TQEX @ Huawei
Tensor-centric SQL, graph-query, and graph-processing systems across heterogeneous accelerators.
TQEX anchors a broader tensor-centric execution line that maps irregular SQL and graph workloads onto portable tensor runtimes and heterogeneous accelerators. The individual systems are related research threads, not components of a single codebase.
Problem
Specialized accelerators offer substantial compute capacity, but hardware-specific database and graph systems are expensive to port and maintain. Tensor computation runtimes provide a common programming layer across accelerator backends, yet SQL and graph workloads are irregular: variable-length values, joins, sparse topology, dynamic frontiers, and graph indexes do not map directly to uniform tensor operations.
Core idea
Use tensor computation runtimes as the portability layer, then bridge the workload–tensor gap with data-system techniques: compact representations, workload-specific tensor operators, batched execution, compression, and out-of-memory strategies.
This Huawei-affiliated research line spans three workloads:
- TQEx(SQL) bridges relational storage and operators with tensor execution, including variable-length data, joins, aggregates, and multi-XPU processing.
- TenGraph maps interactive graph queries to compact tensor representations and batched tensor computation.
- TGraph provides a tensor-centric programming and execution framework for graph algorithms across tensor runtimes and accelerator backends.
My role
Research and system architecture across this Huawei-affiliated tensor-centric data-systems line, spanning query/runtime design and execution over heterogeneous accelerators.
Evidence
System boundary
TQEx(SQL) is the named tensor-based SQL engine. TenGraph and TGraph are adjacent systems that share the tensor-runtime thesis; they are not presented here as modules of TQEx(SQL).
Research Threads
TQEx(SQL) · SIGMOD 2026
Bridges irregular relational workloads and uniform tensor operations through variable-length storage, tensorized joins and aggregates, and multi-XPU execution.
System connection: Anchors the line with a portable SQL engine built over tensor computation runtimes.
TenGraph · VLDB 2024
Represents graph topology with compact tensors and batches graph-query operations through PyTorch.
System connection: Extends tensor-runtime execution to interactive graph queries.
TGraph · SIGMOD 2025
Provides tensor-based graph operators, compression, and out-of-memory execution across multiple tensor runtimes and accelerator backends.
System connection: Generalizes the approach from graph queries to graph algorithms.