An Informative Intel Industry Analyst Conference 2020

I spent close to three days attending the 2020 Intel industry analyst day last week. All the new information was NDA, I am not going to rehash what has already been said over the past year, but I did want to share some of my high-level opinions exiting the week. Comparatively, there was less new information shared than in previous years and less one on ones to get down to the nitty gritty details. I do not see this necessarily as the company withholding information, but rather, the inconveniences brought on by Covid-19 and the lack of face to face communications. Net-net, my view of Intel has improved in some areas and remains the same in others.  If you have not read my previous opinions on Intel’s strategies and products, please check out the resources at the very bottom of the article.

My goals of the week were to re-assess Intel’s strategy, get an update on its fab position, its products, and position in AI training. 

Strategy

I still think that Intel’s strategy is the right one. It is focused on six pillars of innovation and remaining an IDM. Intel has expanded the definition of compute, which Intel calls the “XPU” across multiple types of compute, the CPU, GPU, FPGA, and ASIC. I have believed in this approach for 20 years and I still believe it is the right path. And I also think right now, IDM (Integrated Device Manufacturer) is still the right move for Intel, with the ability to tweak and tune the designs. If a company is executing well on an IDM strategy, it can also be the lowest cost, and it has the most flexibility. The challenge is on Intel’s 10nm execution but try not to confuse that with the merits of strategically being an IDM.

Intel strategy- simplified view 
INTEL

And I think disaggregating chip design makes sense for Intel, even though this is not what the mobile players are doing right now. Qualcomm and Apple, for the time being, are sticking with monolithic designs. AMD has already disaggregated, and Xilinx and NVIDIA are on its way. You see, the larger the die area and the more heterogeneity the SOC, the bigger the need for a 2D, 2.5D, or 3D design. AMD got there first with 2D, and it is paying off well so far. Intel is first on 3D which is early but looks very promising to achieve more density in new designs. I know we’ve seen Intel 3D in lower power designs, but I’m more excited seeing this in datacenter products for mix and match IP.

Intel CSO Saf Yeboah-Amankwah outside perspective 
INTEL

Intel Fabs

When you ask most people in the tech industry about the state of Intel fabs, I think they would say they were “broken”. I understand this sentiment, but I believe it is too simplistic and not accurate. Intel has had big issues with 10nm execution, delayed over two years, and issues with 7nm execution, which have been delayed another six months. This is an issue, but fundamentally what people are missing are what the company is doing with 14nm and the real progress on 10nm. Intel is servicing 90% of the server and PC industry mostly with a super-optimized 14nm process and silicon designs. Intel even had to “back-port” features and parts from 10nm to 14nm. In the notebook market on 10nm process, Intel is competitive and would say most competitive with battery life. Intel has never shared yield in the 30 years I’ve known them, and the company did not share 14nm or 10nm yields with the group.

I did have the chance to talk 1:1 with Dr. Ann B. Kelleher, SVP and GM of Technology Development at Intel. She is responsible for the research, development and deployment of next-generation silicon logic, packaging and test technologies, and before that led manufacturing where she managed a significant expansion of 14nm supply and ramping the 10nm process. Here is the high-level of what she shared with me that I can share here. Note, I cannot share everything that was discussed.

  • Overall: Fabs are at full capacity, including Ireland, Israel, Arizona and Oregon. The rumored cost issues based on fab underloading is just not happening. Every open space is being converted to factory space including a café in Israel. 14nm is at peak performance and capacity, better than any process she had seen in her 24 years at Intel. Yields continue to improve on 14nm and 10nm.
  • 10nm: Crossover between 14nm and 10nm volume will be sometime in 2021, meaning Intel is adding more 10nm capacity as we speak. Comparing 10nm Ice Lake onwards (not 10nm Cannon Lake), 10nm is within a quarter or so of where 14nm was in its lifetime. Israel, Arizona, and Oregon are all doing 10nm in high volume. SuperFIN has been instrumental in increasing 10nm performance.
  • 7nm: We can expect an official 7nm update in January, but clearly there is an intense focus on meeting product schedule commitments.
  • More fabs: Building even more fab space in Oregon and Ireland and starting additional site prep to have room for 7nm and 5nm.

Net-net for me is the Intel fabs are full, 14nm is performing great, 10nm is better than rumors, we have to wait until January on 7nm, but the focus is there, and Intel is building enough capacity as quickly as it can, short-term and mid-term.

External Fabs

Now let’s talk about external fabs. 

First off, Intel has used external fabs a lot historically if it got into a jam or if it made an acquisition and that design was externally fabbed. We saw this in chipsets, FPGAs and LTE modems. So, when the world went nuts after CEO Bob Swan suggested it was going to look at external foundries, I was puzzled, as it already did that. While I think it could have been more strategically communicated, I suppose Swan had to tell the street coincident when he knew it. What I do understand is the surprise at the notion that Intel would take some of leading edge designs to a foundry. That is new. 

I spent nearly 10 years at an OEM (chip customer) and 11 years at a chipmaker and for the most part, customers don’t care where a part is fabbed as long as it is on time, feature complete and at a good cost. Customers want predictable cadence. