Executive Summary
Aible clients consistently get AI projects up and delivering business value in under 30 days. In tests, Intel and Aible demonstrated that the most common Generative AI (GenAI) workload—GenAI with RAG—performed well on serverless CPUs at projected costs up to 55x lower than standard, server-based AI services.1
Challenge
According to a 2024 Gartner® press release, “On average, only 48% of AI projects make it into production, and it takes 8 months to go from AI prototype to production.”2 Gen AI may be making headlines, but businesses don’t seem to be benefiting.
Solution
The Aible AI platform combines user-friendly tools for GenAI with low-cost serverless, CPU computing infrastructure. By simplifying GenAI toolsets and lowering costs, Aible makes it possible for developers to iterate quickly, and for business users to achieve business alignment and build AI applications on their own.
Harness the power of AI and accelerate applications across your entire enterprise. Discover how AI powered by Intel Xeon can work for you.
Results
Intel tested the most common GenAI workload—GenAI with RAG—using the Aible platform on three generations of Intel® Xeon® Scalable processors. Tests show widely available 2nd Generation Intel® Xeon® Scalable instances are fully capable of running Gen AI with RAG—without GPUs or other accelerators. Additional tests demonstrated GenAI responses in seconds with up to 20 percent faster performance on 4th Generation and up to 30 percent faster performance on 5th Generation Intel® Xeon® Scalable processors.1
According to Aible’s benchmark analysis, running Aible Gen AI with RAG on serverless CPUs cuts ongoing computing costs up to 55x.3
Aible Serverless Platform Proves CPUs Solve Key Development Barriers for Gen AI
“When I looked at why AI projects fail, it’s always because of disconnect between the data scientists, experts, and business users,” says Aible founder and CEO Arijit Sengupta. “So, what we obsessively focus on is, how can I get the business user in the mix as quickly as possible?”
The high cost of AI computing creates scarcity that makes it prohibitively expensive for AI developers to iterate at scale and fail fast. Tight budgets can squeeze subject matter experts out of the process. Aible tackles computing costs with a serverless approach that delivers GPU-level performance by scaling across multiple CPUs to create a distributed, parallel-processing instance. By using serverless CPU instances, the Aible platform lowers computing costs so that data scientists, subject matter experts, and business owners can be involved throughout the AI development process.
It’s an excellent model, but can serverless CPUs meet the voracious demands of Gen AI? To find out, Intel and Aible benchmarked one of the most common GenAI workloads—GenAI with RAG—with the Aible platform running on 2nd, 4th, and 5th Generation Intel Xeon Scalable processors.
The results were very encouraging. On a single 5th Generation Intel Xeon Scalable processor, GenAI with RAG execution times were as low as 21.02 seconds.1 Testing showed a single, 2019-era 2nd Generation Intel® Xeon® Scalable processor finished in 30.2 seconds.1
Since the Aible serverless approach scales across multiple CPUs, Aible platform customers can expect highly responsive GenAI services from standard Intel® Xeon® instances.
Why Did Intel and Aible Test GenAI with RAG on Six-Year-Old CPUs?
Most serverless instances only expose “lowest common denominator” technologies, which are roughly equivalent to a ten-year-old CPU. CPUs from 2019 provide a more realistic benchmark.
Serverless Compute is Up to 55x3 Less Expensive for Gen AI
To compare real-world costs, Aible priced AI with RAG services on three computing configurations.
- Shared public cloud LLM with a leading hosted VectorDB.
- LLM and VectorDB running on dedicated servers in a private cloud.
- LLM and VectorDB on the Aible platform using cloudbased, serverless CPU instances.
Aible then projected costs for small (12,000 chats/50GB of unstructured data), medium (60,000 chats/250GB of unstructured data), and large GenAI (240,000 chats/1TB of unstructured data) AI with RAG workloads on each configuration. In each case, serverless computingm infrastructure was significantly less expensive—from 9xm up to 55x less.3
How Aible Exploits Serverless Computing Quirks to Increase AI Performance
True serverless computing provides processing power on demand. The user doesn’t have to reserve or pay for a server in preparation for a workload. This means nothing happens—and no costs are incurred—until a user asks a question.
One downfall of serverless is the time it takes to spin up. Aible solves this by taking advantage of the fact that serverless instances persist after they’ve completed a task. By carefully queueing their platform’s active workloads, Aible keeps serverless instances running non-stop—delivering near-dedicated server performance with no time lost to idling down and spinning back up.
Dynamic Model Selection Reduces Costs Even Further
Aible provides proactive control over the ongoing costs of AI services. Administrators can budget costs per chat and costs per user to limit overspending. The system can dynamically change models to match costs to budgets. For example, the system can switch individual users and entire use cases from an expensive, general-purpose LLM to a smaller, less expensive model like Mistral or Llama 8B.
Lower Cost and Friendlier Tools Create New Techniques for Getting AI into Production
The Aible platform’s ease of use and low-cost computing support faster, scrappier approaches to AI development. With Aible, clients can fail fast and get into production without breaking the bank.
- Run 10, 50, or 100 model variations simultaneously.
- Test RAG, fine-tuning, and few-shot learning, all at the same time.
- Test and refine multiple prototypes with business users.
- Bring Gen AI augmentation to data engineering, analytics, and machine learning.
Because they can work faster, Aible clients typically see business value in less than thirty days, some in as few as eight days.4
Intel and Aible – A Partnership That is Bringing AI Everywhere
Aible customers are proof that lowering computing costs and simplifying AI development fundamentally change how AI projects move through development and into production. A leading analyst firm states that it can take over 8 months to complete a generative AI project. “We have customers who have completed their projects in just two days,” says Aible CEO Arijit Sengupta.
Making AI more accessible and successful isn’t just about fast paths to business value. It’s about unlocking AI’s potential and expanding its application for all. “Our motto from the beginning was I-am-able because we believed anyone should be able to create AI that serves their needs,” says Sengupta. “We are solving that fundamental risk point of AI—that most of the projects are going to fail. If we change the way AI is done—move it away from monolithic AI to AI that empowers the individual—we will have succeeded.
About Aible
Aible provides user-friendly tools at lower costs so everyone can develop their own AI and have a say in how it affects their lives.
Learn more at aible.com.
About Intel
Intel efforts in AI include hardware and performanceimproving AI software design and optimization. Intel works with partners of every size to bring AI everywhere.
Learn about Intel AI Tools.