AMD MI300X
AMD's flagship GPU accelerator for AI training and inference, built with chiplet packaging and integrating 192 GB of HBM3 unified memory. Peak BF16 compute reaches approximately 1.3 PFLOPS per card. The accelerator is supported by the ROCm software stack.
Timeline (last 90 days)
The RL-Kernel team ran a 200-step Qwen3-8B GRPO training task on 8 AMD MI300X cards, with results showing zero logprob mismatch between the Megatron training side and the vLLM rollout side.
The AMD MI300X is AMD's flagship accelerator for AI training and inference, featuring Chiplet packaging, integrated 192 GB HBM3 unified memory, and a peak BF16 compute performance of approximately 1.3 PFLOPS per card.
All Facts (2)
The RL-Kernel team ran a 200-step Qwen3-8B GRPO training task on 8 AMD MI300X cards, with results showing zero logprob mismatch between the Megatron training side and the vLLM rollout side.
50%UnverifiedThe AMD MI300X is AMD's flagship accelerator for AI training and inference, featuring Chiplet packaging, integrated 192 GB HBM3 unified memory, and a peak BF16 compute performance of approximately 1.3 PFLOPS per card.
50%