AMD’s New GPUs: 30% More Tokens per Dollar vs Nvidia

At its Advancing AI event in San Francisco, AMD fired a warning shot at Nvidia by unveiling next-gen Instinct MI300X GPUs and Helios compute racks. CEO Lisa Su claimed a 30% superior tokens-per-dollar metric for LLM inference compared to Nvidia’s H100, positioning AMD as a fierce challenger in the booming AI data center market. This aggressive pricing move aims to lure hyperscalers seeking cost-effective alternatives to Nvidia’s dominant CUDA ecosystem. The announcement heats up the battle for the $100-billion AI accelerator segment, with both companies racing to deliver higher performance at lower cost.

AMD introduced the Instinct MI300X GPU and Helios rack systems, claiming a 30% tokens-per-dollar advantage over Nvidia’s H100 in LLM inference. The Helios architecture combines eight MI300X accelerators with AMD’s EPYC CPUs and high-speed Infinity Fabric links, targeting hyperscale AI training and inference. AMD’s benchmarks show that for running models like Llama 2 70B, the MI300X platform delivers more inferences per dollar—a crucial metric as cloud providers grapple with soaring AI compute costs. CEO Lisa Su emphasized open ecosystem support, contrasting with Nvidia’s proprietary CUDA lock-in. The move pressures Nvidia to accelerate its H200 and B100 roadmaps as AMD aims to capture a larger slice of the accelerator market by 2025.

The surge in high-end GPU deployments has significant implications for PCB fabrication. Each MI300X module uses advanced IC substrates with ultra-fine line widths and embedded passives, requiring HDI any-layer techniques. Helios rack backplanes and power boards demand heavy copper PCBs (up to 20 oz) to handle 700W+ per GPU. As AMD shifts to chiplet-based designs, the need for large-format, high-layer-count rigid-flex PCBs with strict impedance control grows. Manufacturers with capabilities in HDI, IC substrates, and ceramic PCBs for thermal management are well-positioned to support this transition. Substrate-like PCBs enable the dense routing required for next-gen AI accelerators, ensuring signal integrity in high-speed interconnects.

LT CIRCUIT provides high-reliability PCBs for AI accelerators and server infrastructure, including HDI any-layer boards, IC substrates, and heavy copper PCBs. As GPU vendors ramp production, its ceramic and rigid-flex solutions help manage thermal and signal integrity challenges in dense compute racks.