For a rare startup behavior of a 100-employee start-up to acquire another Silicon Valley “predecessor” of comparable size, some industry analysts say Wave’s move is bold. Karl Freund, principal analyst for high-performance computing and deep learning at Moor Insights & Strategy, said: “For small start-ups, the expansion of the market through acquisitions before its first product leaves the market is very bold. This is a very Risky behavior, but this is what it needs. "
Although Wave Computing's acquisition of MIPS is considered very bold, in fact MIPS's DNA seems to have been embedded in Wave. Wave's CEO Derek Meyer was formerly MIPS' sales and marketing vice president, and Mike Uhler, vice president of operations, is MIPS' chief technology officer. The vice president is Darren Jones, the former engineering director of MIPS, and the chief legal adviser, Wave Paul Alpern, was the chief business advisor at MIPS. Therefore, unlike most mergers and acquisitions that take a long time to adjust, it is not possible to develop a roadmap.Seven of Wave's seven executives have worked at MIPS, so the Wave-MIPS team does not need much integration time.Wave plans to deliver its first batch of AI systems to the first batch of customers later this month. "By the end of this month, we will also launch a roadmap for the universal AI platform based on MIPS and Wave DPU, most likely in San Francisco. Design Automation Conference (DAC)" Meyer said.When asked when the AI-MIPS will appear, he promised before the end of this year.
Wave claims to accelerate neural network training by GPU 1000 times
Wave Computing is a Campbell, CA company focused on using its massively parallel data flow architectureDevelopmentAfter the AI system startup company was established more than seven years ago, Wave's latest DPU multicore architecture early experience project was finally opened. At last year's high-performance chip summit Hot Chips,Wave Computing's CTO and Dr. Chris Nicol, chief architect of the DPU (Dataflow Processing Unit), said that their product DPU can surpass GPU 1000x in accelerating neural network training, and believes that early user trials can confirm their position on DPU.Taking into account the GPU's position in the current deep learning and training market, this is indeed a bold declaration.
Wave's DPU has 16000 processing elements, more than 8000 arithmetic units, does not coordinate through the CPU, all cores operate at 6.7GHz (average), and uses a coarse-grained reconfigurable architecture. The DPU has a unique self-timing mechanism. When no data passes, the DPU enters a sleep state. More generally, the DPU can be viewed as a hybrid FPGA and multi-core processor that can handle the static scheduling of data flow graphs for thousands of elements. This design is very different from other deep learning hardware startups.
Nicol pointed out that the accelerator architecture (especially the GPU) has two problems. One is that there is a delay in loading the new kernel. The other is to solve the first problem. When using the MCU to move programs in and out at runtime, the program itself decides when to The MCU communicates but there is also a program cache on the chip. The end result is that no CPU architecture can achieve greater performance in an offload model. On the software side, Nicols said: "Deep learning is actually a data flow diagram that is programmed on top of deep learning software. Running on a processor like ours, you can assemble a data flow diagram at runtime. For example, at runtime The data flow graph is taken from TensorFlow and converted directly at runtime, executed without a CPU, and mapped onto the data flow chip."
Wave Computing DPU
Since data center training currently uses GPUs, and this market is currently dominated by Nvidia, whether Wave’s success in the data center market has affected Nvidia’s position in this market is very much concerned. As one of the more than 20 companies that Silicon Valley entered earlier in the AI space, Wave is also an early company to bring its products to market. Wave claims to outperform its competitors in an innovative way.
Can MIPS sold by Imagination help Wave?
Wave now positions itself as a broader player in the field of artificial intelligence, and its technology and intellectual property will enter the edge of the training and processing market. MIPS has shifted from high-performance CPUs to energy-efficient processors for IoT devices and other low-power applications. And, while the DNN data center training market is very hot today, the future of smart IoT devices will depend on the efficient edge training or reasoning of DNNs.Wave's acquisition of MIPS can build platform and license IP for these applications, which means it can extend product applications from the cloud to the terminal.
