The 6th-generation AMD EPYC plus Instinct MI455X AI accelerator have been confirmed for stable mass production and shipment, as AMD’s latest announcement revolves around the new Helios rack-scale platform that can become a serious contender to NVIDIA’s rapidly expanding AI factory ecosystem. Other than the CPUs and GPUs, the full solution also comes with Pensando networking and an open rack-scale design made for frontier AI training, inference, and fine-tuning that also provides greater flexibility compared to proprietary ecosystems.
One of the interesting parts is the 6th Gen EPYC Venice CPUs being the very first “Zen 6” that packs more power than ever with its 8x massive compute dies and 2x I/O dies offering up to 256 cores and 512 threads, leading to up to 70%+ performance & efficiency improvement. AMD says that these are optimized for agentic AI workloads, as the trend of AI has gone from purely “models” to now “harness” and “loops”.
Meanwhile, the Instinct MI455X doubles its predecessor’s output at 40 PFLOPs FP4 and 20 PFLOPs FP8 compute, while memory has been jacked up from 288GB HBM3e to 432GB HBM4, with the use of HBM4 standard delivering a superior 19.6 TB/s bandwidth. While it is still not the absolute best when compared to NVIDIA Rubin’s 288 GB HBM4 at 22TB/s, the capacity difference is something that hyperscalers will notice.
There are other tiers of accelerators too – the MI450X is just a bit tamer than the MI455X but still made for AI training and inference, while the MI430X focuses on FP64 capabilities and hybrid computing, and is said to be suitable for HPC.
In terms of who’s gonna welcome AMD Helios into their data centers, Microsoft is among the latest to join in, buffing its Azure cloud computing platform to power workloads like Electronic Design Automation (EDA) and the usual AI stuff. Its decision to purchase Hekios is to provide clients with a heterogeneous AI platform that combines its own silicon with technologies from partners like AMD to optimize performance, cost, and power efficiency.
Microsoft isn’t the first hyperscaler to commit to Helios. AMD has already announced OpenAI, Meta, Oracle, and several global infrastructure partners as early adopters, highlighting growing industry interest in its open rack-scale approach. Since Helios was first introduced, AMD has steadily expanded the ecosystem through collaborations with HPE, Celestica, Supermicro, TCS, and AIC to manufacture, deploy, and scale the platform worldwide.
And instead of controlling the entire stack itself, the company is working with OEMs, ODMs, networking vendors, and cloud providers to build an open AI infrastructure platform based on OCP standards and Ultra Accelerator Link over Ethernet (UALoE). The goal is to make deploying massive AI clusters easier while avoiding vendor lock-in.
Team Red is said to provide more details at the upcoming AMD Advancing AI, where the company is expected to unveil additional AI hardware, software, and customer partnerships.
