Okay, here is the summary of the provided Tenstorrent transcript in the requested format:
Tenstorrent's Second-Generation AI Dev Kit: Wormhole Unveiled
00:00:17: Introducing Wormhole: Tenstorrent showcases its second-generation AI dev kit featuring the Wormhole chip, an advancement over the previous Grayskull generation. Available as PCI cards or complete systems (Quiet Box/Loud Box).
00:01:04: Accessibility Goal: The primary aim of these dev kits is to make Tenstorrent's AI hardware accessible for developers and businesses, supported by open-source software and a community platform.
00:02:33: Hardware Configuration (Dev Kit): The development system shown uses liquid cooling, an AMD Genoa (Zen 4) CPU, 512GB RAM, and four dual-chip 300W Wormhole PCIe cards (8 Wormhole chips total).
00:03:13: High-Speed Interconnect: Chips are connected via 400 Gbps Ethernet per chip, enabling efficient chip-to-chip communication within the box and scaling out to other systems.
00:04:36: Pricing and Availability: Systems like the Quiet Box ($15k) and Loud Box ($12k) are available, alongside individual PCIe cards (150W single-chip: $1k, 300W dual-chip: $1.4k). Systems are expected to ship soon.
00:05:40: Software & Community Updates: Since the initial open-sourcing, Tenstorrent has focused on community engagement (Discord), integrating with PyTorch 2.0, developing a new MLIR-based compiler (Forge), and enabling external contributions via CI.
00:07:21: Native Scale-Out: Wormhole chips feature integrated Ethernet, enabling native scale-out capabilities critical for large AI workloads (training, inference, HPC). The dev kits provide a platform to develop for these scalable systems.
00:09:32: Live Demo (Llama 3.1 70B): A demo showcases the Quiet Box running a Llama 3.1 70B parameter model, tensor-sharded across the 8 Wormhole chips (2x4 mesh), handling 32 concurrent users (batch 32).
00:10:14: Consistent Programming Model: The software and programming approach used on the 8-chip dev kit scales directly to larger Tenstorrent systems (e.g., 32-chip Galaxy racks) with minimal adjustments.
00:14:49: Networking & Memory Bandwidth: Each Wormhole chip accesses 400 Gbps of network bandwidth and 288 GB/s of GDDR6 DRAM bandwidth, leading to an aggregate ~2.3 TB/s DRAM bandwidth across the 8-chip system.
00:16:26: Memory Capacity: Each chip has 12GB GDDR6, totaling 96GB capacity for the 8-chip system.
00:16:42: Block Floating Point (BFP): Wormhole hardware natively supports BFP data formats (BFP8, BFP4, BFP2). This uses a shared exponent across a block (16 numbers) to maintain high dynamic range while significantly reducing mantissa precision, enabling much higher model density.
00:20:40: Future of Low Precision: Tenstorrent is working towards native training in BFP formats, allowing models to be stored and run at very low precision (e.g., BFP2) for maximum efficiency. BFP2 could potentially fit a 400B parameter model on the 96GB dev kit.
00:22:24: Upcoming Server Hardware: Larger 32-chip (4x8 mesh) servers with 4x the network bandwidth and faster DRAM are planned for later this year (estimated price ~$50k).
00:25:45: Potential Alternative Use: The high onboard networking bandwidth suggests Wormhole chips could potentially serve as cost-effective, high-performance network switches.
00:28:01: Broader Strategy & Future: Tenstorrent aims to compete by offering more than just AI acceleration, integrating high-performance Risk-V CPU cores (Ascalon) in future products like Blackhole, while maintaining software compatibility.
Error: value error