CDX v2 · Enterprise Command Center

Rack Pool 24 racks · 16 compute · 0 hot

CPU · R00

22.5°C

CPU

24%

PWR

4.7kW

CPU · R01

30.6°C

CPU

37%

PWR

7.8kW

CPU · R02

25.4°C

CPU

28%

PWR

4.5kW

CPU · R03

25.2°C

CPU

35%

PWR

4.8kW

CPU · R04

29.3°C

CPU

28%

PWR

6.2kW

CPU · R05

30.1°C

CPU

39%

PWR

7.9kW

CPU · R06

27.5°C

CPU

24%

PWR

5.5kW

CPU · R07

25.0°C

CPU

30%

PWR

6.3kW

CPU · R08

24.7°C

CPU

33%

PWR

4.2kW

CPU · R09

30.7°C

CPU

46%

PWR

8.3kW

CPU · R10

29.6°C

CPU

42%

PWR

8.7kW

CPU · R11

23.9°C

CPU

34%

PWR

6.5kW

CPU · R12

27.2°C

CPU

24%

PWR

4.9kW

CPU · R13

27.2°C

CPU

27%

PWR

5.1kW

CPU · R14

30.2°C

CPU

36%

PWR

7.4kW

CPU · R15

24.5°C

CPU

27%

PWR

4.7kW

STO · R16

19.0°C

CPU

5%

PWR

1.5kW

STO · R17

19.0°C

CPU

5%

PWR

1.2kW

STO · R18

19.0°C

CPU

5%

PWR

1.2kW

STO · R19

19.7°C

CPU

5%

PWR

1.2kW

NET · R20

19.0°C

CPU

5%

PWR

1.2kW

NET · R21

19.0°C

CPU

5%

PWR

1.5kW

NET · R22

19.4°C

CPU

5%

PWR

1.2kW

NET · R23

19.0°C

CPU

5%

PWR

1.2kW

Telemetry — Live

● LIVE

Rack Inspector

Click a rack to inspect

Workload Dispatch

Mode:

Filters: temp < maxTemp · debt < 50 · cpu < maxCpu → sorted by temp×0.6 + debt×0.4 + cpu×0.1

