Tenstorrent Wormhole n300s

The Wormhole n300s puts two Wormhole processors on one passively cooled PCIe Gen4 card at $1,399, and because Wormhole carries Ethernet on the die the pair is programmed as one continuous Tensix grid rather than as two devices. That gives 128 Tensix cores, 192 MB of on-die SRAM and 24 GB of GDDR6 at 576 GB/s, drawing up to 300 W. Compute is 466 TFLOPS of dense FP8 and 131 TFLOPS of dense FP16, roughly 1.78 times the single-processor n150 because each processor here runs 64 Tensix cores rather than 72. Two QSFP-DD 200G ports carry the mesh off the card to other cards without going through the host, which is the mechanism Tenstorrent scales with. As with every Tenstorrent part, the real draw is that tt-metal and tt-forge are open source top to bottom.