Lessons learned building cross-architecture analytical engines

Modern server hardware is increasingly heterogeneous, featuring a diverse mix of XPU architectures deployed across multi-vendor CPUs, GPUs, and FPGAs. Historically, database and systems developers have had to rely on either proprietary, architecture-specific solutions or low-level frameworks, which can complicate development and limit broader hardware adoption. This thesis explores hardware-software co-design for data-intensive applications, targeting the unification of programming models using open standards and exploring experimental techniques for automated query synthesis. We apply these approaches by progressing from dynamic relational data warehouses to computational genomics pipelines. First, to tackle the engineering complexity of database workloads, we present SYCLDB, an open-source, portable analytical engine. We then investigate the "synthesize-versus-engineer" debate by using a fully autonomous Large Language Model (LLM) framework (SHADB) to synthesize query-specific GPU execution code. By using AI to establish an empirical performance ceiling, we isolate and lift generalizable optimizations back into SYCLDB. Second, extending this cross-architecture thesis, we present X-BQSR, a holistic redesign of genomic base quality score recalibration pipelines. X-BQSRverifies that hardware acceleration must encompass the entire data path, showing that I/O redesign is essential to prevent data movement from bottlenecking GPU gains. Finally, we establish the architectural performance limits of these designs on emerging RISC-V processors. Together, these systems demonstrate novel cross-architecture algorithms and holistic pipeline optimizations that pave the way for the next generation of scalable, open-standard analytics engines.