How TACC Runs On-Premises AI Research on SambaStack™
The Texas Advanced Computing Center (TACC), an NSF-funded research facility, runs high-performance AI workloads and coding agents on SambaNova’s air-cooled, energy efficient SambaStack™ platform.
Company Background
The Texas Advanced Computing Center (TACC) at the University of Texas at Austin is a National Science Foundation (NSF) funded research center supporting the full spectrum of scientific inquiry — from engineering and the natural sciences to the humanities, social sciences, and cultural studies. Through large-scale modeling, simulation, and real-world data analysis, TACC enables breakthrough discoveries that have contributed to multiple Nobel Prize-winning research programs.
What TACC Does
TACC operates some of the world's largest clusters for open compute, including the Frontera supercomputer, one of the most powerful systems in the world at its launch, and the next-generation Horizon supercomputer, which delivers 10x Frontera's performance.
AI is central to how TACC empowers its user community. Internally, AI coding agents help researchers build and understand software modules, answer complex questions, and accelerate new discoveries. Across thousands of users spanning a wide range of disciplines, AI enables researchers to analyze vast quantities of complex data and surface insights that would otherwise be out of reach.
Challenge: High-Performance AI Within Strict Power and Space Constraints
TACC supports thousands of users with an exceptionally diverse range of workloads, but not every application demands supercomputer-class resources. Many workloads, such as coding, run on systems that must fit within strict power, cooling, and floor space constraints while still delivering exceptional performance.
TACC's requirements were specific and demanding:
- Fast, low-latency coding models. Researchers use coding agents to create and manage projects and to analyze previous work to plan next steps. They need access to the latest and best coding models, such as MiniMax M2.7.
- Broad open-source model support. As a research institution, almost everything TACC runs is open source so users can easily collaborate. Any system in their data center must run a broad range of open-source models.
- Energy efficiency as a core priority. While TACC's largest supercomputers consume up to 15 MW, the center also needs to run lighter workloads cost-effectively, making power consumption a decisive factor when evaluating systems.
- A fit for existing data center infrastructure. New systems needed to slot into current power, cooling, and floor-space limits without costly retrofits.
Solution: SambaStack™ for Efficient, Open, Low-Latency Inference
TACC uses SambaStack™ to power a broad range of AI workloads. Its OpenAI-compatible API makes integration straightforward, allowing TACC to incorporate SambaStack into existing infrastructure, including Splunk-based logging and other internal services, without disruption.
“The technology is fantastic," said Niall Gaffney, director of Data Intensive Computing, TACC. "We’re always about better solutions for solving problems.”
For TACC's engineers, that OpenAI-compatible endpoint meant no rewrite of existing pipelines because SambaStack dropped straight into their toolchain.
Most models get the right outcome the first time
Niall Gaffney, director of Data Intensive Computing at TACC, discusses advantages of SambaStack.
SambaStack integrates well with pretty much every other system out there, giving TACC the benefit of both worlds.
The smaller footprint of SambaStack - both energy-wise and physically - while maintaining high performance are attractive.
Challenge:
Hume specializes in building the most realistic voice AI models for developers and enterprises. These models are based on LLMs, so they understand both language and a person’s voice at the same time. Their mission is to bring empathy to AI and to align AI with human well-being. To that end, the speech-LLMs they develop are capable of understanding both the tone and meaning of the spoken word. Applications for this include audio chatbots, customer service, and more.
They recently launched the highest quality speech-LLMs for text-to-speech (Octave) and speech-to-speech (EVI 3). Much of the quality comes from the models’ ability to understand language and to adjust its tone of voice naturally in response to the input. This enables a more natural conversation, which can improve user perception.
Most voice systems today have separate text-to-speech, speech-to-text, transcription, and other models connected together because they were better at each individual task, but with the latest advances in speech-language models this is no longer the case. Moreover, each of these steps adds latency to the process. Conversational human latency is 200 ms and anything longer than 1 second will sound less human. Hume AI and SambaNova have worked together to develop a solution that delivers the highest performance at the lowest latency possible.
Solution:
Hume and SambaNova have worked together to deploy Hume’s speech-language models on SambaCloud, enabling the best speech-to-speech and text-to-speech models in the world to run at conversational latency without any reduction in quality. Together, Hume AI and SambaNova provide enterprises with access to text-to-speech and speech-to-speech APIs with response times on the order of 100 ms to 300 ms, marrying hyperrealistic quality with human-like conversation latency.
