Moore Threads MTT C256 Supernode: 256-GPU Single-Layer Scale-Up Architecture & S5000 Chip Deep Dive

At WAIC 2026, Moore Threads publicly demonstrated the MTT C256 supernode for the first time, turning its roadmap into a shipping product. The system organizes 256 GPUs (MTT S4000/S3000 series) within a single Scale-up network fabric, using a compute-switch integrated high-density design that eliminates the bandwidth loss and forwarding latency inherent in traditional multi-layer network topologies. It targets three core AI workloads: model training as the building block for 100,000-GPU clusters supporting pre-training, post-training, and reinforcement learning; inference service with native large-scale MoE adaptation and million-token context support; and inference decoding for high-concurrency, real-time scenarios such as AI coding assistants.

Single-Layer 256-GPU Scale-Up Breakthrough

The central innovation is a single-layer Scale-up network that lifts the industry-standard 64-GPU limit directly to 256 GPUs. This fundamentally mitigates the bandwidth attenuation and communication latency problems of multi-layer architectures. A single cabinet achieves 128-GPU full mesh; two standard cabinets complete the 256-GPU single-layer interconnect.

Moore Threads Full-Family Chip Portfolio

The product line centers on full-function GPUs. The MTT S5000 is in volume production, while the next-generation "Huagang" architecture is under planning. The "Changjiang" SoC adopts a unified memory architecture, enabling zero-copy data movement between CPU and GPU across graphics, AI, and other workloads to maximize task execution speed.

MUSA Unified System Architecture

MUSA (Moore Threads Unified System Architecture) is a self-developed fused GPU hardware/software stack covering unified chip architecture, instruction set, programming model, software runtime libraries, and driver framework, aiming to deliver high-performance parallel computing across diverse scenarios.

MTT S5000 Technical Analysis

The MTT S5000 serves as the core hardware for Prefill-as-a-Service compute pools in heterogeneous PD (prefill/decode) separation architectures, pairing with GPU A (decode generation pool). Its key technical attributes:

High-density FP8 compute : Matches prefill's compute-intensive profile; 8-card single-node throughput reaches 95,920 tok/s at 64K input length.

Long-context stability : Across 64K–400K input range, P50 TTFT (time-to-first-token) remains stably within a 6-second SLO.

Native FP8 KV Cache : Precision-aligned with the decode pool, supports page-granularity direct transfer, eliminating format-conversion overhead.

MUSA ecosystem CUDA compatibility : Source-level migration; supports mainstream frameworks including PyTorch, SGLang, and vLLM.

GPUDirect RDMA + Mooncake : 400G RDMA networking enables efficient cross-node KV Cache transfer.

Full-precision support : FP8 through FP64; among the earliest domestic GPUs to support FP8.

The article positions the MTT S5000 as a general-purpose AI accelerator card purpose-built for the Prefill compute pool in disaggregated inference architectures.