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)YesYesYes
Data-onlyNightly onlyYesYes
NAS (SMB + NFS)Nightly onlyNoYes
Metal (macOS)YesYesYes
CUDA (Linux)Nightly onlyYesYes
Allocator variantsNightly onlyYesYes
ODBC connectorLocal build onlyYesYes
HTTP user-defined functionsLocal build onlyYesYes
WASM user-defined functionsLocal build onlyYesYes

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). From 2.4.0, the default Enterprise Linux binaries and the <version>-enterprise and <version>-enterprise-models images link jemalloc. 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 allocator 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 startup line of spiced names it, for example (allocator: jemalloc).

ReleaseDefault and models imagesAllocator variant image
2.4.0 and laterjemalloc-sysalloc: the system allocator
2.2.1 to 2.3.xThe system allocator-jemalloc: jemalloc

Contact us about an snmalloc or mimalloc build.

snmalloc

Optimized for concurrent workloads.

jemalloc

Allocator that may perform better for certain memory allocation patterns. From 2.4.0, the <version>-enterprise and <version>-enterprise-models images link jemalloc. For 2.2.1 to 2.3.x, the Marketplace images with the -jemalloc suffix carry it:

docker pull 709825985650.dkr.ecr.us-east-1.amazonaws.com/spice-ai/spiceai-enterprise-byol:2.3.5-enterprise-jemalloc

mimalloc

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

System Allocator

Uses the system's default allocator (glibc malloc on Linux). From 2.4.0, Marketplace images with the -sysalloc suffix carry it:

docker pull 709825985650.dkr.ecr.us-east-1.amazonaws.com/spice-ai/spiceai-enterprise-byol:2.4.1-enterprise-sysalloc

For 2.2.1 to 2.3.x, the <version>-enterprise and <version>-enterprise-models images link 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