Distributions

Spice.ai Enterprise runtime distributions for different workload requirements.

Spice.ai Enterprise provides multiple runtime distributions optimized for different workloads. All distributions are available in Enterprise; some are restricted or nightly-only in open source.

Supported Platforms

PlatformArchitectureMinimum CPU Features
Linuxx86_64AVX2, FMA, BMI1/2, LZCNT, POPCNT
Linuxaarch64 (arm64)NEON, FP16 (FEAT_FP16), FHM (FEAT_FHM)
macOSaarch64 (Apple Silicon)Native

Distribution Availability

Distribution / CapabilityOpen SourceSpice CloudEnterprise
Default (Data + AI)✅✅✅
Data-onlyNightly only✅✅
NAS (SMB + NFS)Nightly only—✅
Metal (macOS)✅✅✅
CUDA (Linux)Nightly only✅✅
Allocator variantsNightly only✅✅
ODBC connectorLocal build only✅✅
HTTP user-defined functionsLocal build only✅✅
WASM user-defined functionsLocal build only✅✅

Inline SQL user-defined functions (from: sql) are available in every distribution. The HTTP and WebAssembly tiers are shipped pre-built in Cloud and Enterprise distributions; open source users can enable them by building locally with the http-functions and wasm-functions cargo features. See User-Defined Functions for the full reference.

Default Distribution

Includes all standard data connectors, embedded data accelerators (Spice Cayenne, DuckDB, SQLite), AI/ML model inference (LLMs, embeddings), and search capabilities (vector and BM-25 full-text search). It links the system allocator; see Allocator Variants to run against a different one.

Data-Only Distribution

Excludes AI/ML model support. Provides a smaller binary size and reduced attack surface for workloads that only need data federation and acceleration.

NAS Distribution

Adds SMB and NFS data connector support. Enterprise-only for production use.

GPU-Accelerated Distributions

Metal (macOS)

GPU-accelerated AI/ML inference on Apple Silicon.

CUDA (Linux)

CUDA GPU-accelerated model inference. Supported compute capabilities:

Compute CapabilityGPUs
80A100, A30
86RTX 30xx, A40, A10
87Jetson Orin
89RTX 40xx, L40, L4
90H100, H200

The AWS Marketplace ECR registry does not carry a CUDA image; it publishes the Default, models, and jemalloc variants (see Docker). Contact us for a CUDA-enabled Enterprise deployment.

Allocator Variants

Different memory allocators can significantly impact performance depending on workload characteristics.

The allocator is selected when the runtime is built, and a build links exactly one. The Default distribution enables no allocator feature, so it links the system allocator.

The AWS Marketplace ECR registry publishes the jemalloc variant. Contact us about an snmalloc or mimalloc build.

snmalloc

Optimized for concurrent workloads.

jemalloc

Alternative allocator that may perform better for certain memory allocation patterns. Marketplace images carry the -jemalloc suffix and are published from 2.2.1-enterprise onwards:

docker pull 709825985650.dkr.ecr.us-east-1.amazonaws.com/spice-ai/spiceai-enterprise-byol:2.3.2-enterprise-jemalloc

mimalloc

Microsoft's mimalloc allocator, designed for performance and security.

System Allocator

Uses the system's default allocator (glibc malloc on Linux). This is what the Default distribution links, so the <version>-enterprise and <version>-enterprise-models images use it.

Choosing a Distribution

Use CaseRecommended Distribution
General purpose with AI capabilitiesDefault
Data federation only, minimal footprintData-only
Network attached storage (SMB/NFS)NAS
macOS with GPU accelerationMetal
Linux with NVIDIA GPUCUDA
Memory allocation tuningAllocator variants