IBM unveils Power9 chip and systems built for AI and machine learning, claims an increase for workloads on common AI frameworks by up to almost 4x
Ron Miller / TechCrunch :
Context & Ripple Effects
IBM's Power9 launch is the silicon payoff to a claim it made just months earlier, when it touted deep learning performance scaling across 64 servers at 95% efficiency — Power9 is the processor generation meant to deliver that kind of throughput on common AI frameworks like TensorFlow. The company is positioning its Power systems line as an AI platform rather than general-purpose servers.
The move also sets the template for the line's future: Power10 followed in 2020 as IBM's first commercial 7nm processor aimed at big data and AI, and the research pipeline (a 2nm test chip, then the memory-light NorthPole prototype) shows IBM treating AI silicon as a continuous program rather than a one-off.
First-order effects
- Enterprises running common AI frameworks on Power systems get a claimed workload speedup of up to almost 4x, giving IBM's server business an immediate performance argument for AI buyers choosing between platforms.
- IBM's Power line gains a distinct identity as AI-specific hardware, competing directly for machine learning deployments that would otherwise default to GPU-based systems.
Second-order effects
- Rival chipmakers are pushed into the same per-generation multiplier contest — AMD's later MI350X launch claiming up to 4x AI compute over its prior generation shows the benchmark-inflation pattern Power9 helped establish.
- AI framework vendors gain leverage, since 'up to 4x on common frameworks' claims make optimization for those frameworks a selling point that hardware vendors compete over.
Third-order effects
- If the cadence holds — Power10 in 2020, Power11 arriving in 2025 with ransomware-detection features — processor generations become the marketing rhythm of enterprise AI infrastructure, with each vendor obligated to publish a headline multiplier per cycle.
- The longer arc points toward specialized AI silicon diverging from general-purpose CPUs, with IBM hedging across both: Power systems for deployed AI workloads and research chips like NorthPole exploring architectures that sidestep external memory entirely.
The trend: Enterprise AI is turning processor launches into a recurring benchmark arms race, where CPU and GPU vendors alike stake their relevance on claimed multipliers for common AI frameworks.