创新背景

微小边缘设备上的AI功能受到电池提供的能量的限制。因此，提高能源效率至关重要。在当今的人工智能芯片中，数据处理和数据存储发生在不同的地方——计算单元和存储单元。这些单元之间频繁的数据移动消耗了人工智能处理过程中的大部分能量，因此减少数据移动是解决能量问题的关键。

创新过程

斯坦福大学的工程师们提出了一个潜在的解决方案：一种新型的电阻随机存取存储器（RRAM）芯片，它在存储器本身内进行人工智能处理，从而消除了计算和存储单元之间的分离。他们的“内存计算”（compute in memory，CIM）芯片被称为Neuram，大约有指尖那么大，在有限的电池电量下比现有芯片能做的更多。

为了克服数据移动瓶颈，研究人员实现了所谓的内存计算（CIM），这是一种新的芯片架构，直接在内存中而不是在单独的计算单元中执行AI计算。Neuram使用的存储技术是电阻式随机存取存储器（RRAM）。它是一种非易失性存储器——即使断电也能保存数据的存储器——已经出现在商业产品中。RRAM可以在较小的面积内存储大型AI模型，并消耗很少的功率，使其非常适合小尺寸和低功耗的边缘设备。

尽管CIM芯片的概念已经确立，并且在RRAM中实现AI计算的想法并不新鲜，这是第一次将大量内存集成到神经网络芯片上，并通过硬件测量呈现所有基准结果。

Neuram的架构允许芯片以低功耗和紧凑的面积足迹执行模拟内存计算。它是与加州大学圣地亚哥分校的Gert Cauwenberghs实验室合作设计的，后者是低功耗神经形态硬件设计的先驱。该体系结构还支持数据流方向的可重构性，支持各种人工智能工作负载映射策略，并可以与不同类型的人工智能算法一起工作——所有这些都不会牺牲人工智能计算精度。

为了证明NeuRRAM人工智能能力的准确性，团队测试了它在不同任务中的功能。他们发现，MNIST数据集的字母识别准确率为99%，CIFAR-10数据集的图像分类准确率为85.7%，谷歌语音命令识别准确率84.7%，贝叶斯图像恢复任务的图像重建误差降低了70%。

效率、多功能性和准确性都是广泛采用该技术的重要方面，从硬件到软件的整个堆栈的协同优化是关键。NeuRRAM是一种物理概念验证，但在准备将其转化为实际的边缘设备之前，还需要更多的开发。

创新价值

这项工作为RRAM器件工程、编程模型和内存计算神经网络设计的未来研究开辟了若干途径，使这项技术可扩展并可供软件开发人员使用。如果成功，像Neuram这样的RRAM内存计算芯片几乎具有无限的潜力。它们可以嵌入农田，进行实时人工智能计算，以根据当前土壤条件调整灌溉系统。

创新关键点

Neuram使用的存储技术是电阻式随机存取存储器（RRAM）。它是一种非易失性存储器——即使断电也能保存数据的存储器——已经出现在商业产品中。RRAM可以在较小的面积内存储大型AI模型，并消耗很少的功率，使其非常适合小尺寸和低功耗的边缘设备。

Develop new chips to improve AI computing efficiency

Engineers at Stanford University have proposed a potential solution: a new type of resistive random access memory (RRAM) chip, which performs artificial intelligence processing in the memory itself, thus eliminating the separation between computing and memory cells. Their "compute in memory" (CIM) chip is called neuram, which is about the size of a fingertip and can do more than existing chips with limited battery power.
In order to overcome the bottleneck of data movement, researchers implemented the so-called in memory computing (CIM), which is a new chip architecture that performs AI computing directly in memory rather than in a separate computing unit. The storage technology used by neuram is resistive random access memory (RRAM). It is a kind of nonvolatile memory - a memory that can save data even when power is off - and has appeared in commercial products. RRAM can store large AI models in a small area and consume little power, making it very suitable for small-size and low-power edge devices.
Although the concept of CIM chip has been established, and the idea of implementing AI computing in RRAM is not new, this is the first time that a large amount of memory is integrated into a neural network chip and all benchmark results are presented through hardware measurement.
Neuram's architecture allows the chip to perform analog memory calculations with low power consumption and a compact area footprint. It was designed in collaboration with the GERT cauwenberghs laboratory at the University of California, San Diego, a pioneer in low-power neuromorphological hardware design. The architecture also supports the reconfigurability of data flow direction, supports various AI workload mapping strategies, and can work with different types of AI algorithms - all of which will not sacrifice the accuracy of AI calculation.
To demonstrate the accuracy of neurram's AI capabilities, the team tested its functions in different tasks. They found that the letter recognition accuracy of MNIST dataset was 99%, the image classification accuracy of cifar-10 dataset was 85.7%, the google voice command recognition accuracy was 84.7%, and the image reconstruction error of Bayesian image recovery task was reduced by 70%.
Efficiency, versatility and accuracy are all important aspects of widely adopting this technology, and the collaborative optimization of the entire stack from hardware to software is the key. Neurram is a physical proof of concept, but more development is needed before it is ready to be converted into an actual edge device.