The cost of the fringe
To compare the cost of deploying computing resources at different scales, we convert capital expenditures to depreciation by dividing each asset class by the number of years it will be depreciated, and then add the resulting depreciation to the annual operating expenses, so you can get a snapshot of the annual cost structure. For example, power and cooling systems are depreciated over 14 years, while COTS servers are typically depreciated over 3 years.
Capex includes:
- Server capex is mainly the cost of COTS servers and virtualization software.
- Other capex consists of the cost of components such as power distribution and cooling systems.
Opex includes:
- The electrical power needed to run and cool the servers.
- Other opex, mainly the costs of operations and maintenance (O&M).
As an example, we estimate the cost of compute resources for a CSP in Sweden. Initially, the rollout of edge compute is expected to take place at aggregation sites with installed power up to 10 kW, with an average of 8 server units, each with 4 cores. With approximately 8,000 access sites and 1 aggregation site per 10 access sites, there is a virtual processor capacity (vCPUs) of 25,600 (800 sites x 8 servers per site x 4 cores per server) for enterprise applications on edge sites owned by CSP. Capex relies on the required capacity plus redundancy in the edge hardware components to meet the reliability requirements for edge services or applications. The geographic distribution can also be leveraged to improve system availability by avoiding a single point of failure. We categorize the capex into server capex and other capex because of the faster cycle of server performance improvement compared to others. Servers typically depreciate over 3 years, while investments in power and cooling systems are depreciated over 14 years. Upgrading aggregation sites with edge computing capability, with an average of 8 units of servers, can use up to 1.6 MW (800 sites x 8 servers per site x 250 W per server) to run the servers. With an assumed energy efficiency factor of 2, an average of 3.2 MW is needed to power all aggregation sites. The cost of compute resources on each aggregation site is estimated to be approximately $20,000. Therefore, the USD per critical watt for an edge site is USD 20,000/(8 servers x 250W/server) = USD 10/W. These costs are comparable to USD per critical watt for building a large-scale data center.
Opex is the sum of electricity costs and O&M. For the current study, we assume that this ranges from $0.10-0.15 per kWh. For O&M, it estimates the cost of full-time employees required to manage and maintain the distributed edge servers.
We constructed four different scenarios to estimate and compare the cost of compute resources, based on USD per vCPU hour.
- Scenario 1 is a baseline assumed cost scenario for a small or medium-sized business that handles its computing needs with its own IT infrastructure.
- Scenario 2 is an estimate of the cost for a large-scale data center to provide the same capacity as in the first case.
- Scenario 3 is built around the provisioning of the capacity that is used in the first two cases by deploying edge computing on the CSP network.
- Scenario 4 is an extension of the third case, adding the cost of implementing a series of measures to reduce power consumption. These include the use of renewable energy, dynamic use of battery/power storage during peak hours and advanced cooling technologies, including a heat exchanger for the server cabinets.
Server capex is the most important parameter for all scenarios except the base case where O&M (other opex) dominates due to lack of scale. Electricity costs are the second largest factor in USD/CPU hours for Scenario 3. This leads to the importance of additional energy efficiency elements in Scenario 4. With an estimate of spend in use cases suitable for edge deployment , the cost of edge computing resources can be only 10 percent more than a large-scale centralized one. Capacity utilization is the most important parameter for increasing the cost efficiency of the edge resources.