In the field of artificial intelligence, training an advanced machine learning model requires a lot of computing resources. With the machine learning algorithm more and more applications in various fields and show superior performance, for the machine learning algorithm professional hardware needs, also become more and more intense.
In 2016, Google announced for the first time specifically for the acceleration of deep neural network computing power and research and development of the chip, in the calculation of performance and energy consumption indicators, TPU performance is far superior to traditional CPU, GPU combination. (We have also published an article last month, analyze the reasons behind the TPU dazzling results)
In the morning of May 19 at the Google I / O 2017 conference, Google officially released the second generation TPU. New chip compared to the first generation of productsPerformance, application, serviceOnce again to achieve a breakthrough.
First, the performance aspects
The new generation TPU can be applied to both high performance computing and floating point computing. And can achieve up to 180 trillion times per second floating-point performance. In contrast, last week, NVIDIA has just launched the GPU Tesla2 V100, only 120 trillion floating-point operations per second.
Compared to the first generation in the function to achieve a breakthrough from scratch, the starting point of the second generation is relatively higher, the development team can also focus on resources to improve the performance of improved TPU. I believe that through the hardware, software optimization, follow-up third generation, the fourth generation in the performance of the possibility of continuous breakthrough is very large.
Second, the application side
The first generation TPU does not specifically mention the combined application, cluster application function, and its own no storage space. The second generation at the conference directly on the display of a 64-generation TPU chip TPU pod operation array. This array of operations can provide up to 11.5 thousand teraflops of floating point computing capacity for a single ML training task, greatly accelerating the training of machine learning models.
There are also professional media mentioned that the new TPU in the left and right sides have four external interface, the left side of the additional two additional interfaces. These interfaces may allow the TPU chip to connect directly to memory in the future, or to a high-speed network, enabling more complex operations. In theory, developers can also design more features on this basis, add more extensions.
Third, the service side
From the name of the Cloud TPU above, you can also intuitively understand that a new generation of TPU will join the Google cloud computing platform, and external cloud services. This means that TPU is no longer just an exclusive service within Google, and will become anyone can easily share, the application of artifacts.
Here to see the place where Google compared thieves, and the first generation just came out, hidden tucked, but also particularly low-key to say only their own internal use. Second generation research and development, directly on the cloud: the hardware is not external sales, service can ah.
If it is direct hardware sales, many small and medium companies (such as our company: wisdom think) may be more inclined to use the GPU: a wider range of applications, according to the need to arrange to deal with different tasks. Large companies (such as Ali, Facebook), ML task is very large, will consider the procurement of TPU, or even out of funds, efficiency considerations, independent team to carry out research and development (such as the Google itself to do TPU, Facebook also had independent research and development Data center hardware device reports);
External sales service, then the first is to activate the demand for small and medium enterprises in this market, users only need to pay according to the use of time, cost savings. In addition, the large enterprise's own resource utilization has also been greatly improved, diluted the cost. For example, Ali, in order to deal with the massive server prepared by the two, in the spare time can provide cloud computing services. Finally, to avoid a direct conflict of interest with hardware vendors (such as NVIDIA).
In general, TPU, especially the Cloud TPU to large Internet companies confirmed the feasibility of independent research and development of hardware.
Performance, for the autonomy of business optimization, saving hardware procurement, data center construction, time consumption and other costs; applications, can be flexible with existing facilities, equipment, portfolio, expansion; services, through the cloud to achieve external resources Sell, earn income.
Corresponding, according to the different business of their own business, the future may be customized hardware equipment will be different. Such as APU for online transaction data processing For online social interaction FPU?
As for how to evaluate the Cloud TPU, probably can be marked as the era of artificial intelligence professional hardware come!
Copyright Notice:Wisdom big dataEdit order