The older HPE documentation reveals a significant GPU memory-addressing restriction, and NVIDIA’s current toolchain no longer builds for several historically supported GPUs. I’m stopping before simulation and saving an INSUFFICIENT EVIDENCE verdict: a four-CPU/GPU system is physically plausible, but an exact compatible configuration, performance advantage, and cost case remain unproven.
Verdict: INSUFFICIENT EVIDENCE. I investigated three four-socket platforms: physical feasibility is plausible, but exact GPU/runtime compatibility, Australian acquisition costs, and any advantage over simpler servers remain unproven; all workloads are NOT RUN, and existing baselines remain unchanged.
Do you authorize preparing the highest-value next experiment—a same-hardware comparison of ordinary NUMA-aware scheduling versus Matrix-inspired scheduling under partial CPU/GPU offload—with execution held until compatibility inputs and any costs are approved?
This is a bounded, deterministic behavioral/network-model study. It is not silicon validation, does not validate physical PHY rates, and does not prove improved reasoning quality. Historical Matrix research schedules remain on hold; this manually authorized study is isolated.
Model contract
Six candidates were compared:
T0 Central Star: all CPU/pillar payload traverses the coordinator/central switch.
T1 Distributed Perimeter: four-pillar bidirectional ring, one CPU injection point, separate control.
T2 Hybrid Matrix: ring plus P0-P2 and P1-P3 express links, one CPU injection point.
T3 Segmented Diagonal: P0-P1-P2-P3 segments plus P0-P2 and P1-P3; no invented P3-P0 edge.
T4 CPU Dual Injection: ring plus CPU injection at opposite P0/P2 nodes and CPU-local memory.
T5 Flux Alternative: two independent non-coordinator payload switch planes, dual-homed CPU and pillars, adaptive least-loaded routing; coordinator/NPU remain control-only.
Workloads W0-W8 covered local, collective, all-to-all, KV/memory-heavy, CPU export, recursive, recurrent, hotspot and migration-storm traffic. W9 injected link, pillar, payload-path and coordinator degradation. Deterministic seeds were 7, 19 and 43.
Fairness regimes:
Equal directed-link capability: capacity 1.0 per directed payload link.
Equal total interconnect budget: total directed capacity normalized to 16 units, with sensitivity at 12 and 20.
Offered-load sensitivity: 0.35, 0.60, 0.85 and 1.05. Energy is a declared normalized proxy using compute baseline, per-hop movement, switch/idle complexity, migration and control overhead, swept by +/-50%.
Baseline results
Baseline: offered load 0.85, high CPU export, page-local-fence migration. Equal-budget result uses budget 16.
Equal link capability
Aggregate mean-throughput ranking:
Table
Rank
Topology
Throughput
1
T5 Flux Alternative
1.007178
2
T4 CPU Dual Injection
0.948111
3
T1 Distributed Perimeter
0.867633
4
T2 Hybrid Matrix
0.867298
5
T3 Segmented Diagonal
0.866737
6
T0 Central Star
0.736978
Equal total interconnect budget (16)
Aggregate mean-throughput ranking:
Table
Rank
Topology
Throughput
1
T1 Distributed Perimeter
1.007951
2
T5 Flux Alternative
1.007734
3
T4 CPU Dual Injection
0.947915
4
T0 Central Star
0.888486
5
T3 Segmented Diagonal
0.878365
6
T2 Hybrid Matrix
0.868171
The T1/T5 aggregate difference is only 0.000217 and is not sufficient to claim a universal winner.
Workload winners under equal budget 16
Table
Workload
Winner
Throughput
p99 latency
W0 Local
T5
0.251414
2.0
W1 Collective
T0
1.026710
2.0
W2 All-to-all
T2
1.042319
2.0
W3 KV/memory-heavy
T0
1.464131
2.0
W4 CPU export
T5
1.618503
2.0
W5 Recursive
T2
0.997220
2.0
W6 Recurrent
T5
1.005175
2.0
W7 Hotspot
T4
1.059190
2.0
W8 Migration storm
T3
0.622740
3.0
Strongest falsification
Added Matrix express/diagonal links do not by themselves solve CPU-export pressure. Under equal budget 16, high export and load 0.85:
T5 CPU-export throughput: 1.618503
T2 Hybrid Matrix throughput: 0.350259
Gap: 1.268244
Therefore, any broad claim that selected diagonal/radial links are sufficient for heavy CPU export is rejected unless CPU injection bandwidth and adaptive routing are addressed.
Surviving mechanisms
Separating control from payload survived coordinator-degradation testing.
Distributed payload paths survived as useful, especially versus T0 under link failure.
