Search Results: Tenstorrent

A developer's deep dive reveals how moving from Python to C++ static autograd with tracing delivers 11x speedups on Tenstorrent's Wormhole hardware, while exploring the surprising limitations of parallelism at GPT-2 scale. The benchmarks expose a fundamental divide between dispatch-bound and compute-bound workloads.

Legendary GPU architect Raja Koduri has unveiled Oxmiq Labs, a new venture developing RISC-V-based GPU IP and a groundbreaking software stack that enables unmodified CUDA workloads to run on non-Nvidia hardware. Backed by $20 million from MediaTek and partnering with Tenstorrent, Oxmiq aims to democratize AI acceleration with its hardware-agnostic platform.

GPU legend Raja Koduri unveils OXMIQ Labs after two years in stealth, backed by $20M in funding and partnerships with Tenstorrent and MediaTek. The startup aims to disrupt the accelerator market with licensable GPU IP featuring RISC-V nano-agents and a 'software first' approach that enables CUDA workloads to run on non-NVIDIA hardware without modification.

AI's explosive growth is bottlenecked by an unquenchable demand for specialized hardware, with NVIDIA dominating 80% of training workloads while startups and hyperscalers vie for a piece of the projected $77 billion market. Trust issues, ecosystem lock-in, and unsustainable costs threaten ambitious challengers like Tenstorrent, setting the stage for a high-stakes battle over the future of AI infrastructure. This deep dive explores the technical and economic forces driving the industry toward a likely consolidation under tech giants, leaving little room for revolutionary upstarts.