Local baselines
Our benchmarks run on a variety of hardware configurations, including commonly available desktop computers. The latter are used as baseline systems for our simulations, providing a reference for users interested in assessing how cloud-based computing solutions compare to their local hardware resources.
We currently benchmark two different baseline configurations: Intel i5-12400F,
and Ryzen 7-7700X, as they are representative of the most common configurations
found in a typical local setup.
Intel(R) Core(TM) 12th Gen i5-12400F with the following specslink:
| Intel i5-12400F | |
|---|---|
| Total Cores | 6 |
| Num. Performance cores | 6 |
| Num. Efficiency cores | 0 |
| Total Threads | 12 |
| CPU max | 4.40 GHz |
| CPU min | 0.80 GHz |
| RAM Memory | 64 GB |
| Price per hour | $ 0.083 |
AMD Ryzen 7 7700X 8-Core with the following specslink:
| Ryzen 7-7700X | |
|---|---|
| Total Cores | 8 |
| Num. Performance cores | n.a. |
| Num. Efficiency cores | n.a. |
| Total Threads | 16 |
| CPU max | 5.40 GHz |
| CPU min | 0.40 GHz |
| RAM Memory | 32 GB |
| Price per hour | $ 0.085 |
Why did we choose these local machine configurations as performance baselines?
The idea behind choosing these baseline machine configurations is to emulate hardware setups that are common among students, small businesses and individual users in general. Current baseline configurations are CPU-only, i.e. they do not include GPUs.
We choose such minimal hardware configurations for two reasons. The first is that, for now, the Inductiva API only offers access to CPU-only cloud VMs, such as the ones listed in our documentation. Second, by choosing CPU-only local configurations we are ensuring that we are choosing minimal-cost hardware options, since one can easily find relatively inexpensive CPU-only computers from most sellers (typically good GPUs are the most significant part of the cost of a machine). Therefore, these local hardware configurations are likely to represent the cheapest possible way for an individual user to execute the chosen benchmark use cases. Therefore, we use these options to set the cost floor for our benchmarks.
How did we estimate the price per hour of running simulations on these local hardware options?
For estimating the cost per hour of running simulation on these baseline machines, we considered three components:
- the cost of purchasing the hardware, and how that cost is diluted over time;
- the cost of the energy needed to keep that hardware running;
- an aggregation of all other costs related to operation and maintenance.
For estimating the first component, we first consulted several websites to see how much machines with similar CPU + RAM characteristics would cost, and we computed the average of the prices we found. Then, we assumed that such a computer would have a life of 3 years, after which we consider that it would become an unusable option for running simulations (e.g. imagine that in 3 years users are expecting to run simulations that are 5x bigger at a 5x faster speed than today, so this hardware would be incapable to match those expectations and would be sent to recycling). Then, to compute the cost per hour of this component, we simply divide the average purchase price we found by the number of hours in 3 years.
Here is an example. Suppose that the average price for a certain baseline machine is €1000. Then the cost per hour of this component is simply:
€1000 / (3 * 365 * 24) = €0.038 per hour = 3.8 cents per hour
For the second component of the cost, the energy cost, we assume that computers with these baseline characteristics will consume an average of 250W of power when running at full speed (we are assuming a power factor of 1). Then, we consulted electricity prices. We understand that this component is location-dependent and quite volatile. Still, we took the price of electricity available in Portugal in April 2024 as a reference point. The value we are considering is € 0.16 / kWh. This means that the electricity cost of running these baseline hardware options is:
0.250 * 0.16 = €0.04/hour, i.e. 4 cents/hour.
The third component is more ambiguous and harder to estimate but is supposed to account for several related costs such as the cost of physical space, the cost of time spent on purchasing and maintaining the hardware, and everything else that can be seen as an accessory cost of having the hardware on-premises. Because this is hard to estimate, we will compute this cost as a 50% overhead over the electricity cost.
Therefore, the total cost per computation hour for our example machine, whose purchase price was 1000 euros, would be:
3.8 + 4.0 + 50% of 4.0 = 9.8 cents / hour
Because we present cloud prices in US dollars, we assumed a conversion rate of 1:1 between euros and dollars.