AMD unveils new MI455 AI chips at CES

AMD showed off its new MI455 AI chips at the CES event in Las Vegas on January 5.

The company introduced these processors as part of its effort to supply data center hardware for clients such as OpenAI.

AMD signed a deal worth tens of billions of dollars with OpenAI in October 2025 to provide these chips for use in server racks.

No additional technical details or pricing for the MI455 chips were disclosed during the presentation, according to Reuters.

OpenAI president Greg Brockman appeared at the event alongside AMD CEO Lisa Su.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

MI455 timeline depends on HBM3E supply and 2.5D packaging

MI455 needs High Bandwidth Memory (HBM), including HBM3E, plus stacked memory and tight packaging. SK Hynix plus Micron face yield problems on 12-high (12-die) stacks, while HBM failures are a top driver of GPU failures in data centers 1.
These accelerators use an interposer-based 2.5D package that puts compute dies and memory side by side on a thin silicon interposer. Chip-on-Wafer-on-Substrate (CoWoS) is one route, and output depends on HBM supply plus packaging slots from partners 1.
AMD has not shared production dates or firm memory allotments. Delivery to OpenAI versus rivals with mature supply chains remains hard to judge.

Software vendors can help ease CUDA-to-ROCm migration

Teams moving from Compute Unified Device Architecture (CUDA) to Radeon Open Compute (ROCm) will lean on HIPIFY, a source-to-source converter for Heterogeneous-computing Interface for Portability (HIP). Case studies report about 90–99% auto conversion in some projects 2, while complex apps still need hand tuning and debugging 3.
Service shops plus systems integrators can fill gaps where HIPIFY stops short, such as CUDA libraries without HIP twins and hot paths that need performance work 4.
Training offers another path, since engineers face a learning curve when switching platforms 5. Firms that build ROCm skills now can become go-to migration partners for enterprises that want options beyond NVIDIA gear.