ShardNet
High-performance scientific computing

Scientific
computing,
re-architected.

Numerical systems, mixed-precision research,
and accelerator infrastructure for next-generation
simulation and AI-scale computation.

Explore our research →
Our mission

We build the computational layer between scientific ideas and real hardware.

01Advanced numerical systems
02Mixed-precision research
03Accelerator infrastructure
↓

Research you can actually run.

ShardNet publishes working numerical experiments, failure boundaries, and machine-readable receipts instead of hiding behind benchmark claims.

01 / Solution fieldLive computation
Poisson · 64² reference gridMixed precision
SN-LAB-001 · Mixed-precision solvers

Poisson equation with iterative refinement.

Float64 establishes the reference. Float32 exposes the precision floor. The mixed path recomputes the outer residual in Float64 while solving corrections in Float32.

Unknowns—
FP64 residual—
Float32 residual—
Mixed residual—
Float32 error vs reference—
Mixed error vs reference—
ReadyResearch note
02 / Failure envelopeMeasured boundary
Hilbert matrix orderAccuracy loss
SN-LAB-002 · Failure envelope

Conditioning is where recovery methods tell the truth.

Hilbert systems become progressively hostile to finite precision. Mixed refinement helps, then fails sharply once the Float32 factorization is no longer a useful correction operator.

Mixed error · n=4—
Mixed error · n=6—
Mixed error · n=8—
Mixed error · n=9—
ReadyResearch noteOpen receipt

From numerical method to physical machine.

The work spans precision systems, scientific kernels, and scale-out execution. Hardware results stay marked pending until they actually run.

Compute path / 01—03Accuracy → execution → evidence
01

Numerical contract

Define the accuracy target and residual checks before choosing a precision path.

02

Kernel mapping

Place the actual computation against memory and accelerator constraints.

03

Machine evidence

Validate accuracy and execution behavior on the hardware that runs it.

Precision systems

Use lower precision only when the result survives.

Decomposition, refinement, reconstruction, and error-aware execution with explicit accuracy contracts.

Read the research ↗
Scientific kernels

Design around the accelerator actually executing the math.

Dense, sparse, spectral, stencil, reduction, and solver kernels for modern matrix engines and memory hierarchies.

Experiment plan ↗
Scale-out execution

Communication and topology are part of time-to-solution.

Multi-GPU decomposition, collective behavior, memory placement, and scaling measured as part of the algorithm.

Discuss a workload ↗