Topology-aware partitioning.
Model, optimizer state, data, and communication are partitioned with real memory and network constraints in the loop.
DISTRIBUTED AI POWER GRID / TRAINING FABRIC
A power grid
for training.
ShardNet turns fragmented accelerators, datacenters, and idle capacity into one schedulable training fabric that can place work, survive churn, and keep evidence about what actually happened.
REGION A / 18 NODES
REGION B / 32 NODES
REGION C / 14 NODES
REGION D / 21 NODES
REGION E / 16 NODES
TRAINING JOB / MODEL-04264 SHARDS ACTIVEcheckpoint epoch 12 · elastic group healthy
PLACEMENTTOPOLOGY-AWARE
CHECKPOINT12 / SEALED
GROUP HEALTH64 / 64
RECOVERYARMED
RESOURCE GRAPH DISCOVERY ELASTIC GROUPS TRAINING CHECKPOINTS RECOVERY HETEROGENEOUS NODES PLACEMENT TRUST ATTESTATION TELEMETRY EVIDENCE RESOURCE GRAPH DISCOVERY ELASTIC GROUPS TRAINING CHECKPOINTS RECOVERY HETEROGENEOUS NODES PLACEMENT
01 / RESOURCE GRAPH
Capacity is only useful when the scheduler understands it.
Workers advertise memory, accelerator type, topology, network, software environment, availability, and trust signals into a live resource graph. Placement starts with the actual shape of the workload.
Placement policy
NODE CLASSMEMORYFABRICSTATESCORE
H100 / 8×640 GBNVLINKREADY0.98
MI350X / 8×2.3 TBXGMIREADY0.96
TPU / POD SLICESHARDEDICIQUEUED0.91
H200 / 8×1.1 TBNVLINKREADY0.95
02 / DISTRIBUTED TRAINING
Shard the job around the network you actually have.
Elastic worker groups.
Workers join bounded groups with explicit roles, health, synchronization, and progress telemetry rather than pretending every node is identical.
Evidence for every run.
Checkpoint lineage, node participation, failures, replacements, throughput, and training state are retained so distributed runs are reproducible and auditable.
03 / RESILIENCE
Assume the network will change underneath you.
Distributed capacity is valuable because it is abundant, not because it is perfectly stable. ShardNet is designed around recovery: checkpoint often, detect failure precisely, replace workers, rebuild communication groups, and continue from known state.
03:18:04checkpoint 12 committedSEALED
03:21:17worker region-d/07 heartbeat missedFAULT
03:21:20replacement candidate selectedREADY
03:21:44group re-formed / state restoredRECOVERED
03:22:01training resumed from checkpoint 12RUNNING
TRUST / 01Identity + policy
TRUST / 02Workload isolation
TRUST / 03Signed receipts
TRUST / 04Telemetry lineage
SHARDNET / OPEN NETWORK
Bring capacity.
Join the training grid.
Datacenters, independent clusters, and specialized accelerator fleets can become schedulable participants in one distributed training fabric.
JOIN NETWORK →