Scorpio Smart Fabric Switch

Scorpio™ X-Series 320 Lanes

Largest Open, Memory-Semantic Fabric Switch
for AI Scale-up

Download Family Brief

Scorpio™ P-Series

Broadest Family of PCIe 6 Fabric Switches
32 to 320 Lanes

Download Family Brief

Purpose-built AI fabric switches

  • Scorpio X-Series 320 Lane: Largest Open, Memory-Semantic Fabric Switch
  • Scorpio P-Series 32 to 320 Lane: Broadest Portfolio of PCIe 6 Fabric Switches

Benefits

  • Simplify large scale-up topologies and improve performance with single-hop latency
  • Maximize token economics with Hypercast™ and In-Network Compute engines
  • Improve efficiency with memory-semantic based open and platform-specific protocols
  • Optimize operational resilience and platform performance with COSMOS software
  • Scale compute capacity with diverse accelerator optionality and system topologies

Scorpio Highlights

Intelligent AI Fabric switches engineered to maximize token economics

Boost Collective Operations

With Hypercast™ and In-Network Compute engines

High-radix, Single-hop Scale-up

Many parallel switches maximize GPU bandwidth

Platform Performance Optimization

Cosmos Platform Logo – Astera Labs Developer Ecosystem for CXL Products

Software stack enhanced beyond fleet management

Why Use Scorpio Smart Fabric Switches?

Purpose-built for the demands of production AI at scale

Performance

  • Optimize Fabric Efficiency

    Memory-semantics enables accelerators to access resources through native load/store operations and eliminates software overhead

  • Maximize Tokens per Watt

    High-radix switch simplifies scale-up topologies, cutting hops and reducing end-to-end latency across the cluster with lower power

  • Improve Time to First Token

    Intelligent fabric with hardware-accelerated Hypercast™ and In-Network Compute engines accelerate collective operations by up to 2x

Scalability

  • Deploy Larger Clusters

    Largest open, memory-semantic fabric switch, Scorpio X-Series supports increased cluster sizes with up to 80 accelerators per switch

  • Increase Compute Capacity

    Broadest family of PCIe 6 fabric switches, spanning 32 to 320 lanes, Scorpio P-series enables flexibility to rapidly scale across diverse topologies

  • Enable Optionality

    Software-defined architecture designed for open and platform-specific protocols to support diverse accelerators and AI platforms

Resilience

  • Maximize Rack Uptime

    Comprehensive failure isolation, non-disruptive firmware updates, and integration into OpenBMC management

  • Accelerate Deployment

    In-depth signal, link, and packet diagnostics provide end-to-end protocol visibility to accelerate debug and improve serviceability

  • Deploy with Confidence

    Rigorous testing with diverse PCIe hosts and endpoints in Cloud-Scale Interop Lab minimizes deployment risk and accelerates time-to-market

Scorpio Smart Fabric Switch Family

Fabric Switches built from the ground up for AI clusters and servers

Scorpio X-Series 320 Lane: Largest open, memory-semantic fabric switch

  • Hardware-accelerated Hypercast™ and In-Network Compute engines
  • High-radix enables rack-scale, low latency, accelerator clusters
  • Platform-specific protocol, performance, and resilience customization

Scorpio P-Series 32 to 320 Lanes: Broadest Family of PCIe 6 fabric switches

  • Optimize AI connectivity with right-sized fabric switches
  • Predictable and high-performance GPU-to-CPU/NIC/SSD data flows
  • Interoperable with a diverse ecosystem of PCIe hosts and endpoints

Example Topologies

High Radix, Multi-level Switch Scale-up

Scorpio X-Series 320 Lane: Expand GPU scale-up with multi-level switching fabric

High Radix, Single-hop Switch Scale-up

Scorpio X-Series 320 Lane: Many parallel switches maximize GPU bandwidth with single-hop latency

16 GPU AI Server

Scorpio P-Series 320 Lane: Maximize performance with single hop latency to all 16 GPUs

8 GPU AI Server

Scorpio P-Series 320 Lane: All-to-all bandwidth across all 8 GPUs and devices with a single switch

8 GPU AI Server

Scorpio P-Series 160 Lane: High bandwidth GPU connectivity with optional switch-to-switch ports

8 GPU AI Server

Scorpio P-Series 64 Lane: High performance GPU to NIC scale-out with modular server design

KV Cache Storage Disaggregation

Scorpio P-Series 160/320 Lane: Increase inference performance with disaggregated KV Cache appliance

Scale-Out NIC/GPU Optionality

Scorpio P-Series 48/64 Lane: High performance GPU-to-NIC peer-to-peer for diverse applications

Front-end NIC Data Ingest

Scorpio P-Series 48/64 Lane: High performance data ingest and storage fanout

Resources

Ordering Information

SeriesPart NumberProtocolLanesPackage Size [mm]
X-SeriesPlatform Specific up to 320 Lanes – Contact Sales
P-SeriesPF50641LPCIe 56429×29
P-SeriesPF60321LPCIe 63221×21
P-SeriesPF60481LPCIe 64821×21
P-SeriesPF60641LPCIe 66429×29
P-SeriesPF61601LPCIe 616045×45
P-SeriesPF63201LPCIe 632062.5×62.5

Why Your Mixture-of-Experts Model Is Only as Good as Your Fabric

Introducing Hypercast™ for Improved Intelligence Benchmarks and Tokens-Per-Watt Performance

 

 

Read the Blog