ALCF AI Testbed

The ALCF AI Testbed provides an infrastructure for the next-generation of AI-accelerator machines. 

The AI Testbed aims to help evaluate the usability and performance of machine learning-based high-performance computing applications running on these accelerators. The goal is to better understand how to integrate with existing and upcoming supercomputers at the facility to accelerate science insights.

We are currently offering allocations on our Groq, Graphcore Bow IPUs, Cerebras CS-3, and SambaNova DataScale system. 

NOTE: There is no need to request an allocation for Metis (SambaNova SN40L). Metis is available to all users through our AI Inference endpoints service.

AI Testbed Links

Request an Allocation on Groq, Graphcore, Cerebras and/or SambaNova SN30

Apply for Time

AI Testbed Guide

Support

AI Testbed Training

Training

Email us for more information

Learn More

Systems

Cerebras CS-3 (Available through an allocation request)

Cerebras CS-3

  • System Size: 2 4 Nodes (Each with a Wafer-Scale Engine) Including MemoryX and SwarmX
  • Compute Units per Accelerator: 900,000 cores
  • Performance of a single accelerator (TFlops): 125,000 (FP16)
  • Software Stack Support: Cerebras Model Zoo, PyTorch
  • Interconnect: Ethernet-based
Cerebras Wafer-Scale Cluster

SambaNova Dataflow (Available through an allocation request)

SambaNova DataScale SN30

  • System Size: 64 Accelerators (8 nodes and 8 accelerators per node)
  • Compute Units per Accelerator: 1280 Programmable compute units
  • Performance of a single accelerator (TFlops): >660 (BF16)
  • Software Stack Support: SambaFlow, Pytorch
  • Interconnect: Ethernet-based
SambaNova DataScale SN30 Gen 2

Metis: SambaNova SN40L (Available to all users)

Metis: SambaNova SN40L

  • System Size: 32 Accelerators (16 Nodes and 2 Accelerators per Node)
  • Compute Units per Accelerator: 1,040
  • Estimated Performance of a Single Accelerator (TFlops): 637.5 (BF16)
  • Software Stack Support: SambaStudio, SambaStack
  • Interconnect: Ethernet-based
Metis, SN40L

GroqRack (Available through an allocation request)

GroqRack Inference

  • System Size: 72 Accelerators (9 nodes  x 8 Accelerators per node)
  • Compute Units per Accelerator: 5120 vector ALUs
  • Performance of a single accelerator (TFlops): >188 (FP16) >750 (INT8)
  • Software Stack Support:  GroqWare SDK, ONNX
  • Interconnect: RealScale TM
GroqRack at ALCF

Graphcore Bow Pod64 (Available through an allocation request)

Graphcore Intelligent Processing Unit (IPU)

  • System Size: 64 Accelerators (4 nodes x 16 Accelerators per node)
  • Compute Units per Accelerator: 1472 independent processing units
  • Performance of a single accelerator (TFlops): >250 (FP16)
  • Software Stack Support: PopArt, Tensorflow, Pytorch, ONNX
  • Interconnect: IPU Link
Graphcore Bow Pod64

Access and Support

Our AI accelerator systems are available to the research community with proven problems to solve. 

Researchers can now submit project proposals for the Groq, Graphcore, Cerebras CS-3 and SambaNova SN30 platforms via the ALCF’s Director’s Discretionary program. 

There is no submission deadline; proposals will be accepted throughout the year. 

Metis (SambaNova SN40L) is available through our AI Inference service.

Please contact support@alcf.anl.gov with any questions.

Success Stories

Bringing FAIR Principles to AI Models

Researchers extend the use of findable, accessible, interoperable, and reusable principles to scientific AI models and software.

Argonne Deploys New Groq System to ALCF AI Testbed

Argonne staff discuss how they use GroqRack system in the science campaigns.

Accelerating Scientific Applications With SambaNova Reconfig

Scientists at ALCF are exploring novel AI-hardware systems, such as SambaNova, in an attempt to address the challenges in scaling the performance of AI models. [Video]