MetaX Xicloud C600

What it is

The Xicloud C600 is a data-center GPU from MetaX. The company describes it as a general-purpose GPU built on its own GPGPU architecture and made on a domestic advanced process.^[1] "General-purpose" here means the chip is meant for computing rather than for drawing images: the same kind of processor that renders graphics can be programmed, through compute interfaces, to carry out the matrix and vector arithmetic that machine-learning models are built from. Building on its own architecture means the processor design is MetaX's own.^[1]

Memory and reliability

MetaX lists ultra-high-capacity GPU memory with end-to-end ECC protection as a defining feature of the C600.^[1] Memory capacity decides how large a model, or how much of a model, can sit on one device; when a model does not fit, it has to be split across chips and the pieces have to talk to each other constantly. ECC stands for error-correcting code, a scheme that detects and repairs the bit errors that cosmic rays, heat and manufacturing variation can cause in memory. "End-to-end" means the protection covers the whole data path rather than only the memory chips, so errors picked up in transit are caught as well. For jobs that run for hours or days across many devices, a single undetected flipped bit can spoil a result or kill a run.

Scaling out with MetaXLink

For multi-GPU work the C600 uses MetaXLink, which MetaX describes as a low-latency interconnect, together with a MetaXLink-E interface that supports supernodes of up to 128 GPUs.^[1] A supernode is a group of accelerators wired closely enough that they can be programmed almost like one large machine. This matters because large models do not fit on a single card: the work is divided up, and the divided parts must exchange data at every layer. The faster and shorter those exchanges, the less time the accelerators spend waiting on one another instead of computing.

Form factor, power and cooling

The C600 is an OAM 2.0 module with a maximum board power of 1,000 W, and it supports both air and liquid cooling.^[1] OAM, the Open Accelerator Module, is a standard for data-center accelerator cards that slot into a shared baseboard, so a server maker can put several accelerators on one board using the same mechanical and electrical design. Power is the other constraint: a module that draws up to 1,000 W concentrates a great deal of heat in a small space, which is why dense systems increasingly move to liquid cooling. Offering both options lets the same part be used in data centers built around either approach.

Software

The C600 runs MetaX's MXMACA software stack, which the company says is compatible with mainstream GPU ecosystems.^[1] A software stack is the layer between the hardware and the programmer: drivers, math and communication libraries, compilers and debugging tools. Compatibility with existing ecosystems matters because a data-center GPU is only useful if the frameworks and models people already use can run on it; the closer the fit, the less code has to be rewritten when moving work onto the hardware.

What the design adds up to

Taken together — its own GPGPU architecture, a domestic advanced process, very large ECC-protected memory, a 128-GPU supernode fabric, a 1,000 W OAM 2.0 module and a choice of air or liquid cooling — the C600 is specified for large-scale data-center deployment rather than for small or embedded systems.^[1] The software compatibility claim points the same way: the aim is to fit into existing GPU-based workflows, so that the hardware can be adopted without rebuilding the software around it.