in High Performance Computing Victor Eijkhout CppCon 2024 Introduction This poster reports on ‘D2D’, a benchmark that explores elegance of expression and perfor- mance in the context of a High Performance implemented using a number of different approaches to parallelism. Implementations are discussed with performance results. Relevance C++ is making inroads into HPC / Scientific Computing, a field traditionally multi-dimensional arrays through ‘mdspan’, it is interesting to explore what C++ can offer for lower level performance critical operations. Scientific computing is an interesting test cases since many algorithms are

Introduction Firsts steps Context Theoretical foundations Outline of an implementation Conclusion High-Performance Numerical Integration in the Age of C++26 Vincent Reverdy Laboratoire d’Annecy de Physique past, other languages do far better in terms of everything: functionality, ease of use, and even performance This talk The goal is NOT to revolutionize everything or show a library that beats everything Verner’s methods Explicit adaptive Runge-Kutta schemes Slightly enhanced methods with interpolants High-accurary interpolation inside integration steps Interpolation with extra coefficients βi yn+u =

is moved into general-purpose registers for computations • Depending on the platform, may see a performance gain at this stageQuat Functions template inline

Binding Constraints Architecture Substitution Construction Conclusion Symbolic Calculus for High-Performance Computing from Scratch using C++23 Vincent Reverdy Laboratoire d’Annecy de Physique des Particules all know about optimization, performance, parallelism, . . . What this talk is not about Complicated maths (you are smart people, you can do it yourself) High-performance computing (you all know about concepts Symbolic calculus (derivatives, integrals) Full blown custom rule-based rewriting High-performance Since formulas have the entire information on the mathematical AST, it’s possible to generate

PERFORMANCE MATTERS (joint work with Charlie Curtsinger, Grinnell College) emeryberger.com, @emeryberger Emery Berger College of Information and Computer Sciences UMASS AMHERSTA short time ago : un.bmp Ogle is too slow! OGLE’84 is too slow!Transistors (millions) Clock Speed (MHz) Performance used to be easy 0.001 0.01 0.1 1 10 100 1,000 10,000 1970 1975 1980 1985 1990 1995 gle loading… No mojitos for me… Back to the present…Transistors (millions) Clock Speed (MHz) Performance not easy anymore 0.001 0.01 0.1 1 10 100 1,000 10,000 1970 1975 1980 1985 1990 1995

Being Friendly to Your Hardware Performance Engineering A gentle introduction to hardware for software engineers 2Where does C++ run? 3On an abstract C++ machine 4On an abstract C++ machine? multilayer componentsMemory • What about HBM? Not 1024 bits, 8 x 128 instead. Bandwidth will only be high when there is sufficient stream of commands inflight • What about remote memory? CXL and vendor In most practical cases at boot time only Same capacity, different composition => different performance profile From JESD 79-4 DDR4 specificationMemory • Memory system is in the uncore • Cores act

Research Tech at Tower Research Capital ○ High frequency trading firm based out of NYC ● Develop low latency trading systems (C++) ○ Nanoseconds matter ● Develop high throughput research systems (C++ and volume in terabytes ● Program analysis research and functional programming in a past life ● Love performance, software abstractions, and clean APIsWhy Python? ● Writing extensive APIs in Python - low boilerplate We’re at CppCon :) Why Python? Why C++?● Why? ○ Avoid reimplementing complex code for Python ○ Performance ○ Back and forth with user’s python code ○ Interoperability with data structures in Python -

Equipment Effectiveness (OEE). Energy companies must balance EV charger loads and manage grid performance in real time. Banks need to analyze audit logs from their website and application in real time frauds. Logistics companies need real-time tracking and historical analysis of shipments, fleet performance, and warehouse operations to optimize delivery times, reduce costs, and improve supply chain efficiency Technology platforms require real-time monitoring and analytics to personalize experiences and ensure performance. 32. The Interconnection of Analytics, Search, and AI Analytics, search, and AI are deeply interconnected

User + Usage + CapEx Growth = Unprecedented • AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising • AI Usage + Cost + Loss Growth Page 293 USA – LLM #1 China USA – LLM #2 AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising 3 Cost of Key Technologies Relative competitive. Breakthroughs in large models, cost-per-token declines, open-source proliferation and chip performance improvements are making new tech advances increasingly more powerful, accessible, and economically