There is no problem with Wave's logic to acquire MIPS, but can acquisitions really help Wave succeed on the terminal AI chip? MIPS was founded in 1984. At the end of 2012, when CEVA and Imagination Technologies rushed to acquire MIPS, many industry observers wondered what Imagination would do with MIPS. After all, the MIPS with a revenue of about US$60 million this year was at a loss. status. When Imagination finally acquired MIPS at an overwhelming $100 million price, industry observers believe that this high-priced acquisition shows that the MIPS CPU is important for the future development of Imagination.
However, contrary to expectations, the acquisition of MIPS did not bring Imagination the expected results, but also dragged down Imagination's performance. Last May, Imagination began selling MIPS, which shrank from $100 million to $65 million in less than five years. According to Kevin Krewell, Principal Analyst at Tirias Research, “MIPS is still a classic CPU design with a scalable and mature software ecosystem. So despite the reduction in MIPS applications, there are still market opportunities.” Others believe that MIPS CPUs The architecture has certain performance and efficiency advantages. For example, multi-threading technology, MIPS can also play an important role in some real-time, power-sensitive applications, such as LTE, artificial intelligence (AI) and Internet of Things (IoT). More importantly, it is reported that MIPS has a team of about 200 engineers developing CPU technology, some of which are even more advanced than Arm can provide - such as multi-threading.
However, due to the fact that the market is currently dominated by Arm and x86 architectures, MIPS does not have much ecosystem, and many analysts also raise the question of who wants to buy MIPS. However, in September last year, Imagination announced that it sold the company to Chinese-owned Canyon Bridge and agreed to sell its MIPS CPU business to Tallwood Venure Capital. Although we do not know the final price for acquiring MIPS by Wave, the acquisition of talents and patents is beyond doubt for the enhancement of Wave's strength. It is still necessary to observe whether it can play a greater role in terminal artificial intelligence.
Does the intelligent terminal of the IoT terminal continue to shine brilliantly or is it replaced?
At present, Arm almost monopolizes the CPU IP market of mobile devices, but MIPS and RISC-V seem to be playing a role in IoT smart terminals. RISC (Reduced Instruction Set Computing), namely the RISC (Instruction Set Computing) system, corresponds to the complex instruction set (CISC). The number of RISC instructions and addressing modes are shorter, the execution efficiency is higher, the manufacturing process is simple and the cost is low. It was first born in the 1980s. The MIPS host, but then Arm also used this processor architecture.
The MIPS architecture was born several years earlier than Arm. At the beginning, the gap between MIPS and Arm was also small. The former also accounted for about 30% of the RISC architecture microprocessor market. However, when the rise of the mobile Internet, MIPS started to fall behind due to technical limitations. , And academic development style also makes the MIPS business process lags far behind Arm, of course, MIPS eventually defeated by Arm had many complicated reasons.
However, after being acquired by Imagination, it did not intend to let MIPS die in the mobile phone market, but more in the wearable device market and network equipment, and Imagination management believes that its neglect of management has affected the MIPS development blueprint. But in the era of IoT, MIPS has the opportunity to take advantage of it.
Donald Zhang, founder and CEO of OURS, stated that the needs of the Wi-Fi are high customization, high modularity, extensibility, and support for new technologies. The Arm architecture does not allow new things to be added, nor does it allow customization and modification, and Patent licensing fees. On the contrary, technically, RISC-V has 5-6 times lower power consumption and 5 times more area efficiency than Arm-based processors, allowing developers to have a lot of freedom to optimize for specific applications. In business, RISC -V open source has no patent licensing fees and is very friendly to startups. Therefore, RISC-V is also comparable to and rivals Arm. Even if IoT does not replace Arm, RISC-V will become a very important player.
So, does RISC-V and MIPS replace Arm in IoT smart terminals, or at least become an important player?