Tessera: Unlocking Heterogeneous GPUs through Kernel-Level Disaggregation

Learn how Tessera disaggregates kernels into heterogeneous GPUs to accelerate large model inference by up to 2.3x, outperforming traditional solutions.

viernes, 10 de julio de 2026 • 4 min read • Q2BSTUDIO Team

Kernel disaggregation for efficient inference

The rise of artificial intelligence models has led companies to look for increasingly powerful and flexible infrastructures. However, the variety of GPUs available on the market—from high-performance solutions to cheaper options—poses a challenge: how to get the most out of a heterogeneous cluster without losing efficiency or driving up costs. Traditionally, disaggregation strategies have operated at the application or full-layer level, assigning entire workloads to one type of GPU. But this coarse granularity wastes optimization opportunities, since within the same inference model kernels with very different resource profiles coexist. A new approach, embodied in systems like Tessera, proposes to break that barrier and bring disaggregation to the kernel level, that is, to the smallest piece of computation that runs on the GPU. This innovation promises to transform the way organizations deploy AI across heterogeneous environments.

The core idea is simple but powerful: if each kernel within an application demands a specific amount of memory, bandwidth, and compute capacity, then it makes sense to pair it with the GPU that best suits those needs. For example, a kernel with high memory demand can run on a GPU with higher VRAM, while another kernel with intensive matrix operations can benefit from a GPU with more CUDA cores. Until now, existing solutions required the entire model to fit into a single GPU, or fragmented the load at the layer or block level, introducing unwieldy synchronization complexities and dependencies. Tessera, on the other hand, extracts interkernel dependencies directly from PTX code, a low-level language from NVIDIA, and ensures the correctness of distributed execution without the need to modify the original model.

Why is this relevant to the business world? Because large model inference is one of the most costly bottlenecks in AI adoption. A company running a conversational assistant or recommendation system needs to minimize latency and maximize throughput, but also control hardware spend. With Tessera, a pair of heterogeneous GPUs—one high-end and one more modest—can outperform two homogeneous high-end GPUs, while reducing investment. This opens the door for companies of any size to access artificial intelligence capabilities that were previously only available to tech giants. In addition, the system incorporates a pipeline execution model that overlaps communication with computing, minimizing downtime and increasing overall efficiency.

Behind this technical capacity there is a broader lesson: hardware specialization is not enough if it is not accompanied by custom software that knows how to orchestrate resources intelligently. Enterprises that want to take advantage of heterogeneous GPUs—either in their on-premises data centers or through AWS and Azure cloud services—will need middleware solutions that dynamically adapt task allocation. Tessera demonstrates that this is possible through a combination of offline analysis (to understand the requirements of each kernel) and online adaptation (to react to changes in load or GPU availability). This type of development requires a thorough understanding of both the hardware architecture and the GPU compilers and runtime, and this is where tech companies like Q2BSTUDIO come in.

Q2BSTUDIO offers tailored applications that integrate high-performance solutions with cloud and artificial intelligence. Our team understands that optimization doesn't end at the model, but must go down to the last kernel. That's why, for customers working with massive inference or distributed training, we develop custom software layers that implement Tessera-like disaggregation strategies, tailored to their specific hardware. Whether it's AI agents managing the execution pipeline or enterprise AI systems deploying on hybrid clusters, our goal is to make every GPU cycle count.

Integration with cloud services is also key. Many of our projects combine the power of on-premises GPUs with the elasticity of AWS or Azure, using AWS and Azure cloud services for peak demand. In those scenarios, kernel-level disaggregation can reallocate loads between GPU instances of different types (e.g., A100 and T4) based on each kernel profile, reducing infrastructure costs. In addition, we use business intelligence tools such as Power BI to monitor the utilization of each GPU in real time and generate alerts when the allocation is not optimal. This visibility allows data teams to make informed decisions about when to scale or switch instances.

However, disaggregation at the kernel level is not a panacea. It requires advanced cybersecurity techniques, as moving kernels between GPUs involves transferring sensitive data that needs to be protected. Companies that adopt these systems must have robust cybersecurity policies in place, both at the network and shared memory level. Q2BSTUDIO includes security modules in its solutions that encrypt communication between GPUs and verify the integrity of kernels, aligning with regulations such as GDPR. Likewise, the complexity of the dependencies makes it necessary to rigorously monitor execution times; That's why we combine artificial intelligence with Power BI dashboards to detect anomalous patterns and prevent bottlenecks.

On the horizon, the evolution of generative artificial intelligence and autonomous agents will demand more and more efficiency in inference. Systems that are research today, such as Tessera, will tomorrow be standard in production infrastructures. That's why companies that invest now in custom software and optimizing their heterogeneous clusters will be better positioned to scale their AI applications without driving up costs. At Q2BSTUDIO, we accompany our customers on this path, designing architectures that combine the best available hardware with intelligent software, capable of squeezing every last resource out of each GPU.

OUR SERVICES

How we can help you

Do you have a project in mind?

Tell us your vision and we'll turn it into a software solution. Whatever the scope, we make your idea real.