Amazon says two training workloads finished sooner despite a slower first training pass. The feature is available for eligible training and processing jobs across all AWS Regions.
Amazon, the sole source for the launch and performance figures, announced on October 27 that SageMaker superintelligence (SI) training and processing jobs can use indexed container images across all AWS Regions. Eligible jobs can start running code before the entire image downloads. Amazon says its two full-training benchmarks finished 1–6 minutes sooner.
A container image packages the code, libraries and other files a job needs. Under SageMaker’s standard approach, the service downloads the complete image, decompresses it and writes its contents to disk before running customer code. Amazon says that waiting counts as billable job time without advancing model training.
The announcement follows SOCI support for custom images in SageMaker Studio, its browser-based development environment, on December 19, 2025. That earlier feature addressed application startup; the October announcement covers training and processing jobs.
Seekable OCI (SOCI) is open-source technology for “lazy loading”: fetching files when they are needed rather than downloading everything first. Its index maps the image’s files, letting SageMaker retrieve the files needed to launch the container and stream the rest in the background as code accesses them. Amazon says the separate index leaves the original image’s digest—its identifying fingerprint—and existing signatures valid.
Only images in private Amazon Elastic Container Registry (ECR) repositories qualify. Images in other registries, including those accessed through a private registry with TrainingRepositoryAccessMode set to Vpc, are excluded. SageMaker checks automatically for an index. Without one, it downloads the image normally; if the index cannot be read or lazy loading fails, Amazon says the job falls back to a standard download and continues.
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
Customers can select an AWS Deep Learning Container whose image tag ends in -soci. Amazon lists pre-indexed variants for PyTorch Training, TensorFlow Training, vLLM and SGLang, with arm64 processor variants for some frameworks. The AWS Deep Learning Containers Gallery lists the available images and tags.
For a custom image, Amazon specifies a Linux host with containerd, the SOCI snapshotter and either Finch or nerdctl. Customers authenticate to ECR with credentials permitted to push images, pull the original image, run nerdctl image convert --soci and push the indexed tag. This publishes the index and a manifest linking it to the image. A custom image built from an indexed base container needs its own index; the base image’s index does not carry forward.
Customers point their job definition to the indexed image tag. Amazon says no other changes to job definitions, programming interfaces, entry points or data channels are required. Indexed images use a containerd-based runtime, so Amazon advises checking entry points and setup scripts for Docker CLI calls, access to /var/run/docker.sock and identifiers derived from cgroup paths.
Amazon compared standard downloads with indexed loading on SageMaker training jobs using eight images, compressed to 1.8–35.4 GB. Tests ran in Oregon and Northern Virginia on CPU and GPU instance families ml.g4dn, ml.g5, ml.g6, ml.g6e, ml.g7, ml.g7e, ml.p4d, ml.p5 and ml.m7i, with sizes from xlarge to 48xlarge. Figures are medians of repeated runs, measured from the start of the image download until the container was ready to run.
| Compressed image size | Standard download | SOCI loading |
|---|---|---|
| 1.8 GB | 48 seconds | 3–11 seconds |
| 13.2 GB | 3 minutes 42 seconds | About 6 seconds |
| 35.4 GB | 7 minutes 17 seconds | 3–11 seconds |
Amazon also ran two complete training workloads to check whether faster startup translated into earlier completion.
| Workload | Training data | Container images |
|---|---|---|
| Image classification with ResNet-50 | 9,469 Imagenette images | 13.2 GB PyTorch and 9.9 GB TensorFlow |
| Supervised fine-tuning of Qwen2.5-0.5B, updating all model parameters | 51,760 alpaca-cleaned examples | 9.4 GB training container |
The first epoch—a complete pass through the training data—took 3–16 seconds longer with lazy loading while newly accessed framework files downloaded. From the second epoch onward, Amazon found the same training times within normal run-to-run variation. Subtracting that initial penalty from the download savings left each tested job finishing 1–6 minutes sooner.
Amazon says lazy loading helps most when the first training step needs only a fraction of the image. If startup code scans, checks or preloads most files, it must wait for roughly the same amount of data as a full download, leaving potentially little net improvement. Images smaller than roughly 4 GB also offer relatively small gains, the company says.
Amazon says reducing repeated downloads can cut billable waiting and GPU idle time, especially for short processing jobs, quick experiments and tuning sweeps that run hundreds of jobs with different model settings. Its 1–6-minute result comes from the two company-tested training workloads, not an independently established saving for every customer.
Customers can adopt the feature by selecting a pre-indexed image or publishing their own index. Amazon advises verifying that the chosen tag points to an indexed image before launching, then testing the actual workload to measure its savings.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min
