Wave Computing made a lot of progress in 20 18, and launched the first data stream processing unit, acquired MIPS, created MIPS Open, and delivered the first batch of systems to a few customers. Although the Wave architecture has some very interesting functions, we expect large-scale real experience feedback from users.
Wave is not an accelerator plugged into the server, it is an independent processor for graphics computing. This method has advantages and disadvantages. On the positive side, Wave will not be affected by the memory bottleneck faced by accelerators such as GPU. On the negative side, installing Wave equipment will be a brand-new upgrade, which needs to completely replace the traditional X86 server and make it a competitor of all server manufacturers.
I don't think Wave can beat NVIDIA at any point, but the architecture is very well designed, and the company has indicated that there will be customer feedback soon.
Figure 1: Wave is a system built from the 4-node "DPU" shown above. Fluctuation calculation
Chart core
Graphcore is a British unicorn startup with abundant funds (financing 3 1 billion dollars, currently valued at10.7 billion dollars) and teams all over the world. It is building a new graphics processor architecture, and its memory and logic unit are located on the same chip, which should be able to achieve higher performance. The release date of the team's products is not clear for the time being, but they said in April last year that they were "almost ready to release", and the latest information in 5438+February indicated that production would start soon.
Graphcore's list of investors is impressive, including Sequoia Capital, BMW, Microsoft, Bosch and Dell Technology.
I learned about the structure of the company and was very impressed. From edge devices to "colossus" dual-chip packaging, it is used for training and reasoning in data centers. At the recent NeurIPS event, Graphcore demonstrated its RackScale IPU Pod, which provided more computing power than 16 petaflops in a 32-server rack. Although the company often claims that it will provide 100 times higher performance than the best GPU in its class.
According to Graphcore, the 4 "colossus" GC2 (8-chip) server can provide mixed precision performance of 500 TFlops (trillions of operations per second). A single NVIDIA V 100 can provide 125 TFlops, so theoretically four V 100 can provide the same performance.
As usual, differences can be found in the details. The peak performance of V 100 is only available when the reconstructed code performs TensorCore's 4x4 matrix multiplication, which is a limitation cleverly avoided by Graphcore architecture. Not to mention that V 100 consumes 300 watts of electricity and a lot of cash.
In addition, Graphcore supports on-chip interconnection and "on-chip memory" method, which can achieve excellent performance beyond TFlops benchmark recognition. In some neural networks, such as generating confrontation networks, memory is the bottleneck.
Similarly, we will have to wait for real users to evaluate this architecture with real applications. Nevertheless, the list of investors and experts of Graphcore and the sky-high valuation of Taiwan Province tell me that this may be a good thing.
Figure 2: Graph Core shows the photos processed by ImageNet dataset. Visualization can help developers know where their training process occupies in the processing cycle.
Havana laboratory
Israeli startup Habana Labs announced at the first AI hardware summit last September that it was going to launch the first chip for reasoning, and its record performance was used for convolutional neural network image processing. The results show that the processor classifies 65,438+05,000 images per second in the Resnet50 image classification database, which is about 50% higher than NVIDIA T4, and the power consumption is only 65,438+000 watts.
On February 20 18, the latest round of financing for Havana Lab was led by Intel Venture Capital, followed by WRV Capital, Bessemer Venture Partners and Battery Ventures, and the company's financing increased by 75 million dollars compared with the previous 45 million dollars.
It is reported that part of Habana Labs' new financing will be used to stream its second chip called Gaudi, which will focus on the training market and is said to expand to more than 65,438+0,000 processors.
Other startups
I know that more than 40 companies in the world are designing artificial intelligence training and reasoning chips. I found that most companies are doing simple FMA (floating point multiplication and accumulation) and mixed precision mathematics (integer 8 bits and floating point 16 bits and 32 bits). I won't be surprised, because this method is relatively easy to implement and will get some results, but it won't provide lasting architectural advantages for NVIDIA, Intel and a few startups to make different architectures.
Here are some companies that caught my attention:
China artificial intelligence chip start-up company
China has been trying to find a way to get rid of American semiconductors, and the artificial intelligence accelerator may provide the opportunity it has been seeking. China has set a goal of building an artificial intelligence industry worth trillions of dollars in 2030. Since 20 12, investors have invested more than $4 billion in startups.
CAMBRIAN Technology, valued at $2.5 billion, is a unicorn company that released the third generation AI chips in China. CAMBRIAN claims that it can provide better AI performance than NVIDIA V 100 with lower power consumption. They also sell its AI IP and install it in the processors of Huawei Kirin 970 and Kirin 980 as AI acceleration hardware.
Shangtang Technology is probably the most valuable artificial intelligence startup, which is famous for promoting intelligent surveillance cameras in China. The number of these security cameras exceeds 654.38+75 billion, including cameras produced by other companies. Shangtang Technology was established in Hong Kong, and the latest round of financing amounted to 600 million US dollars, led by Alibaba. According to reports, the value of this startup is currently $4.5 billion.
Shangtang Technology has established strategic partnerships with big companies such as Alibaba, Qualcomm, Honda and even NVIDIA. The company has a supercomputer today, running about 8,000 pieces (maybe provided by NVIDIA? ) GPU, and plans to build five more supercomputers to process facial recognition data collected by millions of cameras.
Lei Feng. Com, through Forbes