Upcoming CPU and GPU technologies require increased support for the development community, including eleven new development partners and new tooling.
Concurrently, and not coincidently, Amazon announces availability of Intel Habana Gaudi on AWS.
In 1997, Pat Gelsinger, then Intel CTO, realized that the company needed to ramp up its outreach to the development community, so he founded the Intel Developer Forum. Many were dismayed when new management cancelled the event 20 years later in 2017. Saving money may be good, but walking away from developers is very, very bad. Clearly, the new guys did not understand the business they were in where developers hold the keys to success. They need information, tools, and support.
Upon his return to the company, Mr. Gelsinger is quickly righting that wrong with the first “Intel Innovation” event. He came out with CTO Greg Lavender to express his commitment to the community, saying “Developers are the true superheroes of the digitized world – a world which is underpinned by semiconductors…By the end of this decade, there will be the potential for every human to have 1 petaflop of computing power and 1 petabyte of data less than 1 millisecond away.” Powerful stuff indeed.
And Gelsinger backed that up with new investments in the community, critical as the company brings completely new technologies to market. Technologies that requires an engaged development community to turn silicon into solutions.
What did Intel Announce?
Intel covered the waterfront from Xeon to Ponte Vecchio GPUs, to IPUs, to Habana Labs, and to the OneAPI software that will rule them all. Eventually. And they announced partner engagements with Google (IPUs) , Alibaba (Xeon-based DeepRec Recommendation engines), US DOE (OneAPI), AT&T, Fedex (Tofino), and others.
The event contained something for everyone, with a clear message to the development community: “you can count on us to be open and to provide you with a choice of technologies to meet your needs.” Here’s a snapshot of what the company announced, as it relates to AI.
… we need to segue briefly to Amazon AWS, which announced on the eve of Intel’s event that the cloud service provider is now ready to support Intel Habana Gaudi AI accelerators on their cloud, announced last December. Gaudi on AWS is absolutely huge for Intel, which discussed in a blog how AWS substantiates their claim that Gaudi offers 40% better price performance than NVIDIA GPUs. This blog, from the normally secretive Habana Labs, readily admits that this is an unoptimized model on the GPU, but that the organization is submitting MLPerf benchmark results that incorporate optimization techniques. (ok, they also claim that optimization is hard work, but…) We have previously expressed our dismay at AWS comparisons, which do not give any credence to optimization on NVIDIA in order to support their claims for superior price performance for their own Inferentia silicon. However, we are impressed with Habana’s commitment to providing fair, optimized, apples to apples comparisons using soon-to-be-released MLPerf benchmarks.
Ok, back to Intel Innovation
Intel had many announcements, all interesting and worthy of a more exhaustive analysis. While none are earth-shattering, the real story is that Intel is investing in the ecosystem the developer community engenders. In a nutshell the AI-focused announcements included …
- Intel announced a unified Developer Zone with a new oneAPI toolkit, with new Centers of Excellence to enable developers to access reference designs, toolkits and other assets across AI, client, cloud, 5G/edge and gaming with an open, standard-based, unified programming environment.
- Aurora, the long-delayed DOE Exascale supercomputer, and probably the last Intel-primed supercomputer, will nearly double the exascale mark, delivering two exaflops of double-precision performance. This is primarily a testament to the Ponte Vecchio GPU, but also includes Sapphire Rapids Xeon CPUs. However, we must note the impending announcements from AMD regarding their new GPU currently being installed at Oak Ridge National Labs. It is quickly becoming a three-way race between AMD and Intel versus the leader, NVIDIA for high-performance AI and HPC GPUs.
- Google Cloud and Intel unveiled a partnership for joint development of Mount Evans, Intel’s first ASIC-based IPU. NVIDIA created this market with their DPU, but the frenzy for market share is just beginning.
- Intel is preparing to ship oneAPI 2022 toolkits with 900 new features added since it first shipped last year. This new release adds cross-architecture software development capabilities for CPUs and GPUs through the first unified C++/SYCL/Fortran compiler and Data Parallel Python and expands Advisor accelerator performance modeling. Unfortunately, no mention was made to when oneAPI might support for Habana Gaudi, but honestly that is not a high priority at this early point in Habana technology evolution. No harm, no foul.
- Intel announced that SiPearl is designing a microprocessor that will be used in European exascale supercomputers and has selected Intel’s Ponte Vecchio GPUs as the high performance computing (HPC) accelerator within the system’s HPC node. SiPearl is also adopting oneAPI as the open software specification. We continue to be quite bullish about Ponte Vecchio.
While the Intel Innovations event is still underway, I think it is already safe to conclude that the company is on the right track: fix the manufacturing mess, improve the products using architecture and packaging, all while rebuilding developer relations. While this massive undertaking will take time, progress on Ponte Vecchio GPUs, Habana Labs Gaudi, and of course Xeon CPUs for AI performance is impressive. Pat Gelsinger’s focus on the developer community is sincere, and will be critical to turn all that sand into solutions that solve business and scientific problems. Continued execution will be critical to turn this momentum into renewed leadership for Intel in the increasingly AI-centric world they nearly missed.