Multi-card concatenation technologies like SLI and CrossFire have been distant memories for many years due to technical immaturity, poor game compatibility, and product positioning issues. However, they remain valuable in the professional world. AMD's newly released ROCm 6.1.3 development kit now supports the configuration of multiple GPU cards in a single system.

AMD stated in the changelog that multi-card support allows for more flexible AI desktop solutions in multi-service, multi-user environments.

Currently, the models that support multi-card parallelism are limited to the high-end series of RDNA3 Navi31 cores, which include:

RX 7900 XTX, RX 7900 XT, RX 7900 GRE, PRO W7900 (Dual Slot), PRO W7900, PRO W7800.

Of these, the 7900 XTX and PRO W7900 can operate in dual-card parallel configurations, and the PRO W7900 (Dual Slot), officially supported for the first time, can support up to four cards in parallel.

AMD claims that this capability allows users to better execute ultra-large-scale models, such as the Llama 3 language model with 70 billion parameters. Using four PRO W7900s can provide up to 192GB of GDDR6 video memory.

Additionally, ROCm 6.1.3 supports TensorFlow on select RDNA graphics cards, with initial support for running ROCm through the Windows WSL subsystem.

Note that all of the above is limited to the Ubuntu 22.04.3 HWE operating system and requires pairing with Linux version 24.10.3 graphics drivers.