In a rather puzzling turn of events, Intel has decided to scrap its plans to further develop its promising Nervana Neural Network Processor for the training line. The decision comes mere two months after the severely delayed Nervana NNP family was officially launched. Although unconfirmed, the confusing move is most likely the result of Intel’s recent acquisition of Habana Labs.
A mere two months after Intel acquired Habana Labs for $2 Billion, the former has axed its own Nervana Neural Network Processor project intended for the Artificial Intelligence training line. Incidentally, it is not surprising that Intel scrapped the same. This is because it was rather redundant and counterproductive to keep investing in two competing AI-oriented products. It was only logical that one of the two chips, which are architecturally very different but intended for the very same markets, would be dropped from further development.
Intel Scraps Nervana NNP Family In Favor Of Habana Labs Product But Will Honor Customer Commitments:
Intel appears to have decided to discontinue the development of its own Nervana NNP family of AI processors. However, the Intel NNP-I will continue for a little longer due to customer commitments. Still, Intel will eventually discontinue all development in favor of the Habana chips. Interestingly, Intel also announced 3rd Generation Movidius VPUs, codename Keem Bay. The company has confirmed that the Movidius roadmap for vision processing remains unchanged.
Although the acquisition of Habana Labs by Intel makes a lot of sense, the discontinuation of its own in-house design for the former’s alternative does not. Incidentally, none of the AI Chips from either of the companies have been independently tested and their detailed specifications and benchmarks openly published. Still, the officially released specifications, features and performance attributes of the Nervana NNP family of AI processors and those developed by Habana Labs do offer a lot to compare them.
— cleburnepcrepair (@CleburneTXPC) February 3, 2020
The Intel Nervana products are sold under the NNP brand while Habana used to sell under the HL Series. Both Nervana and Habana offer a standardized PCIe card and an OAM mezzanine module. Simply put, physically both the products appear similar. Furthermore, these chips feature 32 GB of HBM2 memory. The Intel’s Nervana chips have slightly higher TDPs and higher clocked memory, but interestingly, they consume lower power during actual operation.
The Intel Chips are based on the Spring Crest Microarchitecture, while the Habana Labs’ chips are based on the Gaudi Microarchitecture, and this is where the differences become apparent. There are several critical pieces of information missing, which prevent detailed comparison. However, it is quite obvious that Intel’s Nervana’s design is much more complex. Spring Crest implements a uniform 2D mesh of 24 Tensor Processor Clusters (TPCs), which each Cluster comprising of an On-Chip Router (OCR), the control, the MAC Processing Units (MPU), and the memory subsystem. There are about two MPUs per cluster.
— Karl Freund (@karlfreund) January 31, 2020
The Nervana system ensures data remains localized, reducing data movement and improving reuse. Additionally, Intel was also reportedly working on an ICL switch to allow further flexibility, in addition to the parallelism that was already baked directly into the architecture. There are no official scaling benchmarks from Habana but pre-production chips from Nervana reportedly displayed good scaling utilization across 100s of nodes with very low latencies even at sizable transfer sizes.
Why Did Intel Scrap Its Own Efficient, Superior And Highly Scalable Nervana Chips For Habana Labs?
Although exact details are a little difficult to obtain, the Nervana family appears to be technologically superior when compared to the alternatives presented by Habana Labs. Intel claims the Habana unified architecture is a “strategic advantage”. Intel’s decision might have to do with market dynamics surrounding the availability of HBM Memory as Habana Labs’ products can also work with standard DDR4 interfaces.
— Omar Khan (@66owls) February 3, 2020
Experts, however, claim that the decision to abandon the Nervana might have to do with the level of software development. First-generation Nervana NNP’s (Lake Crest) had previously encountered software-related difficulties due to using the Flexpoint data type, forcing Intel to choose bfloat16. Whatever the reason, Habana Labs’ acquisition will allow Intel to gain market share in the growing AI ASIC market, and hence the decision might have been inspired by economics as well.