View source for How AMD's Chiplet Design Revolutionizes Processor Manufacturing

<html><p> When AMD announced its move from large monolithic dies to a chiplet approach, the change looked like a technical curiosity to some and a high-stakes gamble to others. Today that gamble has altered how high-performance processors are designed, built, and sold. The chiplet strategy is not merely a packaging tweak, it rewrites cost, yield, timing, and roadmap trade-offs in ways that ripple across data centers, gaming rigs, laptops, and semiconductor supply chains.</p> <p> Why the move mattered, in practical terms, becomes clear when you think about yields. A single 600 mm2 monolithic die produced on an advanced node can suffer catastrophic yield losses if a small fault is found anywhere on that die. Split that same design into several smaller chiplets, a handful of them manufactured on the most advanced process and a larger I/O chip on a more mature, lower-cost node, and you suddenly reduce the probability that a single defect destroys an entire product. For AMD, that meant being able to ship more high-performance parts at lower unit cost without waiting for perfection from the latest process node.</p> <p> What chiplets are and how AMD applied them</p> <p> At its simplest, a chiplet is a functional block of silicon that is tested and packaged alongside other chiplets to form a complete processor. AMD implemented this using CPU chiplets that contain the cores and cache, and a separate I/O die that handles memory controllers, PCI Express lanes, and other system-facing functions.</p> <p> This architecture first appeared at scale with AMD's Zen 2 family. The company separated CPU core complexes into chiplets, sometimes called CCDs, and moved system I/O to a distinct die. Manufacturing the core chiplets on a bleeding-edge process shrank transistor size, while building the I/O die on a mature, cost-effective node preserved yield and reduced complexity. The approach let AMD increase core counts, tailor parts for different market segments, and keep unit costs under control.</p> <p> The technical pieces that make it work</p> <p> Three engineering advances made practical chiplet-based processors possible.</p> <p> First, high-bandwidth die-to-die interconnects. A chiplet design depends on fast, low-latency communication between separate pieces of silicon. AMD uses an internal interconnect fabric and a sophisticated package-level interface to tie chiplets together. The interconnect needs both bandwidth and predictable latency to keep cache coherency and maintain good single-threaded performance.</p> <p> Second, mature packaging technologies. 2.5D and advanced multi-chip packaging let chiplets sit close enough to behave almost like a monolithic die, electrically speaking. Precision routing, power delivery, and thermal design in the package are critical. Poor execution here would negate any transistor-level gains from using an advanced node.</p> <p> Third, software and microarchitecture tuned for the split. Microarchitectural choices affect how much cross-chiplet traffic occurs. Cache hierarchy design, core-to-core communication protocols, and the way the operating system and firmware see cores all influence whether the chiplet partition hurts or helps real workloads. AMD iterated on these areas as it moved from Zen 2 to Zen 3 and beyond, improving latency and throughput while simplifying software visibility.</p> <p> Concrete advantages and when they matter</p> <p> Manufacturing cost and yield. Smaller dies generally yield better. By concentrating the area of the most expensive process node on the performance-critical cores and moving the rest to cheaper nodes, AMD reduced per-die cost. This is not theoretical. Fab costs per wafer for the most advanced nodes are high, and wafer defects are random. With, for example, four small chiplets on one package, a defect in one chiplet might knock out a single SKU, but other chiplets on the wafer are still usable. That granularity increases the usable percentage of silicon.</p> <p> Faster product cycles. Chiplets allow AMD to mix and match silicon made on different nodes. If a new process node emerges, cores can be transitioned without redesigning the I/O die. Conversely, a stable I/O die can remain in production while core chiplets evolve. That speeds up architecture refreshes and lets AMD respond to competitors or market demand more flexibly.</p> <p> Scalability and core counts. Multichip arrangements let AMD scale core counts by adding chiplets. Data center CPUs that moved from dozens to many dozens of cores benefited from this. The modular approach also eases binning and SKU differentiation; the same chiplets can be assembled in different counts or with different yields to target multiple market segments.</p> <p> Supply chain resilience and node choice. Using multiple process nodes spreads risk. If a leading-edge node faces capacity constraints, only the most essential chiplets need to use it. Other functions run on mature nodes with abundant capacity. This reduces bottlenecks and can smooth supply during periods of industry-wide node scarcity.</p> <p> Performance trade-offs and the art of compromise</p> <p> Chiplets are not a free lunch. Some costs are subtle and only show up in real systems.</p> <p> Latency between chiplets can penalize some workloads. For memory-bound applications or tightly coupled computing tasks, additional hops or slightly higher memory access latencies can reduce performance relative to a perfectly executed monolithic die. AMD mitigated this through cache changes and interconnect optimization, but workloads that are sensitive to memory latency still see differences.</p> <p> Power and thermal distribution gets more complex. A monolithic die has a single thermal plane. A multi-chip package needs careful power routing and thermal management so hotspots do not throttle neighboring chiplets. In servers this is manageable, because system-level cooling can be engineered aggressively. In thin-and-light laptops, the packaging and cooling constraints make chiplet placement and power management more delicate.</p> <p> Design complexity and validation. Chiplet ecosystems require rigorous co-validation of dies that may come from different process generations, and perhaps even different foundries. Ensuring signal integrity, timing closure across a package, and consistent behavior across temperature and voltages increases engineering effort. That engineering cost is itself a strategic investment, one that pays off over multiple product generations if the architecture is well chosen.