DeML OS - 2026-04-16

DeML OS Daily DeML OS 最新前沿分析 DeML OS デイリー

Explore Frontier

04.16

2026

Thu

📄

Paper

Tessera: Unlocking Heterogeneous GPUs through Kernel-Granularity Disaggregation https://arxiv.org/abs/2604.10180

Tiancheng Hu GPU Disaggregation

Hu et al. proposed Tessera, the first kernel-level GPU disaggregation system in their paper. It enhances throughput and cost efficiency for large model inference on heterogeneous GPU clusters via offline analysis and online adaptive scheduling.

Notes

Prior GPU disaggregation is coarse-grained and model-coupled, limiting performance.
Tessera uses kernel-level disaggregation, better matching computation to hardware.
Key: PTX-based inter-kernel dependency extraction, pipelined execution.
Combines offline analysis and online workload-aware scheduling.
On 5 GPUs and 4 models: up to 2.3x throughput, 1.6x cost gains.
Heterogeneous GPU pairs can outperform homogeneous high-end pairs.

Collected by @icerdesign

DeML OS Q & A 问答

Deep Dive 💬

04.16

2026

Thu

😇

What is the key innovation of Tessera compared to prior approaches?

Tessera's key innovation is introducing kernel-granularity GPU disaggregation, matching computation tasks with hardware capabilities at a finer granularity than the traditional coarse-grained model-level approach.

😎

😊

How does Tessera ensure execution correctness after kernel disaggregation?

Tessera extracts precise inter-kernel data dependencies from the program's PTX intermediate representation, builds a dependency graph, and uses it for scheduling to ensure correct computation order and data consistency across different GPUs.

😎

😊

What is the role of Tessera's pipelined execution model?

Tessera's pipelined execution model aims to overlap data transfer (communication) between different GPUs with kernel computation on the GPUs, thereby hiding communication latency, maximizing hardware utilization, and improving overall system throughput.

😎

Prompted by @icerdesign