As Moore’s Law slows, superior packaging has change into the crucial lever driving breakthroughs in AI chip efficiency, in response to DIGITIMES chief semiconductor analyst Tony Huang. Speaking with DIGITIMES Asia, Huang emphasised that heterogeneous integration is now as pivotal to system efficiency as transistor scaling as soon as was.
Without superior packaging, transferring from 28nm to A16 would ship roughly an 80-fold improve in compute density per reticle space, Huang stated. With it, the achieve rises to round 320 occasions.
Bottlenecks driving innovation
He defined that AI workloads have uncovered a number of technical bottlenecks—what he calls the brand new “walls”: the reminiscence wall, addressed via high-bandwidth reminiscence (HBM); the I/O and communication partitions, tackled by decreasing latency via chiplet and interposer applied sciences; the ability and thermal partitions, mitigated via energy effectivity and thermal administration; and the yield and fabrication partitions, eased by smaller chiplets with larger yields.
Technologies akin to 3D SoC stacking, hybrid bonding, and co-packaged optics (CPO) at the moment are key enablers, he added.
Credit: Dgitimes
Market progress outpacing the semiconductor trade
Huang forecasts that superior packaging for AI knowledge heart chips will develop at a compound annual fee of 45.5% from 2024 to 2030, far outpacing the general semiconductor trade’s 8.7% and the broader packaging and testing sector’s 9.5%.
“2.5D CoWoS-type solutions still dominate and will account for more than half of the total market,” he stated. “Their share will gradually decline from 69% to 58% as 3D and hybrid approaches mature.”
He famous that the trade is transitioning from 2.5D to three.5D packaging, citing AMD’s MI300X AI accelerator, which mixes 3D SoIC and a couple of.5D CoWoS and is manufactured by TSMC. Broadcom can be creating a 3.5D accelerator utilizing a face-to-face construction, he stated.
Taiwan’s dominant place
“Taiwanese companies accounted for about 77% of the global data center AI packaging market in 2024,” Huang stated. “Even with intensifying competition from Intel, Samsung, and new US and Chinese ecosystems, Taiwan’s share is expected to stay near 70% by 2030.”
If HBM meeting—nonetheless primarily carried out in South Korea—is excluded, Taiwan’s market share may attain 99%, he added.
Defining the panorama
Huang famous that DIGITIMES defines knowledge heart AI chips in three classes: AI server CPUs; AI accelerators, together with GPUs from NVIDIA and AMD, and customized AI chips like Google’s…
