SambaNova raises $350M to rival Nvidia GPUs

🔑 Enhanced Key Takeaways

• •SambaNova's SN50 chip achieves up to 895 tokens per second per user on Llama 3.3 70B model, outperforming Nvidia B200's 184 tokens per second per user.[2][3]
• •SoftBank is the first customer deploying SN50 in next-generation AI data centers in Japan for low-latency inference services.[2][3][4]
• •Intel's new CEO Lip-Bu Tan serves as SambaNova's Chairman and early investor via Walden International.[3]
• •The partnership includes multi-year collaboration for hardware-software co-design on Intel Xeon infrastructure, with joint co-selling to enterprise channels.[2][3]

🛠️ Technical Deep Dive

• •SN50 is the 5th-gen Reconfigurable Dataflow Unit (RDU) accelerator, optimized for large language and multimodal models on Intel Xeon-based infrastructure.[2][3]
• •Equipped with 2.2 TB/s bidirectional chip-to-chip bandwidth via switched fabric for larger scale-up domains.[2]
• •Supports agentic AI workloads and provides scalable, production-ready inference capacity through integration with Intel's CPUs, accelerators, and networking.[3][4]

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline