Gen AI without the risks

ChatGPT, Stable Diffusion, and DreamStudio–Generative AI are grabbing all the headlines, and rightly so. The results are impressive and improving at a geometric rate. Intelligent assistants are already changing how we search, analyze information, and do everything from creating code to securing networks and writing articles.

Gen AI will become a fundamental part of how enterprises manage and deliver IT services and how business users get their work done. The possibilities are endless, but so are the pitfalls. Developing and deploying successful AI can be an expensive process with a high risk of failure. On top of that, Gen AI, and the large language models (LLMs) that power it, are super-computing workloads that devour electricity.Estimates vary, but Dr. Sajjad Moazeni of the University of Washington calculates that training an LLM with 175 billion+ parameters takes a year’s worth of energy for 1,000 US households. Answering 100 million+ generative AI questions a day can burn 1 Gigawatt-hour of electricity, which is roughly the daily energy use of 33,000 US households.¹

It’s hard to imagine how even hyperscalers can afford that much electricity. For the average enterprise, it’s prohibitively expensive. How can CIOs deliver accurate, trustworthy AI without the energy costs and carbon footprint of a small city?

Six tips for deploying Gen AI with less risk and cost-effectively

The ability to retrain generative AI for specific tasks is key to making it practical for business applications. Retraining creates expert models that are more accurate, smaller, and more efficient to run. So, does every enterprise need to build a dedicated AI development team and a supercomputer to train their own AI models? Not at all.

Here are six tips for developing and deploying AI without huge investments in expert staff or exotic hardware.

1. Don’t reinvent the wheel—start with a foundation model

A business could invest in developing its own models for its unique applications. However, the investment in supercomputing infrastructure, HPC expertise, and data scientists is beyond all but the largest hyperscalers, enterprises, and government agencies. 

Instead, start with a foundation model that has an active developer ecosystem and a healthy application portfolio. You could use a proprietary foundation model like OpenAI’s ChatGPT or an open-source model like Meta’s Llama 2. Communities like Hugging Face offer a huge range of open-source models and applications.

2. Match the model to the application

Models can be general-purpose and compute-intensive like GPT or narrowly focused on a specific topic like Med-BERT (an open-source LLM for medical literature). Selecting the right model at the beginning of a project can save months of training and shorten the time to a workable prototype.

But do be careful. Any model can manifest biases in its training data and generative AI models can fabricate answers, hallucinate, and flat-out lie. For maximum trustworthiness, look for models trained on transparent, clean data with clear governance and explainable decision making. 

3. Retrain to create smaller models with higher accuracy

Foundation models can be retrained on specific datasets, which has several benefits. As the model becomes more accurate on a narrower field, it sheds parameters it doesn’t need for the application. For example, retraining an LLM on financial information would trade a general ability like songwriting for the ability to help a customer with a mortgage application. 

The new banking assistant would have a smaller model that could run on general-purpose (existing) hardware and still deliver excellent, highly accurate services.

4. Use the infrastructure you already have

Standing up a supercomputer with 10,000 GPUs is beyond the reach of most enterprises. Fortunately, you don’t need massive GPU arrays for the bulk of practical AI training, retraining, and inference.

• Training up to 10 billion—modern CPUs with built-in AI acceleration can handle training loads in this range at competitive price/performance points. Train overnight when data center demand is low for better performance and lower costs.
• Retraining up to 10 billion—modern CPUs can retrain these models in minutes, with no GPU required.
• Inferencing from millions to <20 billion—smaller models can run on stand-alone edge devices with integrated CPUs. CPUs can provide fast and accurate responses for <20 billion-parameter models like Llama 2 that are competitive with GPUs.

5. Run hardware-aware inference

Inference applications can be optimized and tuned for better performance on specific hardware types and features. As with model training, optimization entails balancing accuracy with model size and processing efficiency to meet the needs of a specific application. 

For example, converting a 32-bit floating point model to the nearest 8-bit fixed integers (INT8) can boost inference speeds 4x with minimal accuracy loss. Tools like Intel® Distribution of OpenVINO™ toolkit manage optimization and create hardware-aware inference engines that take advantage of host accelerators like integrated GPUs, Intel® Advanced Matrix Extensions (Intel® AMX), and Intel® Advanced Vector Extensions 512 (Intel® AVX-512).

6. Keep an eye on cloud spend

Providing AI services with cloud-based AI APIs and applications is a fast, reliable path that can scale on demand. Always-on AI from a service provider is great for business users and customers alike, but costs can ramp up unexpectedly. If everyone loves your AI service, everyone will use your service.

Many companies that started their AI journeys completely in the cloud are repatriating workloads that can perform well on their existing on-premises and co-located infrastructure. Cloud-native organizations with little-to-no, on-premises infrastructure are finding pay-as-you-go, infrastructure-as-a-service a viable alternative to spiking cloud costs. 

When it comes to Gen AI, you have options. The hype and black-box mystery around generative AI makes it seem like moonshot technology that only the most well-funded organizations can afford. In reality, there are hundreds of high-performance models, including LLMs for generative AI, that are accurate and performant on a standard CPU-based data center or cloud instance. The tools for experimenting, prototyping, and deploying enterprise-grade generative AI are maturing fast on the proprietary side and in open-source communities.

Smart CIOs who take advantage of all their options can field business-changing AI without the costs and risks of developing everything on their own.

About Intel

Intel® hardware and software powers AI training, inference, and applications in Dell supercomputers and data centers, through to rugged edge servers for networking and IoT to accelerate AI everywhere. Learn more. 

About Dell

Dell Technologies accelerates your AI journey from possible to proven by leveraging innovative technologies, a comprehensive suite of professional services, and an extensive network of partners. Learn more.

^[1] University of Washington, UW News, Q&A: UW researcher discusses just how much energy ChatGPT uses, July 27, 2032, Accessed November, 2023