June 21, 2023
By Bob O'Donnell
To the surprise of absolutely no one, one of the biggest announcements from HPE’s Discover event had to do with generative AI. What may catch people off guard, however, is the manner with which the company is entering the GenAI market.
What HPE announced is a new as-a-service offering it’s calling HPE GreenLake for Large Language Models (LLMs). Expected to be available later this year in the US and early next year in Europe, the latest HPE GreenLake offering is designed to let customers tap into a multi-language LLM model called Luminous from a German-based company called Aleph Alpha. Like other competitive services, HPE focused on how the goal for this new service is to let companies use their own data to train the underlying multipurpose generative AI model and customize it for their own needs.
What’s particularly interesting about the news is how the company is positioning the service and the technologies that are enabling it. HPE is calling it a public cloud that’s specifically intended to let companies tap into it as they have the need and with the specific resources they require. In other words, a classic public cloud-type service—which isn’t something you’d necessarily expect HPE to do in a brand-new field like generative AI.
What’s even more surprising, however, is that it’s being powered by HPE’s supercomputing assets. Essentially, HPE built a supercomputing-as-a-service solution—an interesting development in its own right—and then populated it with a specific application. In fact, HPE made the point to say that Generative AI is just the first of many different applications that will be made available through the new architecture that they created. Future applications will cover climate modeling, healthcare and life sciences, financial services, manufacturing, and transportation.
Given HPE’s heritage in supercomputing—remember, it bought Cray Supercomputers in 2019—the move makes sense, especially with the company’s even longer efforts to create HPE GreenLake as-a-service compute offerings (those started in 2018). Nevertheless, it’s still interesting to think about the implications of the move and the potential benefits it offers compared to the many other competitive generative AI service offerings.
At its Discover event in Las Vegas, HPE laid out a number of reasons why it chose to enter the Generative AI market in this way. First, from a hardware perspective, the company believes that the unique architecture of the Cray supercomputers along with HPE AI software stack that the service will run on will give it an edge when it comes to model performance and training reliability. In particular, the company discussed how the faster interconnect speeds found in supercomputing architectures will allow the GPUs to be fed data at a rate that’s up to 16x faster than traditional server architectures. This, in turn, should keep the GPUs functioning at a more consistent rate and that should translate into better performance and price/performance ratios.
In addition, the company said it found that for companies building large-scale GenAI models, success rates of running these models could be as low as 15% on traditional server architectures versus closer to 100% on supercomputers. It will take real-world benchmarks to see if these theoretical benefits convert into real-world ones, but it will certainly be interesting to watch.
Another interesting approach the company is taking is to put each AI workload into its own computing instance. Unlike most traditional public cloud workloads, where multiple applications run on a single node, HPE says that each workload that uses the service will have its own dedicated compute infrastructure. In part, this is apparently due to the nature of supercomputing architectures, but it actually offers a privacy benefit that some customers will inevitably find reassuring.
HPE’s choice of using Aleph Alpha’s Luminous model as part of its service is an interesting one given that the German-based company has very little recognition—in the US market in particular. Some companies may have concerns regarding its small size and newness to the market. However, HPE does have a large presence in Europe and the Luminous model natively supports 5 languages out of the gate, making it well suited for European customers. In addition, the Aleph Alpha CEO said the company has a new way of addressing explainability of the model’s results. It’s going to take more real-world testing to see how well this claim bears out, but once again, it offers an intriguing point to consider.
From a software perspective, HPE is also planning to leverage some of its supercomputing-specific software to help with the process. Specifically, the company’s Machine Learning Development Environment is designed to ease the process of creating models, and the HPE Machine Learning Data Management Software is intended to ease the process of prepping the data needed to train AI models. These tools work in conjunction with Luminous to give enterprise developers the capabilities they need to customize the general-purpose model for their specific needs.
The last interesting technological piece of the puzzle is the fact that HPE has chosen to run this service in co-location facilities that offer nearly 100% renewable energy. Given how much energy is often needed for the GPU-heavy computing infrastructure to run generative AI models, this is a non-trivial and valuable initiative. The first location is QScale, based in Quebec, Canada, with others expected to follow suit.
The market for generative AI is getting noisier and noisier each day, and it’s becoming harder for companies involved in this market to stand out from this noise. Given these challenges, HPE is going to have to push its offering and the messages around it in a big way to avoid getting lost. The supercomputing angle is certainly an interesting and unique one, but its ultimate impact is going to be very dependent on what the real-world price/performance benefits turn out to be. If those capabilities have the impact that they’re touted to and if the company can successfully scale the offering should the demand arise, it will certainly be an exciting opportunity for HPE moving forward.
Here’s a link to the original article: https://www.linkedin.com/pulse/hpe-melds-supercomputing-generative-ai-bob-o-donnell
Bob O’Donnell is the president and chief analyst of TECHnalysis Research, LLC a market research firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on LinkedIn at Bob O’Donnell or on Twitter @bobodtech. |
