After acquiring two cloud AI chip startups in succession to consolidate its most important data center market, Intel finally made a choice between the two tigers.
Habana beat Nervana for its excellent scalability technology and design, and its ability to deliver in volume. But this does not mean victory in the end. For Intel, as long as he sees the opportunity to surpass Nvidia, he will not hesitate, the key is only to choose the right chip.
Recently, foreign media broke the news of Intel's new actions in the field of AI. Forbes said that Intel will stop TSMC 's 16nm foundry Nervana chips, mainly the Nervana NNP-T series. Intel did not announce the matter with much fanfare, but has informed analysts and customers via email:
The development of the Nervana NNP-T training chip (formerly codenamed Spring Crest) has been stopped, but the promise made to customers by the inference chip Nervana NNP-I (codenamed Spring Hill) will be honored.
The Habana product line offers the powerful strategic advantages of a unified, highly programmable architecture for both inference and training. By moving to a single hardware architecture and software stack to enable data center AI acceleration, our engineering team can work together to deliver more innovation to our customers more quickly.
On Friday, deep learning analyst Karl Freund also tweeted that Intel will completely stop the NNP-T products of deep learning chip startup Nervana that Intel acquired in 2016 and focus on Habana Labs.
1. Why buy two cloud AI companies?
Although Intel has layouts in the AI field for terminals, edge computing, and the cloud, the top priority is still to take advantage of Intel's strong position in cloud data centers to ensure its position in the cloud AI chip market. This is also the important reason why it successively acquired two cloud AI chip companies, Nervana and Habana.
On December 16, 2019, rumors of Intel's acquisition of Habana Labs were finalized. This acquisition continues Intel's "big-handed" style-$ 2 billion, second only to Mobileye's second largest acquisition.
At that time, Habana had two products, Gaudi AI Training Processor and Goya AI Inference Processor. It is worth noting that the Gaudi artificial intelligence training processor is already providing samples for specific hyperscale customers, and the Goya artificial intelligence inference processor is commercially available.
Among them, Gaudi is the micro-architecture used by Habana to accelerate training. Designed using the TSMC 16-nanometer process, the chip integrates eight TCP and GEMM engines in a cluster. Goya is Habana's microarchitecture for accelerating inference. Goya is manufactured on TSMC's 16-nanometer process, which is actually a simplified version of Gaudi.
However, long before the acquisition of Habana, in August 2016, Intel acquired deep learning technology startup Nervana for $ 408 million.
The main product of this company is the Nervana neural network chip. Its architecture design is very special. For the characteristics of AI requiring high-performance memory, it abandoned the standard cache system and switched to software to manage the memory system.
Unlike Habana, however, Nervana does not have mass production products. Three years after the acquisition of Nervana, Intel launched Nervana NNP-T and Nervana NNP-I.
The Nervana NNP-T series, codenamed Spring Crest, is mainly for AI training. It is produced using TSMC's 16nm process and has a core area of 680mm2. It integrates 27 billion transistors and is equipped with 32GB of HBM2 memory. The frequency is 1.1GHz. TDP 150-250W. Tesla series GPU acceleration chip.
In contrast, the Nervana NNP-I series AI chip is much smaller, codenamed Spring Hill, which is mainly oriented to AI inference applications. The CPU part is Intel's 10nm process Ice Lake core, which consumes between 10-50W and has M. 2 and PCIe specifications, more compact and flexible.
2. Who is more reliable?
Just one month before the announcement of Habana, Intel reiterated its delivery plan for Nervana chips. At that time, some analysts believed that the acquisition may indicate that customers were not satisfied with Nervana hardware and were unwilling to use Nervana again. Since Nervana cannot keep up with the rapidly growing market, Intel will continue to explore other options.
Now, Intel stopped TSMC's 16nm foundry Nervana chip, which also confirms this speculation. Intel said it made the decision after seeking feedback from its engineers and major customers. Feedback indicates that the second-generation Nervana designs, codenamed Spring Hill and Spring Crest, simply do not meet the requirements of those high-performance workloads.
These customers also mentioned that Habana is a preferred platform to compete with Intel. Among them, Facebook has made it clear to Intel: "You need a better chip."
First, Habana network technology is likely to be one of the key reasons for Intel's decision to abandon Nervana in favor of Habana technology.
Nervana's Neural Network Processor (NNP-T) scales using a proprietary interconnect, while Habana's Gaudi can scale to thousands of nodes over a standard 100Gb Ethernet. Moreover, Gaudi even supports remote direct memory access RDMA, a feature that enables software to access memory in the entire fabric without increasing the burden on the remote CPU.
This structure can significantly improve the performance of training large-scale neural network models to cope with the trend of doubling the size of neural network models every 3 and a half months, handling the trend of increasingly complex AI tasks.
In addition, some analysts pointed out that full consideration of software and hardware coordination when designing the underlying compiler and software architecture also helped Habana chips achieve better scalability. According to official figures, its distributed overall performance can be close to linear even when the number of processors is greater than 600. Compared to the Nvidia V100 GPU, the training performance is improved by nearly 4 times, which is a very remarkable result.
Finally, Habana chips are not only more powerful, they have been shipping since the end of 2018. In contrast, Nervana not only took three years from acquisition to product delivery, but also the product was repeatedly delayed.
Buying out a competitor is a logical move by Intel. The success of Habana does not mean that he will not be replaced one day. After all, Groq is doing its best, and more companies are bringing their chips to market.
No comments:
Post a Comment