Supercomputers are very complex assemblies of hardware and software components. One does not simply buy a specific number of standard cabinets from a vendor and be done with the supercomputer procurement. All systems are tailored to specific requirements and are thus unique.
Most of the users who have access to a supercomputer are not familiar with the procedure of acquiring one. A purchase project takes months and requires a team effort of many people from various fields. They need to have a budget, understand the relevant technologies and what the users need, and know the most important applications as well as their characteristics and usage patterns. The team must also have a strategic vision, benchmark new technologies, discuss with the vendors, understand and follow the procurement law, etc.
A computing center needs a supercomputer based on its specific needs, as the users and applications they run are different. Thus, a vendor should provide hardware and software solutions that match these needs as closely as possible. For example, one organization may need only CPUs, while another wants to complement the CPUs with GPUs. Additionally, application performance typically depends on the interconnect speed, so some would need to invest more money to improve the communication performance between the nodes. However, also the processor performance, the amount and speed of memory, and the storage characteristics can be adjusted. Thus, every solution is truly unique and according to what the customer wants.
To assess the performance of different solutions quantitatively, vendors are typically required to provide performance commitments representing and reflecting the expected usage of the upcoming system for a collection of applications. An interesting characteristic of supercomputer procurements is that computing centers typically buy technology that will only be available later, sometimes even after a few years. This is due to the rapid development of the technology and the time it takes to run procurement and build, install, and commission a large-scale system. Therefore, it does not make sense to buy a system that is obsolete on the day it is opened for users.
Hence, the vendors responsible for procurement need to be able to predict the performance of a system that does not yet exist. In the best case, the vendor may have some early engineering samples of the processors and other components, but the benchmark applications are typically analyzed very carefully, and then advanced models are used to foresee and provide the needed performance commitments.
For the vendors to be able to design and propose solutions, a request for proposals (RFP) is prepared with all relevant details and requirements. It is also possible to ask for more information through a request for information (RFI), after which qualified vendors are invited to submit proposals. In the proposals, the vendors present their solutions in great detail, including performance projections for the benchmarks. The proposals are then evaluated carefully based on how they meet the requirements of the RFP. Typically the most important decision criterion is some price-performance metric based on the system’s benchmark commitments, price, and operating costs.
Source: Elements of Supercomputing / NCC Finland -> Course: Running and using a supercomputer -> Section: How to buy a supercomputer