Xsight Labs launches 800G E1 DPU

A DPU is the chip that takes networking, storage, security, and infrastructure chores away from the main server CPU. That matters more now because (xsightlabs.com)sooner, or buy a faster one wrapped in proprietary tooling. Xsight Labs is trying to break that tradeoff with the E1, an 800G D(xsightlabs.com)ped out or previewed. (xsightlabs.com)lleled-fully-software-defined-network-architecture)) ### What is the E1, exactly? The E1 is an Arm-based DPU SoC built on TSMC 5nm with 64 Neoverse N2 cores, 32MB of L3 cache, and dual 400G Ethernet ports for 800 Gbps total throughput. Xsight also lists standard Linux support plus XDP, DPDK, and SPDK, which is the big clue about what kind of product this is trying to be. (xsightlabs.com) ### Why does “software-defined” matter here? Most DPUs still inherit a lot from smartNIC design — fixed pipelines, specialized blocks, and vend(xsightlabs.com)ead of a proprietary stack. Basically, it wants to feel more like developing on Linux than programming around a sealed networking appliance. (xsightlabs.com) ### How fast is the thing? The raw number is easy: 800G. The more interesting part is how Xsight says it gets there. Independent cover(xsightlabs.com) through the Arm cluster with inline engines handling parsing, classification, crypto, and other offloads, so the chip is not just a NIC glued to some management cores. That is unusual — and it is the core architectural bet. (xpu.pub) ### What is Xsight claiming on efficiency? Xsight’s product pages say the E-series delivers double the performance at half the power, and the E1 page lists 70W typical power. The company’s earlier announcement framed (xsightlabs.com)oud and edge AI data centers. Those are vendor claims, so the real test is customer deployment data, but the direction is clear: replace chunks of server overhead with one denser network processor. (xsightlabs.com) ### What can you build around it? Not just one card. Xsight is selling the E1 as a bare SoC and also as server and add-in-card platforms. It has also teamed with(xpu.pub)ers because many buyers do not want to redesign an entire server just to trial a new DPU. They want something they can slot into an existing validation flow. (xsightlabs.com) ### Why does AI infrastructure care? Because AI clusters are increasingly bottlenecked by (xsightlabs.com)ot just GPU math. Xsight and Hammerspace have been pitching the E1 for “warm storage” designs where GPUs can reach flash across the network without bouncing through layers of x86 servers. If that model works in production, it could flatten parts of the data path and cut some expensive infrastructure fat. (xsightlabs.com). Nvidia BlueField, Intel infrastructure offload products, and various smartNIC approaches already sit inside real data-center roadmaps. So Xsight does not just need a faster chip — it needs a better developer story, cleaner integration, and proof that “open” also means supportable at scale. That is why availability matters more than specs alone. (xsightlabs.com)another faster NIC. It is a serious attempt to make an 800G DPU act like an open Linux machine with very high network throughput. If Xsight can back the architecture with real deployments, this could change how data-center teams scope offload, storage, and east-west networking around AI clusters. (xsightlabs.com)