← 返回全部时间线
TheElec(韩国电子产业媒体)原文发布 09-27 18:32Radar 收录 09-27 20:46查看原文 ↗

高通将1位AI模型引入最新骁龙可穿戴平台

Qualcomm Brings 1-Bit AI Models to Latest Snapdragon Wearable Platforms
中文摘要

高通宣布在其最新的骁龙可穿戴平台中加入对1位人工智能模型的支持,目标是在超小型设备上也能流畅运行端侧AI。Ziad Asghar。该平台面向可穿戴设备,属于芯片原厂的产品能力更新,与元器件供应链中的可穿戴芯片方案相关。

原文深度解读

本文报道高通在其最新面向可穿戴设备的骁龙平台中加入对1位人工智能(AI)模型的支持,目标是在内存极其有限的超小型设备上流畅运行端侧AI。涉及主体包括高通及其高管Ziad Asghar、高级产品经理Andrew Reister,以及合作方PrismML;时间点为当地时间9月23日。关键数字包括:相比8位模型内存需求可降至八分之一;若8位AI模型需要1GB RAM,1位模型可用125MB运行;部分智能眼镜此前配备4GB DRAM,近期产品使用2GB。与电子元器件供应链的关系在于,该平台属于芯片原厂面向可穿戴设备的产品能力更新,涉及骁龙AR1、AR1+和Snapdragon Sound Elite Gen 2等芯片方案,并可能影响可穿戴设备对DRAM等存储器件的需求与成本结构。

  • 高通在面向可穿戴设备的最新骁龙平台中加入对1位AI模型的支持,目标是在超小型设备上也能流畅运行端侧AI;涉及平台包括Snapdragon AR1、AR1+以及Snapdragon Sound Elite Gen 2。
  • 高通XR、可穿戴与个人AI高级副总裁兼总经理Ziad Asghar于当地时间9月23日表示,1位AI推理对内存极其有限的小型设备是关键技​​术,相比8位模型,内存需求可降至八分之一。
  • 原文解释AI模型位数代表精度:数据中心和高性能服务器通常使用16至32位浮点格式FP16和FP32,端侧AI模型则量化到约4至8位;位数下降会显著降低模型文件大小和内存(DRAM)需求,计算也可从复杂浮点乘法转向加减等更简单操作,从而提高处理速度并降低功耗。
  • 原文给出量化示例:如果8位AI模型需要1GB RAM,1位模型可用125MB运行;Asghar还表示,部分智能眼镜此前配备4GB DRAM,近期发布的产品使用2GB,高通也在进行优化以进一步降低需求,并预计该技术将显著降低可穿戴设备成本。
  • 原文指出1位AI的优势伴随取舍:量化会缩小表示权重的范围并产生误差,误差累积可能导致模型响应准确率下降、出错可能性增加。
  • 高通与PrismML合作以尽量减少准确率损失;Asghar表示,量化AI模型通常确实会导致一定准确率损失,但PrismML的专有技术即使在从4位模型转向1位模型时也能将准确率维持在很高水平。

供应链影响

  • 若1位AI模型在可穿戴平台中落地,端侧AI对DRAM容量和带宽的需求可能下降,从而可能影响可穿戴设备中存储器件规格与用量结构,但实际影响取决于终端产品设计与AI负载。
  • 原文提到智能眼镜DRAM从4GB降至2GB,并称高通正继续优化,这可能意味着未来可穿戴设备对高容量DRAM的依赖降低,但具体供应链影响取决于平台采用速度和整机厂商方案。
  • 由于1位模型可减少内存需求并可能降低可穿戴设备成本,芯片平台与存储器件之间的价值分配可能发生变化,但这一影响取决于量化精度保持能力、终端功能需求与市场竞争。
  • 高通与PrismML合作以维持1位模型准确率,表明端侧AI供应链可能更依赖模型量化与软件优化能力,而不仅是硬件规格提升,但相关生态影响取决于技术授权与平台兼容性。
  • 原文指出量化可能带来准确率下降和出错风险,因此1位AI在可穿戴设备中的实际采用节奏可能受限于应用对可靠性的要求,供应链影响取决于具体应用场景与误差容忍度。

系统解读,仅作为判断线索,不构成备货、出货、涨价或投资建议。

正文内容 · AI翻译

Ziad Asghar,高通高级副总裁兼XR、可穿戴与个人AI总经理,展示Snapdragon Sound Elite Gen 2。(照片:THE ELEC,Maui)

