Accelerated Storage for AI Training

Every second your GPUs wait for data is wasted money. TAG is a local caching proxy that delivers near-local throughput for AI training — with zero code changes.

Request early access View documentation

Key Metrics

• 99.4% GPU Utilization
• 5.7× Faster Warm Epochs
• 0 Code Changes

Architecture

Inside the training instance.

TAG sits between your training code and cloud storage, caching hot data on local NVMe SSDs. Click each layer to explore the full stack.

1. GPU cluster
   Compute running on premises or any cloud, at massive scale.
2. Training code
   Standard S3 API interface. Drop-in replacement requiring zero code changes to existing training scripts.
3. Tigris Acceleration Gateway
   Intelligent caching proxy runs as a sidecar on your training instance to accelerate data access.
4. Local cache
   High-speed NVMe SSD pool for frequently accessed data. Local NVMe performance with cloud-scale capacity.
5. Tigris Object Storage
   S3-compatible globally available storage with unlimited scalability and 99.99%+ availability.

LAYER.01 // COMPUTE

• H100_01
• H100_02
• H100_03
• H100_04

LAYER.02 // SOFTWARE LOGIC

import tigris as tg

model = tg.Train(
  "s3://model-data"
)
# Accelerated standard API

LAYER.03 // ACCELERATION GATEWAY

TAG

CORE ENGINE

LAYER.04 // LOCAL CACHE

LAYER.05 // PERSISTENCE

TIGRIS S3-COMPATIBLE STORAGE

How It Works

A local cache that speaks S3.

TAG runs as a sidecar on your training instance. Epoch 1 fetches from Tigris, epoch 2+ reads from local NVMe at disk speed. Drop-in S3 API — zero code changes required.

• Training Instance
• Training Code
• Acceleration Gateway
• Local NVMe Cache
• Tigris Object Storage

Near-local throughput

NVMe-speed reads after the first epoch. Training data served from local disk, not the network.

Zero code changes

Drop-in S3 API compatibility. Point your training script at TAG and it handles the rest.

Intelligent prefetching

Anticipates data access patterns to keep your GPU pipeline full and idle time near zero.

Cache Anywhere

Store once, access anywhere.

Deploy TAG across regions and clouds. Each instance caches locally while Tigris handles global replication — your training data is always close to your GPUs.

Performance

Keep GPUs saturated during training.

TAG delivers up to 200× the throughput your GPU can consume — ensuring your training pipeline is never bottlenecked by storage.

Entitlement throughput by shard size×N = headroom over GPU demand (134 samples/sec)

Read: Benchmarking ML Training Throughput on Tigris

Accelerate your training pipeline.

TAG is available in early access. Get near-local storage performance for your AI training workloads — across clouds.

Request early access Explore AI workload docs