|The cost for the new NVIDIA Ampere A100 GPUs that power the DGX A100 nodes in HiPerGator AI will be measured as a multiple of the Normalized GPU Unit (NGU), which can be purchased as hardware or leased at the rates published on this page.
Although the NVIDIA Ampere A100 GPU is significantly more powerful, the average application performance ratio between the new A100 and the RTX 2080ti and RTX 6000 GPUs is not yet known. As a result, until July 31, 2021, the HiPerGator scheduler will treat the A100 as equivalent to the other GPUs. This means that an investment in 1 NGU will give access to 1 A100 GPU or 1 RTX 2080ti GPU or 1 RTX 6000 GPU.
We expect that access to 1 A100 GPU will require 2, 3, or 4 NGUs. In planning for proposal budgets that include use of the A100, you should keep this in mind. We will evaluate the data in July 2021 and update the scheduler policy accordingly.
|Purchases made by UF faculty members and collaborators are heavily subsidized.
Commercial pricing is also available for qualifying organizations.
UFIT’s Research Computing resource access is available either by purchase of hardware or through a charged service. See the Allocations policy page for more details. Note that when purchased as a service, the cost of resources is 10% higher than equipment purchased for its lifetime. Training and consultation services are also available to cluster users.
Select the price sheet appropriate for your group and decide what level of investment will meet your research needs:
For the sake of alternative cost comparison, we use Amazon Web Services’ Elastic Computing 2 (EC2) price points. The EC2 instance type most similar to UFIT RC’s Normalized Compute Unit (NCU) is the c4.large instance, which comes with 1 core (2 vCPU) and 4GB RAM (3.75 GiB) – equivalent of 1 NCU purchased from UFIT RC – and utilizes high frequency Intel Xeon E5-2666 v3 (Haswell) processors optimized specifically for EC2.
The fully burdened cost to the University of Florida for delivering 1 core hour on HiPerGator is $0.031. This includes the cost of the UF Data Center, electrical power, and staff salaries, which are subsidized and not charged to UF researchers.
AWS offers various purchasing models for EC2 instance types. Below are 4 examples as of September 25, 2017 (note that prices are subject to change):
- On-Demand: $0.10/hour
- Reserved Instance (RI), Standard 1-Year Term, All Upfront: $0.059/hour
(this is the effective hourly rate with upfront cost of $515)
- Reserved Instance (RI), Standard 3-Year Term, All Upfront: $0.039/hour
(this is the effective hourly rate with upfront cost of $1,013)
- Spot: $0.0232/hour
(this rate has not reached $0.025/hour in 4 different AZs of NoVa region in over 2 months)
Since the “Spot” price is not a guaranteed rate, the “Reserved Instance” (RI) price will be used as the c4.large “Price Per Unit” in the cost comparison below.
Compare the costs of purchasing equivalent HiPerGator and EC2 compute resources for 1-year and 3-year terms. Note that, for both 1-year and 3-year terms, the total cost of all resources purchased from UFIT RC is less than the price of a single c4.large instance from AWS. The table below compares the AWS prices with the fully loaded commercial prices for HiPerGator. UF faculty pay significantly subsidized prices for the use of HiPerGator.
|UF Research Computing vs Amazon Web Services||Compute Unit||Commercial Price Per Unit|
1 compute core
1 compute core
There are many advantages to fully mature cloud services (such as those provided by AWS) that are not provided by HiPerGator. This is part of the reason the AWS prices are higher. These advantages include:
- The latest cores, meaning faster processing
- Hardware options unavailable from UFIT RC (GPUs, FPGAs, high RAM-core ratios, etc.)
- Software options unavailable from UFIT RC (e.g. HiPerGator runs Red Hat Enterprise Linux 7. A researcher may need a different operating system, desire root access, need key software already configured in marketplace, etc.)
- Bursting for larger jobs (Need 5,000 nodes for the weekend? No problem!)
- No waiting in the queue for those with funding
- Off-site backups/redundancy/high-availability
- Public data sets (geospatial/environmental, genomics/life sciences, datasets for machine learning, regulatory and statistical data, etc.)
Thanks to Bill Richmond, Senior Solutions Architect at AWS, for helping construct this cost comparison.
Research Computing products and services are designed to support change and innovation. If you cannot see what you are looking for, please contact us and let us know how we can help meet the requirements of your group’s research.