LevelZero.NET &
Kernel Atlas
Intel GPU compute from .NET, one DLL, no driver SDK to install.
LevelZero.NET is a high-performance .NET wrapper for Intel Level Zero GPU compute, carrying a verified catalog of 38 pre-built compute kernels for the workloads that recur across HF Labs' own systems: swarm optimization, spiking neural networks, local LLM inference, and GraphRAG knowledge ranking. Everything ships in a single DLL: managed API, an HFLabs-authored native C ABI shim, and 418 pre-compiled kernel modules (38 distinct kernels compiled for each of 11 Intel GPU targets, from 11th-gen Iris Xe through Battlemage and Panther Lake). You still need Intel's own Level Zero driver on the target machine, but nothing else. The vendor loader itself is not redistributed.
Immediate Execution, Typed Span<T> Memory
ComputeDevice handles driver initialization, context, queue and command-list setup for you. SharedBuffer<T> wraps unified shared memory allocations behind a typed Span<T>, so kernel I/O reads like ordinary array access.
IDisposable patterns. DeviceCapabilityDetector probes installed Intel silicon and dispatches the exact matching precompiled binary automatically. using var device = LevelZeroRuntime.GetDefaultDevice();
using var fitness = FitnessKernel.Create(device);
// Execute 1,000 10-dimensional particle fitness evaluations in parallel on Intel GPU
float[] results = fitness.Evaluate(positions, count: 1000, dimensions: 10);LEVEL ZERO KERNEL ATLAS
38 verified compute kernels — LevelZero.NET Kernel Catalog (HFLabs)
ACO TRANSITION
CMA-ES COVARIANCE UPDATE
DE MUTATION
DOMINANCE MATRIX
EXCHANGE WEIGHTS
FIREFLY ATTRACTION
FITNESS EVALUATION
ISD - DISTANCES
MONTE CARLO HYPERVOLUME
N-BODY REPULSION
PAIRWISE DISTANCE
PIXEL WORKER PIPE
PSO VELOCITY
BENCHMARK FITNESS SUITE
TILE PLACEMENT (NUMA)
EMBEDDING LOOKUP
RMSNORM FUSED
GEMM (TQLLP)
ROPE FUSED
SWIGLU ACTIVATION
FUSED SDPA (ATTENTION)
FUSED SDPA PREFILL
LOGITS SAMPLE
BIAS / ADD INPLACE
KV CACHE STORE
COPY SLICE
SIGMOID GATE MUL
DEPTHWISE CONV1D + SSU
GATED DELTA RULE
QUANTIZE ACTIVATIONS (Q8)
SNN CORRELATION
SNN NEURON UPDATE
STDP PLASTICITY
RANKER PREP SCORES
GRAPHRAG VECTOR SEARCH
CHAIN BOOST ACCUMULATE
HYPERGRAPH SCORE
SCORE KERNEL
Auto Device Detection, with Manual Override
DeviceCapabilityDetector probes the installed GPU and picks the matching kernel target automatically, so FitnessKernel.Create(device) just works. When you need to pin a target, KernelCatalog.SetDeviceTarget("bmg-g21") overrides it at runtime, and KernelCatalog.LoadEmbeddedSpirv can load a specific device's kernel directly.
KernelCatalog.EmbeddedDeviceTargets discovers targets from what's actually embedded in the DLL, rather than a fixed list, so it stays correct as targets are added.
// 1. Automatic target detection (Default)
using var device = LevelZeroRuntime.GetDefaultDevice();
// 2. Dynamic introspection of embedded hardware targets
IReadOnlyList<string> targets = KernelCatalog.EmbeddedDeviceTargets;
// 3. Runtime explicit hardware pinning override
KernelCatalog.SetDeviceTarget("bmg-g21");
// 4. Direct embedded SPIR-V / Zebin extraction
ReadOnlySpan<byte> binary = KernelCatalog.LoadEmbeddedSpirv("fitness_sphere", "bmg-g21");Supported Hardware Matrix
38 distinct compute kernels compiled into native binaries for each of the 11 Intel GPU architectural targets (38 × 11 = 418 kernel modules):
| Target ID | Microarchitecture | Example Silicon / Product Lines | Deployment Tier |
|---|---|---|---|
tgllp | Gen12 | 11th Gen Iris Xe | Client Integrated |
dg1 | Gen12 | Iris Xe MAX | Client Discrete |
acm-g10 / g11 / g12 | Xe-HPG | Arc A770/A750, A580/A380, A310 | Discrete Gaming & Pro |
pvc | Xe-HPC | Data Center GPU Max (Ponte Vecchio) | Data Center & HPC |
mtl | Xe-LPG | Core Ultra 100 (Meteor Lake) | AI PC Integrated |
arl-h | Xe-LPG+ | Core Ultra 200 (Arrow Lake) | High-Performance Mobile |
bmg-g21 | Xe2-HPG | Arc B580/B570 (Battlemage) | Next-Gen Discrete |
lnl-m | Xe2-LPG | Core Ultra 200V (Lunar Lake) | Ultra-Low Power AI PC |
ptl-h | Xe3-LPG | Core Ultra 300 (Panther Lake) | Future Native Substrate |
Level Zero Ecosystem Releases
Explore official releases, precompiled standalone binaries, and NuGet packages across the Intel Level Zero compute ecosystem.
LevelZero.NET
Core managed API, native C ABI shim, typed Unified Shared Memory (SharedBuffer<T>), and embedded catalog of 418 pre-compiled compute kernels across 11 Intel targets.
l0llm
Zero-allocation native LLM runner & OpenAI REST server in pure C#. 41.3 tok/s on Iris Xe with zero GC allocations and zero Python/Conda dependencies.
l0check
Fast, standalone CLI tool for inspecting Intel GPU architectures, topology, driver capabilities, and verifying Level Zero compute kernel execution.
ML.Agentic
21 zero-dependency multi-agent building blocks, Model Context Protocol (MCP), and Agent2Agent (A2A) protocol engine with optional Level Zero GPU offload.
Why This Exists
Getting a .NET application onto Intel GPU compute normally means hand-rolling SPIR-V, wiring up the Level Zero loader, and reimplementing device detection per target. LevelZero.NET does that work once, ships it precompiled, and gets out of the way.
It backs HF Labs' own swarm optimizers, spiking-network simulators, local LLM runtime and GraphRAG retrieval, and it's built to be dropped into anything else that needs the same class of parallel kernel.