HOLL-I OPTIMIZED COMPUTING FOR AI
HOLL-I (Highly Optimized Logical Learning Instrument) is a new service at NCSA that offers University of Illinois researchers access to an extreme scale machine-learning capability, to complement other available resources at NCSA, such as Delta and HAL.
Leveraging the power of a Cerebras CS-2 Wafer Scale Engine (WSE), and with access to NCSA’s shared project storage on TAIGA, HOLL-I is capable of training large machine-learning models in short order. HOLL-I’s unique architecture offers higher-speed processing than anything currently available on campus. The CS-2 WSE is composed of 850K AI optimized compute cores providing an unprecedented computational power.
HOLL-I’s AI-specific targeting provides an economic choice for obtaining specialized compute resources without an allocation on or in addition to other systems like HAL or Delta. The service is operated by NCSA on a cost-recovery basis with access open to all University staff and researchers, NCSA Industry Partners and NCSA collaborators. Allocations will be reserved via the NCSA XRAS service.
To get started with allocations, request an NCSA Kerberos account. Account creation may take up to 24 hours once requested.
Once your user account has been created, you can submit proposals for allocation requests by visiting the NCSA XRAS Submit portal.
See documentation for link to login.
- 1 CS-2 WSE Accelerator
- 40 GB RAM on accelerator, loading/dumping at 120GB/s
- 850K AI optimized compute cores
- Support hardware and services for the CS-2:
- 9 Dell R6515 job prep and i/o nodes
- 64 AMD Epyc cores
- 768 GB DDR4 RAM
- 3TB local NVME storage (2 x1.6 raided)
- 1 Arista 100Gbe switch with 64 ports
- 192TB local ZFS storage for scratch
- Access to TAIGA shared NCSA LUSTRE project space
- TAIGA storage for project data pursuant to TAIGA guidelines
- You must provide a short description of the project (1 paragraph) and answer a few questions about the intended use of the system. If the description is unclear to us, we will contact you for more information.
- This is a cost-recovery service, and unless you receive a director’s allocation, you agree to finance the costs and compute time involved in your project.
- You agree to acknowledge the resource with the credit statement listed below.
- You also agree to adhere to the End User License Agreement between UofI and Cerebras.
ACKNOWLEDGE USE OF THESE SYSTEMS
To acknowledge the use of these systems in your research, please include the following credit statements with any publication you produce. We encourage researchers to let the Center know of any newly published work facilitated with our systems:
This work utilizes the National Center for Supercomputing Applications Highly Optimized Logical Learning Instrument (HOLL-I), supported by the Center for Artificial Intelligence Innovation, as well as the University of Illinois at Urbana-Champaign.
Fees below are per-minute costs for University of Illinois users and their collaborators that have a university funding account number (CFOP) or for Industry Partner use. Allocation/access cannot be shorter than three months with time purchased to be used within a three-month period. User decisions to terminate allocation/access must be provided one month prior to termination date. Costs include access to the full cluster of job support nodes and the CS-2 accelerator. Jobs are scheduled via a batch scheduler to ensure fair access to the unique resources across all node-types. Rates are reviewed and recalculated biennially to ensure superior infrastructure and services. Billing will occur monthly for the prior month’s allocation/use.
INTERNAL UNIVERSITY USER CS-2 COMPUTING
Full use of CS-2 and support nodes for U of I faculty and affiliates:
$4.31/minute (10-minute minimum per job)
EXTERNAL NCSA INDUSTRY PROGRAM PARTNER CS-2 COMPUTING
Full use of CS-2 and support nodes for Industry Partners:
$5.67/minute (10-minute minimum)
Contact NCSA for rates applied to account numbers residing outside the University of Illinois and not part of the Industry Partner Program.
Access to the entire CS-2 accelerator with the quantity of users dictated by the Principal Investigator. Billed monthly for prior month’s use. For consistent billing, users have the option to pay for a set threshold of CS-2 minutes per month, regardless of whether they use all the CS-2 minutes they anticipated.
1 TB per project on TAIGA will be provided free with an allocation for project usage. Additional project space will be available for purchase on TAIGA. Archival storage is also available through NCSA’s Granite service.
Per TAIGA storage rate: Additionally, access to the HOLL-I system provides 5TB of scratch space per project and 5GB of home space per user at no cost with access to any shared data on TAIGA with a max of 10 users per project as determined by the principal investigator. Additional Project data is expected to be purchased and stored on TAIGA.
CONSULTING, TRAINING, DOCUMENTATION, and ADVANCED SUPPORT
- Access to training, documentation, CS-2 vendor support, consulting, and advanced application support
- Performance analysis, software optimization, efficient use of accelerators, I/O optimization, data analytics, visualization, use of research computing resources by science gateways and workflows
- Security administration (e.g. software security vetting, external audits, compliance with University Security and Privacy Directive)
- System and security documentation (e.g. for proposals, contracts, agreements and reporting)
- Standard system administration (e.g. installs, updates, access management, infrastructure maintenance and testing)
Additional support beyond what we offer may be written into any proposed funding opportunities as personnel directly on your budget.
PUBLICLY RELEASED SOFTWARE
Projects and Publications
News & Events
- April 13, 2022, Three New Supercomputing Resources, NCSA
- May 2, 2022, Seminar, Cindy Orozco Bohorquez and Adam Lavely, Cerebras Systems
- Cerebras Systems, PyTorch Training