices = Prices {
e_sample: f64::NAN,
e_read: f64::NAN,
e_write: f64::NAN,
reflash_hz_cap: None,
source: "no published or measured per-operation energy for this machine",
};
/// Whether these prices describe anything.
#[must_use = "false means these prices describe no machine, and pricing a run against them produces a figure that looks exactly like a real one"]
pub fn is_stated(&self) -> bool {
self.e_sample.is_finite() && self.e_read.is_finite() && self.e_write.is_finite()
}
}
/// Z1-class prices from arXiv:2608.01615 Table IV (SPICE estimates for taped-out, uncharacterized
/// silicon; "measured" in the paper's prose is a misnomer the appendix itself contradicts).
pub const Z1_SPICE: Prices = Prices {
e_sample: 7.09e-15,
e_read: 1.692e-12,
e_write: 153.6e-12,
reflash_hz_cap: Some(1.0),
source: "Z1-class SPICE estimates, arXiv:2608.01615 Table IV — taped-out but uncharacterised \
silicon, not measured. Applies to that device model and to nothing else.",
};
/// Operation counts accumulated by a run.
#[derive(Clone, Copy, Debug, Default)]
pub struct Ledger {
pub samples: u64,
pub reads: u64,
pub writes: u64,
}
impl Ledger {
/// Energy under `p`, or `None` when `p` states no prices.
///
/// `Option`, not a number, and not zero. A machine whose per-operation energy nobody has
/// published does not cost nothing, and a caller that has to unwrap this cannot accidentally
/// print a figure for a device that has none.
pub fn joules(&self, p: &Prices) -> Option<f64> {
p.is_stated().then(|| {
self.samples as f64 * p.e_sample
+ self.reads as f64 * p.e_read
+ self.writes as f64 * p.e_write
})
}
/// The wall-clock floor this many full-graph reflashes implies, or `None` if the device states
/// no cap.
///
/// A run that reflashes faster than the hardware sustains is not fast; it is unphysical, and
/// pricing it describes something that could not have happened.
pub fn reflash_seconds(&self, p: &Prices, graph_nodes: u64) -> Option<f64> {
let hz = p.reflash_hz_cap?;
if graph_nodes == 0 || hz <= 0.0 {
return None;
}
Some(self.writes as f64 / graph_nodes as f64 / hz)
}
/// Fractional breakdown (sample, read, write) of total energy under prices `p`.
pub fn shares(&self, p: &Prices) -> (f64, f64, f64) {
if !p.is_stated() {
return (f64::NAN, f64::NAN, f64::NAN);
}
let s = self.samples as f64 * p.e_sample;
let r = self.reads as f64 * p.e_read;
let w = self.writes as f64 * p.e_write;
let t = (s + r + w).max(1e-300);
(s / t, r / t, w / t)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn write_to_sample_ratio_is_the_finding() {
// The structural number behind the robotics verdict: one write = how many samples?
let ratio = Z1_SPICE.e_write / Z1_SPICE.e_sample;
assert!((ratio - 21664.0).abs() < 100.0, "write/sample = {ratio}");
let rr = Z1_SPICE.e_read / Z1_SPICE.e_sample;
assert!((rr - 238.6).abs() < 5.0, "read/sample = {rr}");
}
#[test]
fn unstated_prices_produce_no_number_rather_than_a_wrong_one() {
// The failure this prevents: every fabric in the tree once declared Z1_SPICE, so a Hitachi
// CMOS annealer and a laptop CPU both reported Extropic's pre-silicon SPICE estimates as
// their own energy. Nothing was lying; nothing had been asked whose numbers those were.
let l = Ledger { samples: 1_000, reads: 10, writes: 1 };
assert!(l.joules(&Z1_SPICE).unwrap() > 0.0);
assert_eq!(
l.joules(&Prices::UNSTATED),
None,
"a device with no published per-operation energy has no joules figure, and zero would \
be a claim that it costs nothing"
);
assert!(!Prices::UNSTATED.is_stated());
assert!(Z1_SPICE.is_stated());
assert!(
Z1_SPICE.source.contains("SPICE") && Z1_SPICE.source.contains("not measured"),
"the provenance has to travel WITH the numbers: {}",
Z1_SPICE.source
);
}
#[test]
fn a_program_load_is_charged_as_a_write() {
// On this hardware class a write costs ~21,700 samples, which is the ledger's whole thesis.
// No implementation charged it, so every figure the stack produced was a sample-and-read
// story with the expensive term silently zero.
use crate::fabric::{Cpu, Device};
use crate::ftp::Program;
let mut cpu = Cpu::default();
assert_eq!(cpu.ledger().writes, 0);
let sched = crate::schedule::Schedule::default();
let p = Program::from_graph(&crate::ising::lattice2d(4, 1.0), &sched);
assert!(cpu.program(&p).is_empty(), "a pairwise lattice loads on the CPU device");
assert_eq!(cpu.ledger().writes, 16, "one write per node flashed");
}
#[test]
fn the_reflash_cap_turns_writes_into_a_wall_clock_floor() {
// A workload that reflashes the whole graph faster than the device sustains is not fast,
// it is unphysical -- and pricing it describes a run that could not have happened.
// `reflash_hz_cap` was declared and read by nothing.
let l = Ledger { samples: 0, reads: 0, writes: 500 };
// 500 node-writes over a 100-node graph is 5 full reflashes; at 1 Hz that is 5 seconds.
assert_eq!(l.reflash_seconds(&Z1_SPICE, 100), Some(5.0));
assert_eq!(
l.reflash_seconds(&Prices::UNSTATED, 100),
None,
"a device that states no cap implies no floor"
);
}
}