高通已在其最新的面向可穿戴设备的骁龙平台中加入对1位人工智能(AI)模型的支持,目标是在超小型设备上也能流畅运行端侧AI。

高通XR、可穿戴与个人AI高级副总裁兼总经理Ziad Asghar于当地时间9月23日表示,1位AI推理对于内存极其有限的小型设备而言是一项关键技​​术。他说:“与8位模型相比,内存需求可以降至八分之一。”

高通表示,其已在Snapdragon AR1和AR1+以及Snapdragon Sound Elite Gen 2中引入支持1位AI模型的架构。多模态AI模型是一种机器学习模型,可处理多种类型或模态的数据,包括文本、音频和图像。在智能眼镜上,这可能使摄像头能够直接识别用户前方的物体,并让设备通过音频在实时对话中对其进行描述。

AI模型的位数代表其精度。更高的位数使模型能够更精确地表示内部计算数据和权重。数据中心和高性能服务器中使用的AI模型通常采用16至32位浮点格式,即FP16和FP32。端侧AI模型则将这些模型压缩或量化到约4至8位。端侧AI在设备本地完成所有计算。

从左至右,高通高级产品经理Andrew Reister和高级副总裁Ziad Asghar参加问答环节。(照片:THE ELEC,Maui)

随着位数下降,AI模型的文件大小和内存(即DRAM)需求会急剧下降。计算也可以从复杂的浮点乘法转向加减法等更简单的操作,从而提高处理速度,同时降低功耗。如果8位AI模型需要1GB RAM,那么1位模型可以用125MB运行。

Asghar表示,一些智能眼镜此前配备4GB DRAM,而近期发布的产品使用2GB。他说:“我们也在进行优化,以进一步降低这一需求”,方式是使用1位AI模型。高通预计该技术将显著降低可穿戴设备的成本。

1位AI的优势也伴随取舍。量化会缩小用于表示权重的范围,从而产生误差。随着这些误差累积,模型响应准确率可能下降,出错的可能性也会增加。

高通与PrismML合作,以尽量减少准确率损失。Asghar说:“量化AI模型通常确实会导致一定准确率损失,这是事实。但PrismML的专有技术即使在从4位模型转向1位模型时,也能将准确率维持在非常高的水平。”

Ziad Asghar, Qualcomm senior vice president and general manager of XR, Wearables and Personal AI, presents Snapdragon Sound Elite Gen 2. (Photo: THE ELEC, Maui)

Qualcomm has added support for 1-bit artificial intelligence (AI) models to its latest Snapdragon platforms for wearables, aiming to run on-device AI smoothly even on ultra-compact devices.

Ziad Asghar, Qualcomm senior vice president and general manager of XR, Wearables and Personal AI, said on Sept. 23 local time that 1-bit AI inference is a critical technology for small devices with extremely limited memory. "Compared with an 8-bit model, memory requirements can be reduced to one-eighth," he said.

Qualcomm said it has introduced an architecture supporting 1-bit AI models in Snapdragon AR1 and AR1+ as well as Snapdragon Sound Elite Gen 2. A multimodal AI model is a machine-learning model that processes multiple types, or modalities, of data, including text, audio and images. On smart glasses, that could allow the camera to recognize an object directly in front of the user and the device to describe it through audio in a real-time conversation.

The bit count of an AI model represents its precision. Higher bit counts allow the model to represent internal computational data and weights more precisely. AI models used in data centers and high-performance servers typically use 16- to 32-bit floating-point formats, or FP16 and FP32. On-device AI models compress, or quantize, those models to around 4 to 8 bits. On-device AI performs all computation locally on the device.

From left, Qualcomm Senior Product Manager Andrew Reister and Senior Vice President Ziad Asghar take part in a question-and-answer session. (Photo: THE ELEC, Maui)

As the number of bits falls, the AI model's file size and memory, or DRAM, requirements decline sharply. Computation can also shift from complex floating-point multiplication to simpler operations such as addition and subtraction, increasing processing speed while reducing power consumption. If an 8-bit AI model requires 1GB of RAM, a 1-bit model can operate with 125MB.

Asghar said some smart glasses previously came with 4GB of DRAM, while recently released products use 2GB. "We are also working on optimizations to bring that down further" by using 1-bit AI models, he said. Qualcomm expects the technology to significantly reduce the cost of wearable devices.

The advantages of 1-bit AI come with trade-offs. Quantization narrows the range used to represent weights, creating errors. As those errors accumulate, model response accuracy can decline and errors can become more likely.

Qualcomm worked with PrismML to minimize the loss of accuracy. "It is true that quantizing an AI model typically results in some loss of accuracy," Asghar said. "But PrismML's proprietary technology can maintain accuracy at a very high level even when moving from a 4-bit model to a 1-bit model."

↑