IT之家 9 月 23 日消息,夏威夷时间 9 月 22 日 9 点(北京时间 9 月 23 日 3 点)召开的 2026 骁龙峰会上,IT之家带来前方现场报道,高通宣布将会向骁龙芯片下放 HBC 工艺。
IT之家注:HBC 全称为 High Bandwidth Compute,直译为“高带宽计算”,于 2026 年 6 月召开的投资者日上公开,最初面向数据中心机型,在本次活动中宣布将来将会应用于骁龙芯片。
该架构是一种近内存计算(near-memory computing)架构,将计算芯片放置在堆叠式 LPDDR DRAM 下方,借助 3D 集成和硅通孔(TSV)缩短数据传输距离,目标是缓解 AI 推理中的“内存墙”难题。
对于普通使用者而言,高通骁龙芯片将来如果整合 HBC 架构,手机可以更快、更省电地运行更大的端侧 AI 模型,而且可以在本地运行更多 AI 能力。
当下限制手机端 AI 的瓶颈,不再仅是 NPU 的乘加计算能力,还考验模型参数和中间数据在内存与处理器之间搬运的速度。HBC 借助把部分计算放到内存附近,减少数据来回移动,有望降低延迟和内存功耗。
高带宽内存本身不等于更大容量,但如果 HBC 同时提高有效带宽、减少数据搬运,并改善每瓦表现,手机就有机会运行参数量更大、上下文更长的模型。
2026 骁龙峰会专题
本文来源:https://www.ithome.com/1/006/003.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
On September 23, IT Home reported that at the 2026 Snapdragon Summit held at 9:00 a.m. Hawaii time on September 22 (3:00 a.m. Beijing time on September 23), IT Home brought live coverage from the scene, and Qualcomm announced that it would bring HBC technology to Snapdragon chips.
IT Home note: HBC stands for High Bandwidth Compute, literally translated as "high-bandwidth computing." It was unveiled at the Investor Day held in June 2026, initially for data center models, and at this event it was announced that it would be applied to Snapdragon chips in the future.
This architecture is a near-memory computing architecture that places compute chips beneath stacked LPDDR DRAM, using 3D integration and through-silicon vias (TSV) to shorten data transmission distances, aiming to alleviate the "memory wall" bottleneck in AI inference.
For ordinary users, if Qualcomm Snapdragon chips integrate the HBC architecture in the future, phones will be able to run larger on-device AI models faster and more power-efficiently, and also run more AI capabilities locally.
The current bottleneck limiting on-device AI is no longer just the NPU's multiply-accumulate computing power, but also the speed at which model parameters and intermediate data are moved between memory and the processor. By placing some computing near memory, HBC reduces data back-and-forth movement, which is expected to lower latency and memory power consumption.
High-bandwidth memory itself does not mean larger capacity, but if HBC simultaneously improves effective bandwidth, reduces data movement, and improves performance per watt, phones could have the opportunity to run models with more parameters and longer context.
2026 骁龙峰会专题
