Specifications
Flagship Model
Xiyun C500 — PCIe general-purpose GPU accelerator card
Compute, FP16 / BF16
280 TFLOPS dense
Compute, INT8
560 TOPS
Compute, TF32
140 TFLOPS
Compute, FP32
18 TFLOPS
GPU Memory
64 GB HBM2e with end-to-end ECC protection
Board Power
350 W maximum board power
Form Factor
Full-height, full-length, dual-slot PCIe card
Interconnect
MetaXLink — 2-way or 4-way card-to-card
Virtualization
Software-defined partitioning down to 1% granularity
Software Stack
MXMACA, CUDA-idiom compatible with an SDK and driver suite
Next-Generation Model
Xiyun C600 — 144 GB HBM3e, OAM form factor, domestic 7 nm process
Inference Card
Xisi N100 — 16 GB GDDR6, 80 TFLOPS FP16 / 160 TOPS INT8, 75 W, PCIe 4.0
C-Series Portfolio
C500, C550, C500X, C588 and C600
Additional Series
N-Series (N100/N260/N300), G-Series (G100), X-Series (X206/X301/X302)
Supernode Options
Liquid-cooled C500 cabinet, C500X optical-interconnect supernode, C550 3D Mesh supernode
Workloads
LLM training and inference, AIGC, HPC and data processing
Deployment
Cloud AI training and inference data centres, workstations and supernodes
Overview
The MetaX Xiyun C500 is a general-purpose GPU accelerator built on MetaX's own core GPU IP. It pairs 64 GB of HBM2e with end-to-end ECC protection and a full-height, full-length dual-slot PCIe form factor drawing up to 350 W. Dense throughput is rated at 280 TFLOPS FP16/BF16, 140 TFLOPS TF32, 18 TFLOPS FP32 and 560 TOPS INT8 — the multi-precision spread that training and inference workloads expect.
Multi-card scaling uses MetaXLink, which supports high-speed 2-way or 4-way card-to-card topologies, while software-defined partitioning allows the card to be sliced at a granularity as fine as 1% for multi-tenant inference. The MXMACA software stack is deliberately CUDA-idiom compatible so that teams can migrate existing kernels and frameworks with minimal rewrite; the driver and SDK are distributed directly by MetaX.
MetaX's portfolio extends well beyond a single card. The C-Series spans C500, C550, C500X, C588 and the HBM3e-based C600 on a domestic 7 nm process with an OAM form factor; the N-Series (N100/N260/N300), G-Series (G100) and X-Series (X206/X301/X302) cover other price and performance tiers. Above the card level MetaX ships turnkey supernode designs including a liquid-cooled C500 full cabinet, a C500X optical-interconnect supernode and a C550 3D Mesh supernode.
Key Benefits
Training-class memory footprint at 64 GB HBM2e with ECC; multi-precision throughput from 18 TFLOPS FP32 up to 560 TOPS INT8 on one card; flexible 2-way/4-way MetaXLink scaling for small-cluster builds; 1% granularity virtualization for inference multi-tenancy; and a CUDA-idiom MXMACA stack that shortens porting work for teams standardising on non-NVIDIA silicon.
Applications
LLM training and fine-tuning; large-scale inference serving; AIGC and generative workloads; HPC and scientific simulation; data processing and analytics pipelines; private-cloud and sovereign AI data centres; and GPU workstation deployments on the N260 platform.
Request a Quote — METAX XIYUN C500 & C600 — DOMESTIC GPGPU ACCELERATOR CARDS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →