![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
vakСегодня на конференции Linley Spring Processor Conference был впервые представлен публике стартап SiMa.ai, где я имею честь трудиться. Вот тут можно почитать подробнее.
sima-machine-learning-moves-to-the-edge-wp.pdf
"SiMa.ai was founded in early 2019 to solve these problems. Sima means edge in Sanskrit, and the company targets edge applications that require high-performance vision processing without the high power and cost of GPU-based accelerators. To do so, the company has developed the MLSoC architecture, which encapsulates a single-chip solution for embedded systems that require machine-learning capability."
"The chip’s unique feature is the SiMa.ai ML accelerator (MLA). This custom block has enough performance to run complex neural networks, but it consumes much less power than traditional GPUs. The MLA employs a new architecture that shifts complexity from the hardware to the software, thus reducing power consumption. The design is optimized for the matrix computations that are common in many types of neural networks, improving efficiency over CPUs (which are optimized for scalar computation) and GPUs (vector computation). The MLA architecture can scale from 50 TOPS to 200 TOPS and beyond, enough for even the most demanding edge applications. The company is also developing a tool chain to simplify porting ML applications to the chip."
"The ML accelerator enables customers to add machine learning without breaking their power budget. In fact, the MLSoC is designed to deliver industry-leading power efficiency of 10 TOPS per watt. By comparison, leading GPUs offer about 1 TOPS per watt, and special-purpose accelerators can achieve as much as 5.2 TOPS/W.
This incredible efficiency translates directly to strong performance on real neural networks. For example, SiMa.ai expects the MLSoC to achieve 2,280 images per second (IPS) on ResNet-50 inference when running at a batch size of one (batch=1), which is typical for real-time video analysis. Nvidia rates its Xavier processor at 1,390 IPS on the same test. But while Xavier consumes 29W at this speed, the MLSoC will achieve its strong performance while using about 4W. In other words, the MLSoC is rated at 570 IPS/W, versus just 48 IPS/W for Nvidia’s best edge processor. The leading accelerators top out at 395 IPS/W."
sima-machine-learning-moves-to-the-edge-wp.pdf
"SiMa.ai was founded in early 2019 to solve these problems. Sima means edge in Sanskrit, and the company targets edge applications that require high-performance vision processing without the high power and cost of GPU-based accelerators. To do so, the company has developed the MLSoC architecture, which encapsulates a single-chip solution for embedded systems that require machine-learning capability."
"The chip’s unique feature is the SiMa.ai ML accelerator (MLA). This custom block has enough performance to run complex neural networks, but it consumes much less power than traditional GPUs. The MLA employs a new architecture that shifts complexity from the hardware to the software, thus reducing power consumption. The design is optimized for the matrix computations that are common in many types of neural networks, improving efficiency over CPUs (which are optimized for scalar computation) and GPUs (vector computation). The MLA architecture can scale from 50 TOPS to 200 TOPS and beyond, enough for even the most demanding edge applications. The company is also developing a tool chain to simplify porting ML applications to the chip."
"The ML accelerator enables customers to add machine learning without breaking their power budget. In fact, the MLSoC is designed to deliver industry-leading power efficiency of 10 TOPS per watt. By comparison, leading GPUs offer about 1 TOPS per watt, and special-purpose accelerators can achieve as much as 5.2 TOPS/W.
This incredible efficiency translates directly to strong performance on real neural networks. For example, SiMa.ai expects the MLSoC to achieve 2,280 images per second (IPS) on ResNet-50 inference when running at a batch size of one (batch=1), which is typical for real-time video analysis. Nvidia rates its Xavier processor at 1,390 IPS on the same test. But while Xavier consumes 29W at this speed, the MLSoC will achieve its strong performance while using about 4W. In other words, the MLSoC is rated at 570 IPS/W, versus just 48 IPS/W for Nvidia’s best edge processor. The leading accelerators top out at 395 IPS/W."
no subject
Date: 2020-04-08 06:41 (UTC)no subject
Date: 2020-04-08 14:47 (UTC)Угадай с первого раза, лозунг какой компании из 9 букв это мне напоминает.
no subject
Date: 2020-04-08 17:40 (UTC)no subject
Date: 2020-04-08 15:30 (UTC)Матрицы против векторов, оригинально.
no subject
Date: 2020-04-10 07:17 (UTC)По порядку величины - как у мозга ;-)