Artificial intelligence (AI) has emerged as a driving force behind tech innovation, with advancements in large language models (LLMs), computer vision, and predictive analytics transforming almost every type of industry.
However, the rapid adoption of AI brings significant environmental and financial challenges.
Training a single LLM with 174 billion parameters – such as with OpenAI’s GPT-3 – generates 552 metric tons of CO2 emissions.
That’s the same as the yearly emissions of 60 average U.S. homes or 125 roundtrip flights from Beijing to New York.
As AI workloads continue to grow, their associated costs and carbon footprint demand sustainable solutions.
Refurbished servers have emerged as a viable offering, providing significant cost savings, scalability, and reduced environmental impact.
But can they meet the technical demands of AI workloads? This article explores the potential of refurbished servers in addressing the infrastructure needs of the AI revolution.
AI Workloads: Infrastructure Demands
AI tasks depend heavily on compute and storage resources.
Whether training deep learning models, processing vast datasets, or deploying models for inference, servers provide the backbone of AI infrastructure.
- Training: Training large AI models requires feeding massive datasets into neural networks. Distributed training systems use clusters of servers equipped with GPUs or TPUs to accelerate this process. For example, training GPT-3 cost OpenAI an estimated $4.6 million in compute expenses alone.
- Data Storage and Processing: AI models rely on scalable storage to handle diverse datasets, including images, text, and videos. Distributed file systems or object storage (e.g., Ceph) are often implemented to meet these needs.
- Inference: AI inference – running trained models to generate predictions – is resource-intensive, especially at scale. Systems must handle millions of requests per second, requiring efficient hardware configurations.
With these demands, organisations must balance performance, cost, and sustainability when choosing server infrastructure.
Refurbished Servers: The Sustainable Alternative
Refurbished servers address some of the key challenges of AI infrastructure: high cost, carbon emissions, and scalability.
By reusing and upgrading existing hardware, businesses can meet AI demands while minimising environmental impact.
Here’s how:
1. Provide Cost Savings
Refurbished servers offer substantial cost reductions, often priced at 50%-80% less than new hardware.
For example, the HPE DL580 Gen10 Rack Server supports up to 4x ‘Double Height’ (2x PCI Slots used) GPUs of the more powerful variants like the NVIDIA A40’s and the NVIDIA A100’s, or up to 8x NVIDIA L4 variants for entry level AI tasks.
2. Sustainability
The ICT industry is responsible for 2% of global carbon emissions, equivalent to the entire airline industry.
Refurbished servers mitigate this by extending the lifecycle of existing hardware, reducing e-waste and cutting the embedded carbon footprint from manufacturing by up to 75%.
Microsoft, for instance, has implemented Circular Centres to reuse up to 90% of server components. This reduces hardware procurement costs by 27% and significantly limits material waste.
3. Scalability and Customization
Refurbished servers are highly customisable, allowing businesses to tailor them for their specific AI workloads:
- Upgraded NVMe drives provide faster data pipelines.
- High-performance GPUs (e.g., NVIDIA A100 or H100) help accelerate deep learning tasks.
- Expanded memory configurations can support large-scale datasets.
The Carbon Footprint of AI and Refurbished Hardware
Training an LLM like GPT-3 can produce 30,000 kilograms of CO2 emissions, underscoring the need for sustainable solutions.
Refurbished servers contribute to greener AI infrastructure by reducing the manufacturing demand for new servers.
Additionally, optimizing refurbished systems with energy-efficient components, such as Intel Xeon Scalable processors or modern SSDs, can cut power consumption by 30%-50%.
Microsoft’s Greensku framework exemplifies this approach, leveraging AI to optimize server reuse and reduce the environmental costs of data centres.
Similarly, Dell Technologies has validated that refurbished HPE ProLiant Rack Servers equipped with 4th Gen Intel Xeon CPUs deliver 53% greater energy efficiency per workload compared to older models.
Refurbished Servers in Action: Meeting AI Demands
Refurbished servers have proven their ability to support the demanding requirements of AI tasks.
Below are some practical examples:
1. High-Performance Storage
The HPE DL580 Gen10 Rack Server supports 16 drive bays, offering scalable storage for massive AI datasets. High-speed read/write capabilities make it ideal for data-intensive AI pipelines.
2. Advanced GPUs for Training
Servers like the HPE DL580 Gen10 support up to 6 GPUs, enabling parallel processing for deep learning tasks. NVIDIA GPUs drastically reduce training times by accelerating matrix computations.
3. Energy Efficiency
The HPE DL380 Gen10 can support up to 2x ‘Double Height’ (2x PCI Slots used) GPU’s of the more powerful variants like the NVIDIA A40’s and the NVIDIA A100’s or up to 4x NVIDIA L4 variants for entry level AI tasks.
4. Enhanced Security
- The Dell PowerEdge R640 offers advanced security features like Secure Boot and Cryptographically Signed Firmware, ensuring AI workloads involving sensitive data are protected from unauthorized access.
Cost and Lifecycle of AI Infrastructure
- New Servers: Running a server with 8x AI Optimised GPU’s – like the A100 – would cost roughly £5,800 for a year.
- Refurbished Servers: By contrast, refurbished systems provide comparable performance for 50%-80% less, with lifecycle extensions of 3-7 years through proper maintenance and component upgrades.
This cost efficiency allows businesses of all sizes to scale AI operations without breaking the budget.
The Future of Sustainable AI Infrastructure
Refurbished servers align with the growing demand for sustainable, high-performance computing. Industry leaders like Microsoft, Lenovo, and Dell are already incorporating refurbished hardware into their operations:
- Microsoft: AI-driven diagnostics optimize the reuse of GPUs, CPUs, and storage drives, reducing waste and improving Total Cost of Ownership (TCO).
- Lenovo: Rigorous testing ensures refurbished servers meet enterprise-grade standards, extending hardware lifespans while maintaining performance.
Refurbished servers are uniquely positioned to power the AI revolution by offering the perfect blend of performance, cost-efficiency, and sustainability.
They provide the necessary computational power to handle demanding AI workloads, reduce the environmental impact of infrastructure, and extend the lifecycle of valuable hardware.
As organizations prioritize scalable, green solutions, refurbished servers will play an increasingly critical role in shaping the future of AI.
For those ready to explore this option, Intelligent Servers offers tailored refurbished hardware with significant cost savings and reliable performance.
By embracing refurbished servers, businesses can innovate responsibly, powering the AI revolution without compromising their environmental or financial goals.
How Intelligent Servers can help
If refurbished servers sound like the way to go for your business, then you can purchase them at Intelligent Servers.
We offer refurbished products priced at up to 80% less than the price of new products. We also provide free worldwide delivery, and all our refurbished products come with a three-year warranty as standard – including free remote technical support for the duration of your warranty.
Have any questions about our products? Get in touch using the live chat on our website or call 01423 223430 to speak to our UK support team.
