In the keynote of the entrepreneur company NGD, we saw the prototype concept of their and Microsoft Research Institute. The concept based on in-suite processing integrates a calculated acceleration unit in storage control. This concept and most of the schemes are similar to the speed of acceleration, using the advantages of comparing the data stored more than the data store to achieve the calculated OFFLAODING, thereby reducing delay.
The application of NGD acceleration is currently a hot CNN, similar to Image, face recognition, and EDGE ends such as vehicle license identification. Unlike the Google TPU POD in the form of the cloud data center, the application in the EDGE end is more stringent for real-time requirements. Before 5G appears, the path from the terminal device sends raw data to the cloud processing or too long, so the EDGE calculation and The so-called fog calculation is still very powerful. The advantage of using the EDGE end is mainly: 1. Delayed end, 2. Reduces the amount of data, no raw data, feature data is enough.
On FMS 2018, similar solutions also have Marvell, Marvell announced a conceptual chip that combines SSD controllers and NVIDIA's NVDLA, pushed a device that can do deeply study in devices.
NVDLA information is here http://nvdla.org/primer.html, when you are released in 2017, everyone exclaiming a lot of AI's startups, because NVDLA uses NVIDIA to study in deep learning. Technical advantages, open source such a Deep Learning Accelerator design, most investment in NVIDIA is training market, for reasoning, need to adapt through new hardware and architectures. From the perspective of the occupation of the market, open source accelerator is important for increasing the right to speak.
The specific information of this concept chip seems to have not seen this concept chip on the booth of Marvell. Just a news from FMS2018 to see some information. At present, the field of this AI SSD focus is a data analysis market, and the acceleration function of AI SSD can be used in the application of the specific cloud or EDGE end large data analysis and data labeling. As can be seen from the internal block diagram of NVDLA, NVDLA is only an image application for CNN.
Not MARVELL, many enterprise-level controller manufacturers are considering this direction; Netint.ca from PMC Canada Class also shows the SSD controller of their PCIe 4.0 in FMS.
A VIDEO processing unit is integrated in SSD control. For HD H.265, think that the current mainstream scheme is still the CPU, so the single Intel's X86 server level scheme is between 4-6, while using hardware acceleration can be on performance and power consumption Winning.
At the Wuhan Storage Semiconductor Summit, the National Science of A-share listed company also proposed the concept of Near Data computing, how to reflect the optimization of Ai Workload in SSD design. At present, in my Tianchao, AI has the most hot, and the leaders of all levels are most important to security, face recognition, behavioral judgment, numerous head, and so on are technical means for stability unity. The technical challenge is also a lot.
As a storage big data analysis acceleration, you can't talk about BlueDBM. The Koreans of this project graduated from MIT, to UIC IRVIEN, or talking to the big data analysis acceleration, adding a lot of content on the original framework.
It has been calculated and the calculation of Vertex, the next step is, and those who participated in the machine to get started should be unfamiliar. It is recommended that you have technical attention to this star project. The previous studies supported before MIT, and now I don't know if Samsung is supported.
Similarly, the calculation of the previous four speeches is similar. We can see an obvious trend is to increase support for Near Data computing in the SSD controller, integrating some relatively mature solutions, which may be an enterprise SSD controller. New Direction. Finally, use a Korean SSD controller as the end of FMS 2018. They are the first to use RISC-V to do SSD controllers, and their performance and architecture have a great difference between and traditional controllers. It is worth learning. Read the full article, original text title: FMS 2018: Do Ai in storage
Article Source: [Micro Signal: SSDFANS, WeChat public number: SSDFANS] Welcome to add attention! Please indicate the source of the article.
Our other product:
