Experts say chiplet-based designs, which let AMD make a full suite of server chips to compete with Intel, will quickly become a dominant form of chip design
the way their executives tell it. But the approach is getting deployed across the whole industry and has big potential to shake up how the chips of the future work. https://www.protocol.com/...
Context & Ripple Effects
The chiplet idea has been building since 2018, when chip makers began splitting processors into multiple small pieces of silicon to keep up with Moore's Law as a response to slowing single-die scaling. AMD turned it into a server strategy: CTO Mark Papermaster described combining 7nm CPU dies with cheaper 14nm IO dies in EPYC a mix-and-match approach that let the company field a full server line against Intel despite a fraction of its resources.
What changed this week is the verdict: experts now expect chiplet-based designs to become the dominant form of chip design industry-wide, not just an AMD trick. The approach is already spreading beyond CPUs — AMD's upcoming RDNA 3 GPU architecture uses chiplets and targets a 50%+ performance-per-watt uplift over RDNA 2 extending the method to graphics — which is why the technique is being read as a structural shift rather than one company's cost hack.
First-order effects
- AMD gains the ability to assemble a complete server chip portfolio from reusable dies, sustaining the EPYC momentum it promised when it committed to rapidly ramping production after launch ramping EPYC output quickly.
- Intel faces a competitor whose design economics no longer require matching its monolithic die scale — AMD can refresh compute cores on new nodes while leaving IO dies on older, cheaper ones.
Second-order effects
- Cloud buyers, who have long pushed for custom silicon — Intel was selling majority-custom chips into public clouds as far back as 2015 over half its cloud chips were custom designs — get more vendors able to tailor mixes of dies to specific workloads at lower cost.
- NVIDIA's server position is exposed on a second front: AMD already aimed Radeon Instinct accelerators at the machine-learning server market its first ML accelerator line, and chiplet economics make it cheaper to iterate GPU variants against NVIDIA's roadmap.
Third-order effects
- If chiplets become the default, competitive advantage migrates from owning the largest integrated fab process to packaging, interconnect, and die-mixing know-how — lowering the barrier for more companies to field credible server silicon.
- The monolithic processor as the industry's basic unit of design gives way to modular assemblies, changing how performance, cost, and supply relationships are structured across the semiconductor supply chain.
The trend: Chip design is shifting from monolithic processors to modular chiplet assemblies, with AMD's server and GPU roadmaps proving the model and the rest of the industry following.