December 2025: Implemented and delivered SFPI kernel support for stochastic rounding in Tenstorrent's LLK path (Blackhole and Wormhole). The feature adds stochastic rounding kernels for floating-point operations, enabling more accurate numerical behavior in critical paths. Changes touch tt_metal/third_party/tt_llk/tt_llk_blackhole and tt_llk_wormhole_b0 modules, and were validated via CI post-commit checks that passed. This work establishes the foundation for more reliable FP arithmetic in stochastic environments and paves the way for higher-precision results in downstream workloads.

September 2025 performance highlights for tenstorrent/tt-metal focused on reliability, scalability, and training efficiency. Delivered compute kernel performance optimizations with expanded testing coverage for optimizer configurations, added momentum to the fused SGD optimizer, and resolved synchronization gaps in compute callbacks. These changes reduce nondeterministic results, improve convergence speed, and strengthen ML workload stability across diverse configurations.