InferX: AI Lambda-Like Inference Function as a Service

    InferX is an advanced serverless inference platform engineered for ultra-fast, efficient, and scalable deployment of AI models. It enables serving multiple models within a single environment, achieving cold start times of less than two seconds.

    With InferX, you can:

    Whether you're running LLMs, vision models, or custom AI pipelines, InferX delivers unmatched speed and efficiency for next-gen AI applications.

    Ultra Fast Cold Start

    Cold start GPU-based inference in under 2 seconds for large models (12B+).

    GPU Slicing

    Allocate only a fraction of a GPU (e.g., 1/3 GPU) per model to efficiently run multiple workloads in parallel.

    Super High Model Deployment Density

    Serve hundreds of models on a single node (e.g. 30 models and 2 GPUs in the demo), maximizing hardware utilization.

    80+% GPU Utilization

    Based on Just-In-Time scaling and super high deployment density, we can achieve 80% GPU utilization.

    Lambda-like AI Serving

    Automatically scale AI inference workloads with on-demand execution.

    Optimized Performance

    Reduce latency, improve cost efficiency, and streamline AI inference at scale.

    How It Works

    The InferX platform follows a streamlined process to deliver fast, efficient AI inference:

    InferX Cold Start Process

    How InferX achieves < 2 second cold starts

    1

    Snapshot Generation

    When a model is uploaded, InferX captures a complete memory snapshot of the fully initialized model, including GPU memory state, CPU memory state, and runtime context.

    2

    Pre-warm Phase

    Before any inference requests arrive, InferX loads the snapshot into CPU memory or high-speed Blobstore. During this phase, the model consumes zero GPU resources.

    3

    Inference Request

    When a request arrives, the API Gateway checks for a warm instance. If none exists, it signals the scheduler to perform a cold start using the pre-warmed snapshot.

    4

    Ultra-Fast Cold Start (< 2 seconds)

    The snapshot is loaded directly into GPU memory at PCIe bandwidth speeds. Unlike normal startups, there's no GPU initialization or model loading from disk, enabling cold starts in under 2 seconds.

    Traditional Cold Start

    2-10+ min

    Loading time from disk to GPU

    InferX Cold Start

    < 2s

    Direct snapshot restoration at PCIe speeds

    1

    Upload & Deploy

    Users upload their inference docker image, and InferX provides a unique URL. The system generates a snapshot of the container with both CPU and GPU states.

    2

    Pre-Warm Process

    Before inference requests arrive, InferX loads the GPU snapshot into CPU memory or InferX Blobstore with ultra-high throughput (PCIe bandwidth, tens or hundreds of GB/s per node). During this phase, the model consumes ZERO GPU resources.

    3

    Inference Request Processing

    When a request arrives, InferX checks for a warm instance. If available, the request is processed immediately with minimal latency.

    4

    Ultra-Fast Cold Start

    If no warm instance exists, InferX loads the snapshot GPU state to GPU memory from CPU memory or Blobstore. Unlike normal cold starts, there's no GPU initialization, and the start latency is limited only by GPU PCIe bandwidth. This enables cold starts in under 2 seconds, even for 12B+ parameter models.

    Platform Architecture

    InferX AI Function Platform is an AI-native serverless inference platform that serves AI models as AWS Lambda-like functions.

    InferX Platform Architecture

    A Lambda-like AI Function-as-a-Service platform

    User

    API Gateway

    OpenAI-compatible API interface for model inference

    Dashboard

    User interface for monitoring and management

    Scheduler

    Resource allocation and container management

    Blobstore

    High-speed storage for model snapshots

    Model Containers

    Warm and cold start container instances

    GPU Resources

    Shared GPU hardware with 90% utilization

    Key Performance Indicators:

    < 2s

    Cold Start Time

    90%

    GPU Utilization

    3500%

    Deployment Density

    Dashboard

    The graphical user interface exposed to users for model management, deployment monitoring, and performance tracking.

    API Gateway

    The API interface that supports OpenAI-compatible REST APIs for seamless integration with existing applications.

    Scheduler

    A cluster-level scheduler that assigns warm instances to server requests if they exist, or initiates cold starts when needed, optimizing resource allocation across the platform.

    Blobstore

    A high-throughput blob storage system specifically designed to store model snapshots and enable rapid transfer of container snapshot data to GPU and CPU memory during restoration.

    Advantages Over Traditional Systems

    The industry average for GPU utilization in Kubernetes inference platforms is typically only 10-20%. InferX solves the core limitations that cause this inefficiency:

    ✅ Handling Sporadic Requests

    Traditional platforms leave GPU-based inference instances idle when there are no requests. InferX co-deploys multiple models (potentially from multiple tenants) in a shared cluster, and since peak demand times rarely overlap, this dramatically decreases idle time.

    ✅ Improving High Availability

    Traditional platforms must deploy 1-4 idle inference instances for failover and redundancy, especially with 5+ minute cold start latencies. InferX's ultra-fast cold starts eliminate the need for standby instances.

    ✅ Enabling True Auto-Scaling

    For busy inference systems with inconsistent workloads, traditional platforms require idle instances for fast scale-out due to high cold start latency. InferX's sub-2-second cold starts enable true on-demand scaling without pre-provisioning.

    Ready to Transform Your AI Inference?

    Whether you're running LLMs, vision models, or custom AI pipelines, InferX delivers unmatched speed, efficiency, and cost savings for your next-gen AI applications.