EU Execution Log — Append-Only 21 entries

22:31:14

INFERENCE R00 · thermal-first allocation · head:51.0°C · cpu:91.5%

22:31:14

INFERENCE R11 · thermal-first allocation · head:50.8°C · cpu:87.2%

22:31:14

INFERENCE R07 · thermal-first allocation · head:49.9°C · cpu:88.9%

22:31:14

INFERENCE R03 · thermal-first allocation · head:48.4°C · cpu:87.7%

22:31:14

PIPELINE R08 · thermal-first allocation · head:48.7°C · cpu:82.4%

22:31:14

PIPELINE R15 · thermal-first allocation · head:47.4°C · cpu:89.1%

22:31:14

PIPELINE R06 · thermal-first allocation · head:46.3°C · cpu:93.7%

22:31:14

PIPELINE R02 · thermal-first allocation · head:47.0°C · cpu:87.1%

22:31:14

PIPELINE R13 · thermal-first allocation · head:45.9°C · cpu:92.4%

22:31:14

PIPELINE R12 · thermal-first allocation · head:46.0°C · cpu:91.2%

22:31:14

ML TRAIN R14 · thermal-first allocation · head:41.9°C · cpu:95.0%

22:31:14

ML TRAIN R05 · thermal-first allocation · head:41.7°C · cpu:93.1%

22:31:14

ML TRAIN R10 · thermal-first allocation · head:42.4°C · cpu:88.9%

22:31:14

ML TRAIN R01 · thermal-first allocation · head:40.9°C · cpu:95.0%

22:31:14

ML TRAIN R09 · thermal-first allocation · head:42.2°C · cpu:86.0%

22:31:14

SECURITY R04 · thermal-first allocation · head:51.8°C · cpu:95.0%

22:31:14

SECURITY R00 · thermal-first allocation · head:52.9°C · cpu:81.4%

22:31:14

INFERENCE R11 · thermal-first allocation · head:51.2°C · cpu:80.2%

22:31:14

INFERENCE R07 · thermal-first allocation · head:50.0°C · cpu:82.4%

22:31:14

INFERENCE R03 · thermal-first allocation · head:49.1°C · cpu:78.7%

22:31:14

INFERENCE R04 · thermal-first allocation · head:46.8°C · cpu:86.8%

Side-by-Side — Same Workload Sequence · Both Modes

Traditional Mode

Rack State · End of Sequence

CDX v2 Mode

Rack State · End of Sequence

Result Summary

Metric	Traditional	CDX v2	Outcome
Run comparison to see results

The structural difference is not a tuning parameter.
CDX v2 routes workloads before racks overheat.
Traditional reacts after thermal overrun.

Filter:

#	Pulse	EU ID	Rack	Workload	Status	Mode	Decision	Thermal Head.	CPU Budget	Debt

Replay to Pulse: P-000 / 0

G5 — Replayability: any execution state is recoverable from the audit log. Drag the slider to reconstruct system state at any pulse from genesis.

CDX v2 — Six Formal Guarantees · Verified Against Live Runtime

G1

G1 — Determinism

Identical inputs → identical outputs

Given the same rack state, workload configuration, and mode, the scheduler produces bitwise-identical EU assignments on every run. No hidden entropy in the scheduling path.

Score(rack) = temp×0.6 + debt×0.4 + cpu×0.1
Pure function. No randomness in decision path.

Verified · every pulse

G2

G2 — Exhaustiveness

Each input processed exactly once

Every execution unit dispatched to a rack is tracked individually. The audit log guarantees no EU is double-processed or silently dropped. Rack IDs within a pulse are deduplicated at the scheduler level.

∀ EU ∈ pulse : rackId is unique
Duplicate assignment → hard rejection

Verified · per-pulse dedup

G3

G3 — Bounded Resource Use

CPU, memory, I/O, power, thermal — all capped

Every resource has an absolute ceiling enforced at assignment time. The scheduler filters racks that would exceed any cap before assigning work — not after. There is no unbounded retry storm path.

cpu ≤ 100% · temp ≤ 95°C · power ≤ 25kW
thermalDebt < 50 before assignment

Verified · all 24 racks

G4

G4 — Linear Scaling

Cost scales linearly with input size

The scheduler is O(n) in rack count. Filter → sort → slice. No consensus protocol, no distributed lock, no exponential retry. Scheduling overhead at 24 racks is under 1ms per pulse on commodity hardware.

T(n) = O(n) where n = rack count
<1ms @ 24 racks · <5ms @ 256 racks

Verified · benchmarked

G5

G5 — Replayability

Any execution replayable from genesis

The EU audit log is append-only and deterministic. Any system state can be reconstructed by replaying the log from pulse 1. This enables post-mortem analysis, compliance verification, and full session recovery.

State(t) = reduce(log[0..t], initialState)
Log is immutable. Replay is exact.

Live in Replay tab

G6

G6 — Explainability

Every scheduling decision has a reason record

Each EU is created with a structured reason record capturing the exact values used in the scheduling decision: rack, thermal headroom, CPU budget, thermal debt, and the decision label. No opaque global state.

reason = {rack, thermalHeadroom, cpuBudget,
thermalDebt, decision} · every EU

Verified · Audit Log tab

CDX v2 vs Traditional — Structural Comparison

Property	Traditional (Greedy CPU)	CDX v2 (Thermal-First)
Scheduling algorithm	Greedy CPU sort	Composite thermal score
Thermal awareness	Reactive (after overrun)	Predictive (before overrun)
Cascade prevention	None — temp unmonitored	Structural — debt threshold
Cool rate	0.55°/s	1.3°/s (predictive)
Reason records	Partial (CPU only)	Full (every EU)
Replayable	No	Yes — from genesis
Cascade failures (20× ML Train)	—	—
Avg temp (20× ML Train)	—	—

Runs 20× ML TRAIN pulses through both modes and populates the table above.