Blaize Xplorer X1600E & X600M AI Accelerators

Models & Pricing

Specifications

Overview

Blaize approaches edge inference from a different angle than most accelerator vendors. Instead of a fixed-function NPU, its Blaize 1600 SoC is built on the Graph Streaming Processor (GSP) architecture — a fully programmable, graph-native engine designed to execute an entire AI application as a streaming dataflow rather than a sequence of kernel launches. The practical consequence is latency and host-load behaviour that looks much more like dedicated streaming hardware than like a GPU.

The Xplorer X1600E packages that silicon as an EDSFF 1U ruler card: 16 GSP cores, 16 TOPS, 4GB of LPDDR4 and PCIe Gen 3.0 x4, in a form factor that drops directly into standard server bays. Because the card requires near-zero host CPU involvement, a 1U chassis can hold 24 of them for roughly 400 TOPS of inference density without the power and thermal cost of a GPU-based equivalent. Blaize has publicly collaborated with Supermicro on exactly this configuration, and volume pricing has been quoted in the region of $299 per X1600E card — a cost-per-TOPS profile that is hard to match with datacenter GPUs.

For embedded and industrial systems, the X600M M.2 module and the X1600P / X1600P-Q PCIe cards deliver the same architecture in smaller footprints, with a soft ISP option that runs on the Blaize 1600 SoC so no separate image signal processor is required. The Pathfinder P1600 system-on-module and 1600-DK development kit serve customers building their own AI appliances on custom carriers.

Everything Blaize sells is programmed through the Picasso Software Development Platform, which is the strategic point of the portfolio: rather than betting on a fixed neural-network shape, customers write graph-level applications that can be re-optimised as model architectures evolve. For industrial and edge-datacenter buyers who have been burned by accelerator obsolescence, that programmability plus the EDSFF server-bay integration story is the core argument.

Target Applications

On-premises and edge datacenter AI inference, real-time multi-camera video analytics, industrial machine vision, streaming inference within 1U server appliances, edge vision at industrial PCs, and AI appliance designs built on the Pathfinder system-on-module.

Related Products