AMD Launches Helios and Expands AI Portfolio

According to AMD's press release, each Helios configuration combines 72 Instinct MI455X GPUs with 18 EPYC Venice CPUs, along with Pensando networking and the ROCm software stack. AMD described Helios as its first rack-scale AI solution and stated that it is intended for deployment at gigawatt scale by leading AI companies.

AMD said Helios provides up to 30% more inference tokens per dollar than competing systems. Separately, Data Center Dynamics reported that AMD claims 15% higher peak FP4 performance than Nvidia's Vera Rubin platform and 50% more high-bandwidth memory capacity. These are vendor performance claims rather than independently verified benchmark results.

The MI455X is the accelerator inside Helios. Wccftech reported that AMD listed the chip at 40 PFLOPs of FP4 performance, 20 PFLOPs of FP8 performance, 432 GB of HBM4 memory, and 19.6 TB/s of memory bandwidth. System-level AI performance depends not only on accelerator throughput, but also on model architecture, networking, memory behavior, compilation, and serving software. For teams evaluating rack-scale platforms, reproducible workload-level benchmarks, power measurements, and software compatibility remain more decision-useful than peak specifications alone.

AMD estimated that the total compute market could reach approximately $2 trillion by 2030, Reuters reported, compared with an estimated $365 billion in 2025. Su told Reuters that AMD assigns $1.4 trillion of that 2030 estimate to AI-accelerating chips and $220 billion to CPUs.

The original report also states that AMD expects the global server CPU market to exceed $200 billion by 2030, up from about $25 billion currently. Wccftech reported that AMD presented a 46% revenue share in data-center CPUs and linked the forecast to demand for AI accelerators, CPUs, and rack-level systems.

These are AMD market estimates, not independent market forecasts. Still, the figures reflect a broader infrastructure shift: as inference and agentic workloads scale, the relevant deployment unit increasingly includes CPUs, accelerators, high-bandwidth memory, networking, storage, and orchestration software rather than a standalone GPU. Comparable industry transitions have made system integration and software maturity central factors in whether theoretical hardware performance translates into production throughput.

AMD's Helios announcement places ROCm alongside its silicon and networking components as part of the rack-level offering. That matters because porting, kernel performance, distributed-training behavior, and inference-serving support can determine the practical cost of moving workloads between hardware ecosystems.

Reuters characterized the launch as part of AMD's effort to gain ground against Nvidia in data-center AI chips, particularly inference. The most consequential near-term evidence for practitioners will be production deployments and independently reported results across common stacks such as PyTorch, distributed training frameworks, and high-throughput inference servers.