Intel and SambaNova Team Up to Compete with NVIDIA’s Groq Partnership in AI Inference Solutions

In the continuously evolving landscape of artificial intelligence (AI), inference solutions are gaining increasing attention from computing providers. Following the recent partnership between NVIDIA and Groq, the AI community has recognized the pressing need for a diversified approach beyond traditional GPUs. Enter the emerging collaboration between Intel and SambaNova.

Intel’s Xeon 6 CPUs to Serve as the Host for Agentic Systems, Powered by SambaNova’s SN50 Chip for Inference

During NVIDIA’s recent GTC event, the concept of disaggregated inference came to the forefront, with the manufacturer emphasizing the shift away from a purely ‘GPU-centric’ approach. This strategy involves incorporating a variety of advanced computing units into their infrastructure. This evolution is highlighted by the licensing agreement with Groq, which integrated SRAM-based LPUs into Rubin’s LPX racks. Intel and SambaNova have decided to venture into similar territories, introducing an innovative “inference architecture”that merges SambaNova’s RDUs with Intel’s Xeon 6 CPUs.

SambaNova has announced the latest phase in its collaboration with Intel: a heterogeneous hardware solution that combines GPUs for prefill tasks, Intel® Xeon® 6 processors as both hosts and “action” CPUs, and SambaNova RDUs for decode, aiming to deliver superior inference for demanding Agentic AI applications.

– SambaNova

This collaborative arrangement is strategically designed to optimize workloads, with RDUs dedicated to decode tasks while GPUs manage the prefill processes and Xeon 6 CPUs handle orchestration and general computing tasks. Notably, this partnership does not tie down a specific hyperscaler for GPU solutions, leaving room for potential integration with ASICs. However, SambaNova has not elaborated on the specific performance metrics for GPUs within this setup. The company recognizes the SN50 units as pivotal in achieving “end-to-end coding agent workflows, ”asserting their superiority over ARM alternatives.

A person holds a semiconductor chip with visible text, 'sambanova, ' labeled 'CEREBRAS 500K M2019-05' and 'TITANIUM.' — Image Credits: SambaNova

Now, let’s delve into the details of the SN50 chip. Announced in early 2026, this solution features SambaNova’s fifth-generation RDU units, integrating advanced memory technologies including DRAM, SRAM, and HBM. With a substantial 2TB of DDR5 memory, 64GB of HBM3, and 520MB SRAM, this architecture is designed to minimize latency while maximizing throughput and capacity. According to SambaNova, the unique combination of DRAM, SRAM, and HBM facilitates what they term ‘agentic caching, ’ providing a competitive edge in processing efficiency.

A comparison chart titled 'Landscape: Delivering Premium Inference on the Largest Workloads' showing data for Intel, NVIDIA, and Cerebras solutions. — Screenshot

When comparing the strategies of Intel and SambaNova with those of NVIDIA, a notable distinction emerges. Intel’s partnership leans towards a more pragmatic approach, minimizing the requirement for extensive underlying infrastructure typically needed for disaggregated inference. This makes the Intel-SambaNova solution attractive for hyperscalers seeking a more modular rack-scale offering that effectively addresses the “prefill + decode”division.

Furthermore, Intel’s CEO actively participated in SambaNova’s latest funding round, alongside Lip-Bu Tan, an early investor in the company. Although discussions surrounding an acquisition took place, they were reportedly stalled due to board disagreements, leading Intel to opt for a role as a funding participant instead.

Source & Images