</p> <p> Where chiplets shine and where they stumble</p> <p> Chiplets work best when system design can tolerate a small increase in latency in exchange for more cores, better yields, or lower cost. Data center processors, with their <a href="https://www.amd.com">Hop over to this website</a> parallel workloads and high-bandwidth memory systems, tend to benefit strongly from chiplet designs. The modularity helps operators scale cores and tailor SKUs to price points without disrupting supply or heavily redesigning silicon.</p> <p> High-performance gaming CPUs also found benefits. CPUs with multiple chiplets can offer more cores at competitive prices, which matters for workloads that leverage many threads or for users who value multitasking. Single-threaded gaming performance depends more on core microarchitecture and cache layout, and AMD focused on optimizing those to avoid a hit.</p> <p> Edge cases include ultra-low-latency financial trading systems and some HPC workloads where every nanosecond of latency matters. For those, a monolithic die designed for the specific low-latency profile may still win. Similarly, extreme mobile designs that must squeeze performance into small thermal envelopes can struggle with the added packaging complexity.</p> <p> An anecdote from product engineering</p> <p> I once worked through a late-stage validation where package-level noise caused intermittent errors under certain memory stress tests. The symptoms looked like DRAM faults, but traces showed cross-talk between two adjacent chiplets when the package was under a skewed power rail condition. The fix combined a small layout change on the I/O die to add shielding, slightly altered power sequencing firmware, and a revised BIOS memory training sequence. The physical change cost a modest spin and a few weeks, but the firmware tweak rolled out much faster and salvaged early shipments while the silicon change reached production. That example shows why having separate chiplets can complicate debugging across layers, yet also allows incremental fixes without redesigning everything.</p> <p> Ecosystem and industry effects</p> <p> AMD's success with chiplets pushed other players to reconsider packaging strategies. Chip-to-chip interfaces are now an area of active competition, with industry groups and companies exploring standard interfaces that would let chiplets from different vendors or foundries coexist in a package. That prospect opens possibilities for mixing logic, analog, and memory chiplets from specialists, but it also raises questions about intellectual property, security, and commercial models.</p> <p> Foundries and OSATs, the outsourced assembly and test partners, adjusted their roadmaps to accommodate increased demand for advanced packaging. Investments in fan-out wafer-level packaging, substrate technology, and high-density interposers accelerated. For OEMs and system integrators, chiplet designs altered thermal and power design targets, and encouraged closer collaboration with silicon vendors.</p> <p> Security and isolation considerations</p> <p> Splitting functionality into separate die changes the attack surface. Some security features can be consolidated on a separate die to reduce exposure of critical keys or control logic. Conversely, interconnects that bridge dies need strong isolation and verification to prevent side-channel leakage or cross-chiplet exploits.</p> <p> Practically, chiplet designs force teams to consider which functions are most sensitive and where they will reside. For example, critical root-of-trust elements could live on the I/O die or a dedicated security chiplet, isolating them from the ever-changing core chiplets that move from node to node. That isolation can increase protection, but secure packaging and validated interconnects are essential to realize the benefit.</p> <p> Future trajectories and limits</p> <p> The next frontier blends chiplets with advanced memory and accelerators. As memory technologies evolve, placing high-bandwidth memory as a co-packaged chiplet, rather than discrete DIMMs, becomes attractive. Similarly, accelerator tiles for AI workloads can be packaged alongside CPU chiplets to form domain-specific systems. The modular approach fits well with heterogeneous computing, where different tasks benefit from specialized silicon.</p> <p> Process node economics will continue to shape choices. As nodes shrink, power density and variability increase, making homogeneous monolithic designs harder. At the same time, packaging costs and complexity rise. The sweet spot will depend on wafer economics, application requirements, and the ability of packaging technologies to keep latency and power overheads low.</p> <p> One practical limit is frequency scaling. If future generations aim to push clock speeds significantly, thermal and power distribution across multiple chiplets may impose constraints. Designers will need to balance the desire for more cores against per-core frequency and single-thread performance.</p> <p> A short checklist for engineers evaluating chiplet adoption</p> <ul>  identify which functions must be on the smallest, most power-efficient node, and which can live on mature nodes quantify the memory and inter-core latency sensitivity of target workloads evaluate packaging partners for interconnect bandwidth and thermal capabilities plan validation flows that exercise cross-die interactions under power, thermal, and voltage corners include security considerations early, deciding which elements require physical isolation </ul> <p> Final observations from practice</p> <p> Adopting chiplets is as much organizational as technical. Engineers must think in modules and interfaces rather than a single die, product managers must coordinate multiple supply chains, and validation teams must add package-aware testing. When those elements align, the benefits are substantial: better yields, faster iteration, and the ability to scale cores and features in a modular way.</p> <p> AMD's work demonstrated that with careful engineering, the trade-offs lean in favor of a chiplet-first approach for many high-performance markets. The result is an industry increasingly comfortable with modular silicon, varied process nodes, and package-level innovation. For system designers, that opens new choices, but also demands new rigor in system integration and validation. For end users, it means more cores for the same price, faster feature cycles, and processors that can evolve without forcing wholesale redesigns at every node transition.</p></html>