CLOUD VS. ON-PREMISE - Total Cost of Ownership Analysis

Deep learning applications require powerful multi-GPU systems for development and operation, which can be very expensive to rent in the cloud for long-term operations. The question arises: Which infrastructure offers the best compromise between time-to-solution, cost-to-solution and availability of resources?

To illustrate the benefits of an on-premise solution, let‘s compare the acquisition and operating costs of a 4-GPU system with a comparable cloud-based GPU server, the AWS p3.8xlarge instance from Amazon.

Comparing The Costs: Total Cost Of Ownership (TCO)

To be able to present a direct price comparison, we are assuming a 1-year contract for the use of an AWS EC2 P3 instance „p3.8xlarge“ in the region „EU (Frankfurt)“. The lowest rate AWS offers for guaranteed uptime is categorized as Reserved Instance, All Upfront, to be paid 100% in advance. To simplify the calculation, we leave out the additional costs that would be incurred for storage reservation and data throughput at AWS.

Cost for 1 Year

AWS p3.8xlarge
Reserved Instance, All Upfront

76.798,04 €
AIME R400
4x RTX 2080TI, incl. electricity
12.051,80 €
You save 84,3%

AIME R400
4x Tesla V100, incl. electricity
41.226,40 €
You save 46,3%

Profitable From The Second Month On

In our teaser example set-up with the AIME R400 we therefore save 84.3% compared to the cloud provider, without sacrificing performance. The lifetime Return of Investment (RoI) of AIME R400 compared to the monthly cost of an AWS p3.8xlarge All Upfront instance with a one-year contract is shown in the following illustration. You can see that the AIME R400 with RTX 2080TI GPUs is profitable from the second month on and with the ultra high performance of the Titan V100 GPUs from the fifth month on.

Lifetime RoI of AIME R400 compared to the monthly cost of an AWS p3.8xlarge All Upfront instance with a one-year contract.
Lifetime RoI of AIME R400 compared to the monthly cost of an AWS p3.8xlarge All Upfront instance with a one-year contract.

Saving Money While Boosting Performance

The total cost of ownership (TCO) of an AIME R400 server includes the initial cost of the system, as well as energy costs. The electricity consumption of the AIME R400 server is calculated at € 0.28 per kWh at a green electricity supplier, assuming a consumption of 8760 kWh per year for 24/7 continuous operation under full load.

The initial cost of an AIME R400 server depends mainly on the installed GPUs. AIME offers two configurations: 4x RTX 2080TI or 4x Tesla V100 GPUs.

As you can see in the following figure, if you invest in an on-premise machine, you can save up to € 108,510 on a one-year project lifecycle without sacrificing performance.

Comparison of cost and performance for 1 year of a 4-GPU server in continuous use, based on the cheapest rate with the highest possible performance of a comparable AWS Instance (AWS p3.8x Large Reserved Instance, All Upfront)
Comparison of cost and performance for 1 year of a 4-GPU server in continuous use, based on the cheapest rate with the highest possible performance of a comparable AWS Instance (AWS p3.8x Large Reserved Instance, All Upfront)

More Benefits Besides The Costs

Compared to AWS instances offering lower-priced quotes than the p3.8xlarge used here, an on-premise solution provides a more powerful system for a fraction of the cost of a cloud solution.

Running your own deep learning machine brings even more benefits: you get faster, more direct access to the data store, you do not compromise on data quality, and you have enough storage at higher transfer rates. You also protects corporate data that does not need to be uploaded to the cloud.

Comparison of the monthly costs of a 4-GPU server in continuous use, based on the AWS On Demand tariff for a p3.8xlarge instance for 1 year term.

Comparison of the monthly costs of a 4-GPU server in continuous use, based on the AWS On Demand tariff for a p3.8xlarge instance for 1 year term.

Well Balanced & Preconfigured

All AIME components have been selected for their energy efficiency, durability and high performance. They are perfectly balanced, so there are no performance bottlenecks. AIME optimizes their hardware in terms of cost per performance, without compromising endurance and reliability.

AIMEs hardware was designed for their own deep learning application needs and evolved in years of experience in deep learning frameworks and customized PC hardware building. It comes preconfigured with all common AI frameworks, so you can start right out-of-the-box with your calculations.


We look forward to providing you an individual offer with your desired configuration.