For many enterprises, it is critical to deploy in private environments. Hume and SambaNova are providing Hume’s text-to-speech and speech-to-speech models through private deployments to meet these needs.
Response time
Highest quality speech LLMs
“It allows us to have flexibility while still providing a low-energy, low-footprint, low-latency, highly efficient system. It really is the best of both worlds.”
— Niall Gaffney, Director of Data Intensive Computing
Texas Advanced Computing Center
Coding Agents Powered by MiniMax M2.7
Many of the researchers at TACC take advantage of coding agents so that they no longer have to write test modules or comment their code manually. SambaStack's support for leading open-source models enables it to run the full suite of coding agents TACC relies on, including MiniMax M2.7.
MiniMax M2.7 lets researchers generate code for a wide range of tasks and work with substantially larger prompts, which is an excellent fit for the complex, data-intensive work common in research environments. On SambaStack, MiniMax M2.7 delivers the fast, low-latency execution that agentic coding workflows demand.
Air-Cooled Design: Reliability Without the Leak Risk
Reliability and simplicity of deployment also matter. Other systems in TACC's data center require full liquid cooling, and have experienced coolant leaks. SambaStack's air-cooled design eliminates that risk and simplifies installation, dropping into TACC’s existing data center without specialized cooling infrastructure.
Energy Savings vs. GPU-Based Alternatives
Combined with a lower energy footprint than other systems, SambaStack lets TACC run the workloads it needs, reduce operating costs, and stay within the physical constraints of its existing data center.
Why TACC Chose SambaStack
TACC chose SambaStack because it delivered high-performance AI inference while solving the operational constraints that ruled out other systems. Five factors drove the decision:
- Air-cooled design. No liquid cooling means no coolant-leak risk and a straightforward install in TACC's existing data center.
- Energy efficiency. Lower power draw than GPU-based alternatives keeps lighter workloads cost-effective, which is critical for a data center already managing up to 15 MW of supercomputing load.
- OpenAI-compatible API. SambaStack integrated cleanly with TACC's existing infrastructure, including Splunk-based logging, with no pipeline rewrites.
- Broad open-source model support. SambaStack runs the open-source models TACC's collaborative research community depends on, including MiniMax M2.7.
- Low-latency inference for coding agents. Fast execution. With industry-leading performance on MiniMax M2.7, researchers' agentic coding workflows stay responsive.
Gaffney finished by saying, "Working with SambaNova is working with them. Working with the team is what has made the difference."
The result: TACC runs the AI workloads its community needs, reduces operating costs, and fits everything within the power, cooling, and floor-space limits of its existing facility. The best of both worlds: flexibility and efficiency in one system.
FAQs
What is SambaStack?
SambaStack is a full-stack enterprise AI inference platform that runs on-premises, in a dedicated cloud, or in a hybrid deployment. It combines SambaNova's Reconfigurable Dataflow Unit (RDU) chip, SambaRack™ hardware, orchestration software, and OpenAI-compatible APIs into a single stack. The design targets the most efficient full-stack AI inference, from chips to models, for enterprises and research centers building their own AI infrastructure.
What open-source models does SambaStack support?
SambaStack supports a broad range of leading open-source models, including the largest frontier models such as DeepSeek, Llama, Qwen, gpt-oss-120b, and MiniMax M2.7. Open-source support was essential for TACC because its research community relies on open models for easy collaboration. Any system in its data center needed to run that full range of open-source models without lock-in.
What does TACC use coding agents for?
TACC researchers use coding agents to build and manage research projects without writing test modules or commenting code manually. The agents also analyze previous projects to help researchers identify the next steps in their work. MiniMax M2.7 lets them generate code for a wide range of tasks and work with substantially larger prompts. This capability is essential for the complex, data-intensive challenges common in research.
How does SambaStack's OpenAI-compatible API work?
SambaStack exposes an OpenAI-compatible API, so teams can point existing tools and pipelines at SambaStack without rewriting their code. For TACC, this meant dropping SambaStack into its existing infrastructure, including Splunk-based logging and other internal services, with no disruption. Any harness or application already built for an OpenAI-style endpoint can connect to SambaStack directly.
.jpg?width=380&height=220&name=Blog%20-%20SN50%20Runs%20the%20Fastest%20MiniMax%20Speeds%20in%20the%20World%20(1).jpg)
