Qualcomm teams up with Samsung and SK hynix to bet on HBC! A potential new solution for the “memory wall” in AI chips
Qualcomm is rallying Samsung Electronics and SK hynix to jointly advance the commercialization of High Bandwidth Computing (HBC). This means competition among AI chips is shifting from simply chasing raw compute performance to addressing how to move data and overcome memory bottlenecks more efficiently. Qualcomm has recently unveiled its HBC technology roadmap and is positioning it as a key technical direction for data-center AI infrastructure.
Based on the current division of labor, Qualcomm is mainly responsible for designing the compute chips and system integration, while Samsung Electronics and SK hynix handle DRAM stacking. They also plan to work with TSMC to integrate chip manufacturing and advanced packaging steps. The first-generation HBC products are expected to enter commercial shipments in 2027, with further upgrades to follow.
HBC’s biggest selling point is that it aims to tackle the increasingly evident “memory wall” in traditional AI architectures. In simple terms, AI models are getting larger, and GPU compute speeds are improving—but memory bandwidth and data transfer efficiency are emerging as the new bottlenecks. Qualcomm’s HBC approach, by vertically stacking low-power DRAM and integrating it more closely with compute chips, further reduces the distance over which data must travel.
According to data published by Qualcomm, at the full accelerator level, HBC’s unit power bandwidth can reach up to 6× that of HBM, and its unit power storage capacity can reach up to 200× that of SRAM. However, these are technical targets disclosed by the vendors; whether they can truly be deployed at scale in data centers will still depend on real product validation, cost, and ecosystem compatibility.
This development is especially worth watching for Samsung and SK hynix. The two companies are not only continuing to bet on HBM, but also participating in the new HBC roadmap—essentially pursuing both a traditional high-bandwidth memory + next-generation compute-storage fusion architecture at the same time.
In the race for AI compute power, it’s not just about how fast GPUs are anymore—it’s about who can connect “compute” and “memory” more efficiently. HBM is still ramping up production, $HBCP.US has already started grabbing the next opportunity, and the memory war for AI chips may only be entering its next phase.
$SKHYNIX $SKHY
Qualcomm is rallying Samsung Electronics and SK hynix to jointly advance the commercialization of High Bandwidth Computing (HBC). This means competition among AI chips is shifting from simply chasing raw compute performance to addressing how to move data and overcome memory bottlenecks more efficiently. Qualcomm has recently unveiled its HBC technology roadmap and is positioning it as a key technical direction for data-center AI infrastructure.
Based on the current division of labor, Qualcomm is mainly responsible for designing the compute chips and system integration, while Samsung Electronics and SK hynix handle DRAM stacking. They also plan to work with TSMC to integrate chip manufacturing and advanced packaging steps. The first-generation HBC products are expected to enter commercial shipments in 2027, with further upgrades to follow.
HBC’s biggest selling point is that it aims to tackle the increasingly evident “memory wall” in traditional AI architectures. In simple terms, AI models are getting larger, and GPU compute speeds are improving—but memory bandwidth and data transfer efficiency are emerging as the new bottlenecks. Qualcomm’s HBC approach, by vertically stacking low-power DRAM and integrating it more closely with compute chips, further reduces the distance over which data must travel.
According to data published by Qualcomm, at the full accelerator level, HBC’s unit power bandwidth can reach up to 6× that of HBM, and its unit power storage capacity can reach up to 200× that of SRAM. However, these are technical targets disclosed by the vendors; whether they can truly be deployed at scale in data centers will still depend on real product validation, cost, and ecosystem compatibility.
This development is especially worth watching for Samsung and SK hynix. The two companies are not only continuing to bet on HBM, but also participating in the new HBC roadmap—essentially pursuing both a traditional high-bandwidth memory + next-generation compute-storage fusion architecture at the same time.
In the race for AI compute power, it’s not just about how fast GPUs are anymore—it’s about who can connect “compute” and “memory” more efficiently. HBM is still ramping up production, $HBCP.US has already started grabbing the next opportunity, and the memory war for AI chips may only be entering its next phase.
$SKHYNIX $SKHY
