Aviator software:
The only way to pilot your model

Corsair’s architecture was designed with software in mind.
Our Aviator software stack is performant and easy to use, offering a seamless user experience.

dual card optimized

Scale your inference-time compute fearlessly to unlock new levels of autonomy and intelligence with Corsair and JetStream!

Meet skyrocketing AI demands with JetStream™, purpose-built I/O accelerator for AI inference

Power AI operations with millions of requests using our custom NIC built for accelerator-to-accelerator communications.

  • Deploy AI with blazing fast multi-node scale-up
  • PCIe form factor for easy plug-and-play with existing datacenter infrastructure
  • Built for ethernet-based open standards
Learn more about Jetstream
server diagram - d-Matrix

Corsair Dual Card​

  • 6400 mm2 silicon ​
  • 4800 TFLOPs MXINT8 dense
  • 19200 TFLOPs MXINT4 dense​
  • 4 GB Performance Memory @ 300 TB/s​
  • Up to 512 GB Capacity Memory ​
  • 512 GB/s card-to-card DMX Bridge​™

Blend of cutting-edge innovations in a industry standard FHFL PCIe card, connected as a pair with high-bandwidth DMX Bridge™​

Inference Server

  • 8 cards*
  • 19.2 PFLOPs MXINT8 dense​
  • 76.8 PFLOPs MXINT4 dense​
  • 16 GB Performance Memory @ 1200 TB/s​
  • Up to 2 TB Capacity Memory​
  • PCIE based scale-up

High compute and memory density and card-to-card connectivity for efficient AI inference​

*reference configuration

Inference Rack​

  • 8 servers (64 cards)*
  • 128 GB Performance Memory @ 9.6 PB/s​
  • Up to 100B models in Performance Mode​
  • Up to 16.4 TB Capacity Memory
  • 1T+ frontier models in Capacity Mode​
  • Scale-out over PCIe or Ethernet​

Ideal for both high interactivity and offline batched use cases​ deployed at scale.

*reference configuration​

Corsair is the first-of-its-kind AI compute platform. It offers unparalleled performance and efficiency for generative AI inference in the datacenter.

Ideal for tomorrow’s use cases where models will “think” more. Supercharging reasoning, agents and video generation.

Blazing Fast

x

interactive speed

Commercially Viable

x

cost-performance

Sustainable

x

energy efficiency

Performance projections for Llama-70B, 4K context length, 8-bit inference vs H100 GPU, results may vary




60,000 tokens/second at 1 ms/token latency
for Llama3 8B in a single server

30,000 tokens/second at 2 ms/token latency
for Llama3 70B in a single rack

Results are preliminary and subject to change.

High Throughput

Up to 20x higher throughput for generative inference on LLMs

Results are preliminary and subject to change.

Low Latency

Up to 20x lower inference latency for generative LLMs (lower is better)

Results are preliminary and subject to change.

Significant Savings

Up to 30x better TCO to help your bottom line

d-matrix video poster
Play the Video

20x better power
efficiency

20x lower
latency

10x lower
TCO

  • Unique Digital-In Memory Compute Architecture (DIMC™) Purpose-Built for AI Inference Acceleration
  • Integrated Performance Memory for Blazing Fast Interactivity
  • Off-chip Capacity Memory for Offline Batched Inference
  • Block Floating Point Numerical Formats for Inference Efficiency

Aviator software:
The only way to pilot your model

Corsair’s architecture was designed with software in mind. Our Aviator software stack is performant and easy to use, offering a seamless user experience.

Aviator Software

Integrates with broadly adopted open frameworks such as PyTorch and Triton DSL to allow developers to easily deploy their model. Built with open-source software such as MLIR, PyTorch, OpenBMC.

open source spot sized r

How it works

Hover over icons for more information

Our Locations

Tools

A comprehensive suite of tools to deploy and monitor d-Matrix cards, and analyze workload performance

Model Factory

A template of PyTorch models that enables distributed inference across multiple d-Matrix cards

Compressor

A model compression toolkit to enable efficient block floating point numerical formats supported by d-Matrix hardware

Compiler

A state-of-the-art compiler that leverages MLIR ecosystem and produces optimized code targeted for d-Matrix hardware

Inference Engine

A distributed inference execution engine that deploys inference requests efficiently to a cluster of d-Matrix hardware

Host Runtime

A runtime that manages host interaction with the d-Matrix cards

Chip Runtime

On-chip firmware that launches jobs asynchronously from the host and enables high utilization of the d-Matrix devices

Tools – A comprehensive suite of tools to deploy and monitor d-Matrix cards, and analyze workload performance

Model Factory – A template of PyTorch models that enables distributed inference across multiple d-Matrix cards

Compressor– A model compression toolkit to enable efficient block floating point numerical formats supported by d-Matrix hardware

Compiler – A state-of-the-art compiler that leverages MLIR ecosystem and produces optimized code targeted for d-Matrix hardware

Inference Engine– A distributed inference execution engine that deploys inference requests efficiently to a cluster of d-Matrix hardware

Host Runtime – A runtime that manages host interaction with the d-Matrix cards

Chip Runtime – On-chip firmware that launches jobs asynchronously from the host and enables high utilization of the d-Matrix devices

Scalable

System software easily integrates into inference servers for process spawning, pipeline management, and scale-out communication.

scalable spot sized

Frictionless

Compiler and runtime produces highly optimal placements and schedules for graph execution without manual intervention.

frictionless spot sized

Blazing fast

Commercially viable

Energy efficient

product nav block b bg
Contact Sales