CPU-local memory and dual injection survived on latency, energy proxy and hotspot behavior.
A simple perimeter remained highly competitive under a strict total-resource budget.
Adaptive dual-plane routing remained valuable for CPU export and irregular exchange, but its extra switch complexity prevents a universal energy win.
Migration finding
Global quiescence is rejected for ordinary tier migration. Under equal budget 16 and W8:
Table
Strategy
Mean useful-compute fraction
Mean delivered throughput
Global quiescence
0.532633
0.895326
Page-local fence
0.826500
0.621425
Region explicit staging
0.771400
0.807787
The higher delivered migration traffic of global quiescence does not compensate for the large useful-compute penalty. The next model must refine page/region ownership, dirty tracking, atomics, IOMMU invalidation and commit semantics before transparent migration can be accepted.
Latency and normalized energy proxy
At equal budget 16:
Best aggregate p99: T4, 1.777778.
T5 aggregate p99: 2.000000.
Best normalized energy-to-same-work proxy: T4, 1.307120.
T1: 1.335106; T5: 1.444947.
The top-three energy ordering did not change under +/-50% energy-coefficient sensitivity: T4, T1, T5.
Failure findings
Coordinator degradation retained approximately 0.998-1.000 payload throughput for separated-control topologies while control p99 rose to 4.2. This supports, but does not physically validate, control/data separation.
T0 retained 0.777393 throughput under one payload-link failure, with degraded throughput 0.809938 and 0.229288 dropped/lost work.
Pillar-unavailable cases retained roughly 0.564-0.566 because work targeting P2 was declared lost rather than incorrectly completed.
Incorrect and duplicate transactions were constrained to zero in the model; this is a model invariant, not proof that a hardware coherence protocol satisfies it.
Sensitivity and determinism
The winner changes with budget and load:
Budget 12, load 0.85: T4 leads at 0.947367.
Budget 16, load 0.85: T1 leads at 1.007951, narrowly over T5.
Budget 20, load 0.85: T5 leads at 1.008285.
Budget 16, load 1.05: T4 leads at 1.171058.
Repeated-run topology-order agreement across seeds was 0.919192. Broad rankings are mostly stable, but narrow wins must not be overstated.
Claim-evidence ledger
Table
Claim
Status
Evidence
Limitation
A central star is universally inferior
Rejected as universal
T0 wins W1 and W3 under equal budget 16
Workload/model dependent
Express diagonals solve CPU-export bottlenecks
Falsified
T2 0.350259 vs T5 1.618503 on W4
Abstract traffic model
Separate control protects payload under coordinator degradation
Supported behaviorally
~0.998-1.000 retention
No protocol/PHY validation
CPU dual injection is valuable
Supported
Best p99 and energy proxy; W7 winner
Attachment/PHY feasibility unknown
Global quiescence is acceptable for routine migration
Rejected
Useful compute 0.532633
Migration semantics are abstract
One topology is universally best
Rejected
Winner changes by workload/budget/load
Requires real platform calibration
Verdict
RETAIN_SELECTED_MECHANISMS
Retain control/payload separation, distributed payload paths, CPU-local memory, dual CPU injection as a candidate, and explicit coarse/page-local migration mechanisms. Do not retain a claim that the full Hybrid Matrix topology is superior. Keep T1, T4 and T5 as the leading physical-comparison candidates; retain T0 as a necessary negative/control baseline.
Required closing judgments
Best surviving architecture: No universal winner. For a balanced equal-budget baseline, T1 and T5 are statistically/behaviorally near-tied; T4 is strongest for latency, energy proxy and high-load/hotspot cases. The recommended composite hypothesis is T4-style CPU locality/injection plus separated control, compared directly against T1 and T5 rather than assumed superior.
Strongest falsification found: T2 express/diagonal links fail badly on CPU export without adequate CPU injection/adaptive paths.
Highest-risk remaining assumption: Abstract traffic/routing and migration/coherence semantics may diverge from real protocols; PHY, switch and memory-controller behavior are uncalibrated.
Recommended next experiment: Calibrate the simulator against a software-defined physical network and run identical trace-driven W2/W4/W7/W8 tests with forced faults.
Cheapest meaningful physical demonstrator: CPU host, four low-cost FPGA/SoC or development-board pillars, separated coordinator/control node, configurable direct/ring links, and two optional commodity payload switches for T5 emulation.
Can prove: Relative queueing, routing, CPU-injection, control-isolation, failure-recovery and trace-driven migration-policy behavior at demonstrator scale.
Cannot prove: HBM/DDR/CXL/UCIe bandwidth, package-level power, coherent shared-memory correctness at production scale, silicon joules, cooling, or reasoning-quality improvement.