Resource graph
Match the workload to memory, topology, network and software constraints before the job starts.
01 / COMPUTECAPACITY EXCHANGE
One control plane for accelerator capacity that lives in different places.
Search by accelerator, memory, topology, interconnect and region. ShardNet keeps provider inventory behind a common job interface without pretending every machine is equivalent.
SHARDNET / CAPACITY MARKET
ACCELERATORMEMORYSHAPEFABRICACCESS
H100 SXM80 GB8 GPU nodeIB / NVLinkrequest
H200 SXM141 GB8 GPU nodeIB / NVLinkrequest
B200192 GB8 GPU nodeNVLinkrequest
B300288 GB8 GPU nodeNVLinkrequest
MI355X288 GB8 GPU nodeXGMI / 400Grequest
RTX PRO 600096 GB1–8 GPUPCIe / 400Grequest
Availability and commercial terms are queried at request time.NO FAKE LIVE INVENTORY
02 / TRAINDISTRIBUTED EXECUTION
Training state should survive the machine it started on.
ShardNet forms bounded worker groups, records topology and run state, commits checkpoints, detects failure and replaces workers from compatible capacity when the network changes.
US-W / GROUP 18
APAC / GROUP 12
EU-C / GROUP 09
APAC-2 / GROUP 21
Run state / live demo
03:18:04SEALEDcheckpoint-012 committedoptimizer + RNG + model state
03:21:17FAULTworker apac-2/07 missed heartbeatcommunication group degraded
03:21:20PLACEMENTcompatible replacement selectedmemory / fabric / image matched
03:21:44RECOVEREDstate restored / group re-formedresume from checkpoint-012
Worker groups
Build explicit groups with role, rank, health and communication topology tracked by the control plane.
Checkpoint lineage
Persist model, optimizer, data position and random state often enough that recovery is routine.
Elastic recovery
Replace failed workers with compatible capacity and reconstruct the group from a known checkpoint.
03 / STACKCLI + CONTROL PLANE
A small interface over a messy world of providers, schedulers and clusters.
The CLI should feel the same whether a job lands on one node, an 8-GPU box or a larger multi-node reservation. Provider details remain visible as evidence, not exposed as operational busywork.
Start from a job, not a vendor console.
Describe accelerator requirements, image, storage and training command. ShardNet resolves compatible capacity and records the allocation in the run manifest.
SLURMK8SCONTAINERSPYTORCHJAX$ pip install shardnet $ shard run train.yaml --gpus 32 --class h200 resolving capacity... 4 nodes / 32 GPUs / high-bandwidth fabric manifest: run_01JQ7M... checkpoint policy: every 512 steps job submitted ✓ $ shard watch run_01JQ7M step 6144 · workers 32/32 · ckpt 12 sealed
04 / RESEARCHOPEN SYSTEMS
Distributed training gets useful when the failures are part of the design.
ShardNet research focuses on the systems problems that appear once training leaves one homogeneous cluster: topology mismatch, bandwidth variance, churn, state recovery and verification.
Checkpoint economics under volatile capacity.
How checkpoint frequency changes when replacement time and provider churn are part of the expected runtime.
READ NOTE →Heterogeneous placement without pretending hardware is fungible.
Memory, bandwidth, interconnect and software constraints belong in the scheduler's model of the job.
READ NOTE →Receipts for distributed training state.
What has to be recorded to audit who ran which shard, with which environment, against which checkpoint.
READ NOTE →Bring capacity.
Or bring a training job.
ShardNet is designed for operators with real accelerators and teams that need compute without stitching provider consoles together by hand.