Student design team · University of Waterloo

HighPerformanceComputing

UWHPC

A student design team at the University of Waterloo pushing the boundaries of parallel and high-performance computing.

Scroll to power on
00/ Who we are

We're a student design team at the University of Waterloo working on high‑performance computing. We write software that makes demanding workloads run faster and at scale.

01/ Our Work

Projects

What we're building now, and the tools we've built along the way.

P01In Progress

BATS

Beamforming Across Tensor-core Systems

Beamforming combines signals from an array of sensors so that one direction adds up and the rest cancel out. It sits at the core of radio astronomy, radar, and medical ultrasound. Forming many beams across frequency channels is a batched complex matrix multiply, which tensor cores accelerate. BATS is building a distributed tensor-core beamformer across GPUs and nodes, with a planner that chooses how to split the work from the workload, where the data lives, and the interconnect.

Leads Alyan Salamat · Shpat Sahiti

S = 16 · B = 3
-60°-30°0°30°60°sensors · d = λ/2dB, floor −30
per channel fYf = Wf · Xf(B×S)(S×T)
P02In Progress

Maverick

Topology-aware GPU execution of graph neural networks

Two graphs with the same number of edges can produce very different GPU workloads: a handful of high-degree nodes can leave most of the GPU waiting. Maverick analyzes the topology of a static graph once and uses it to choose specialized execution strategies for GNN inference, measured against TorchInductor under torch.compile, with the cost of that analysis counted against the speedup.

Leads James Redekopp · Vijay Goyal

|V| = 24 · |E| = 48
Graph A · balanced
work per node · one group each
mean
4
max
4
idle
0%
P03In Progress

XPU

A header-only C++23 library that lets the same allocation, layout, and math code compile for either CPU or NVIDIA CUDA. Backend-aware buffers, structure-of-arrays storage, OpenMP + SIMD parallel loops on the CPU, and launch helpers on the GPU. It's the performance foundation shared across UWHPC projects.

soa.cu
enum Axis : std::size_t { X, Y, Z, NUM_AXES };

xpu::soa<float, Axis::NUM_AXES> position{N};

float* x{position[Axis::X]};
one allocation · aligned operator new
count()
10
stride()
16
storage_size()
48

CPU: rows padded to 64-byte SIMD lanes

P04In Progress

Variational Monte Carlo

A Variational Monte Carlo engine in CUDA C++ for the homogeneous electron gas. A plane-wave Slater determinant with a Padé–Jastrow factor, sampled by Metropolis moves in a periodic cell, with Jastrow optimization and blocking analysis for honest error bars.

N = 32 · periodic cell
No Jastrow · uncorrelated
g(r) · pair correlation, sampled
r = 0half cell
accept
100%
g(0⁺)
1.02
samples
150
02/ What We Do

Focus Areas

We work across the stack, from numerical methods down to the code generation and systems that make them fast.

  • 01

    Scientific Computing

    Numerical methods and simulation for physics and the computational sciences, including Monte Carlo techniques and large-scale numerical workloads that have to run fast enough to be useful.

  • 02

    Machine Learning Infrastructure

    The systems layer beneath the models: data and training pipelines, GPU utilization, and the throughput and memory limits that decide how quickly experiments can actually run.

  • 03

    Compilers

    How source becomes fast machine code. Intermediate representations, optimization passes, code generation, and understanding compiler behaviour well enough to know when hand-tuning is worth it.

  • 04

    Distributed Systems

    Splitting work across many machines with MPI and related tooling, and handling the coordination, communication, and failure behaviour that comes with it.

  • 05

    Performance Engineering

    Profiling real workloads to find bottlenecks, then closing them with vectorization, cache-aware tuning, and low-level work in C, C++, and Rust.