▶ Watch ↗AI Engineer World's Fair 202420:46
Conference affiliation: Microsoft · 2024
Lachlan Ainley works in product marketing for AI infrastructure, explaining how choices across a computing system affect the workloads customers can run. His work spans Azure’s infrastructure strategy and the Microsoft–NVIDIA Catalyst series, connecting computing platforms with concrete applications of AI.
A Microsoft senior product marketing manager at AI Engineer World’s Fair 2024, Lachlan presented Azure’s infrastructure strategy as a combination of accelerator hardware, host processors, virtual machines, networking, and system topology—the arrangement and connections between machines.
In his 2024 conversation with Snorkel AI researcher Humza Iqbal, Lachlan explained why Azure optimizes these layers together. An accelerator’s usefulness depends on the throughput of its host and the connections carrying data between components. He described a cycle in which Microsoft’s experience supporting large model developers, including OpenAI and Mistral, informs the infrastructure it offers other customers.
Matching accelerators to workloads was central to that strategy. Lachlan discussed NVIDIA GPUs, AMD’s MI300X, and Microsoft’s Maia silicon, identifying Maia’s use in Microsoft’s internal workloads. His account treated new chip generations as opportunities to reconsider which hardware belongs on which job, rather than prescribing one accelerator for every application.
Lachlan’s questions brought the customer’s operating experience into the discussion: how Snorkel got started on Azure, what went wrong, how much flexibility its experiments required, and what changed when it moved from A100 to H100 GPUs. Humza supplied Snorkel’s distributed-training methods and performance comparisons; Lachlan connected that experience to Azure’s infrastructure choices.
One comparison made the stakes clear. Humza reported that, for Snorkel’s tested workloads, two H100s outperformed four A100s at roughly the same cost. Faster inference also let the team generate more synthetic data and iterate more quickly. Lachlan highlighted the resulting ability to accomplish more with fewer GPUs: hardware selection changes the cost and pace of experimentation, alongside the number of machines required.
Lachlan’s credited contributions to the Microsoft–NVIDIA Catalyst series extend that product-marketing work beyond infrastructure. They include an episode featuring Pangaea Data’s work on disease identification and another featuring Basecamp Research’s work on biodiversity and biological discovery. His contribution was to the series presenting these applications; the featured companies developed the underlying products. These examples give his infrastructure work a practical destination: helping audiences understand what organizations can do with the computing capacity they adopt.
▶ Watch ↗AI Engineer World's Fair 202420:46
Loading more talks and highlights…