Abdul Dakkak develops compilers, GPU libraries, and AI inference software that let programmers use high-level languages without giving up hardware-specific performance. His work spans Mathematica’s GPU integration, the Wolfram Language compiler, machine-learning evaluation and profiling, and Modular’s compute stack. He was chief scientist at Modular when he presented Taming the AI Hardware Cambrian Explosion at AI Engineer World’s Fair 2026.
Bringing GPU computing into Mathematica
At Wolfram Research, Abdul led development of Mathematica’s CUDALink and OpenCLLink, bringing GPU programming into an environment used for symbolic mathematics and scientific computing. Developers could load custom GPU functions, manage device memory, and coordinate computation across multiple GPUs. With Ulises Cervantes-Pimentel, he developed instructional material for these workflows and demonstrated interactive volumetric rendering and real-time ray tracing on consumer graphics hardware. The integration made accelerated computation available within an existing scientific workflow.
Abdul’s responsibilities expanded into the Wolfram Language compiler. As a primary developer, he developed its type system and architected its runtime. His 2020 computer-science doctorate at the University of Illinois Urbana-Champaign examined compilation while preserving the language features and runtime behavior users expected. Moving interpreted programs toward compiled execution had to retain the behavior that made existing programs useful. His dissertation evaluated compiled programs against highly tuned, handwritten C implementations.
Abdul subsequently worked as a principal research software engineer at Microsoft Research AI, developing compilers for end-to-end machine learning. His MLPerf-Bench professional biography describes that role and his earlier compiler leadership at Wolfram.
Making performance comparisons explainable
Alongside compiler development, Abdul worked on two related questions: what must be specified to reproduce a model evaluation, and how to trace an application’s performance across its software stack.
Reproducible model evaluation: Abdul co-developed MLModelScope with Cheng Li, Jinjun Xiong, and Wen-mei Hwu. The project specifies and provisions evaluation pipelines across frameworks and hardware so experiments can be repeated and compared. Their research demonstrated that seemingly small pipeline changes can alter measured accuracy, while framework and hardware choices affect performance. Reproducing a result therefore requires more than the model’s weights: it requires the surrounding evaluation procedure.
Across-stack profiling: With Cheng Li and collaborators at IBM, Alibaba, and Illinois, Abdul co-authored XSP. It uses distributed tracing to correlate measurements from the model, framework, libraries, and GPU, with an iterative measurement method that accounts for profiling overhead. Connecting these layers helps explain which underlying kernels contribute to a slow model operation, rather than leaving developers with disconnected measurements.
Portable kernels and the inference stack
At Modular, Abdul co-authored the team’s 2023 account of unified matrix multiplication, developed with Chad Jarvis, Eric Johnson, Hengjie Wang, Ian Tramble, and other Modular engineers. The design addresses the proliferation of separate implementations for different chips, numerical types, matrix layouts, and input sizes.
Its approach combines reusable kernel code with hardware specialization, supports dynamic shapes, and allows neighboring operations to be fused with matrix multiplication. Fusion can reduce memory traffic and dispatch overhead; dynamic-shape support avoids maintaining a separate compiled variant for every input size. These mechanisms explain how the team approaches portability while retaining opportunities to optimize for particular hardware.
Abdul’s public professional profile describes contributions spanning Mojo, libraries, kernels, framework and serving layers, including helping lead Modular’s generative-AI work. The company’s MAX framework brings modeling and serving together and supports NVIDIA and AMD GPUs.
In his 2026 hardware talk, Abdul argues that the growing variety of AI chips makes fragmented software stacks a central obstacle to using available compute. He presents Modular’s response across several layers: Mojo for programming CPUs, GPUs, and accelerators; optimized kernels; and MAX for inference across GPUs, nodes, and mixed hardware. This extends the problem visible in his earlier work: preserving a productive programming interface while giving compilers, runtimes, and kernels room to exploit the machine underneath.
Abdul Dakkak explains how Modular connects Mojo kernels, MAX inference and distributed deployment—and why its performance comparisons need to be read alongside their workloads, hardware and costs.
Modular connects three levels of work: Mojo implements portable device code, MAX executes and serves models, and Cloud operates distributed endpoints.
Mojo’s performance strategy combines portable kernel abstractions with low-level access; open kernels let developers test the claims on their own workloads.
MAX keeps model authoring in Python while exposing multi-GPU, multi-node and heterogeneous execution, including separate hardware for prefill and decode.
Time to first token, throughput and baseline-relative speedups answer different questions. Hardware price must enter the comparison before drawing a cost conclusion.
▶ Watch ↗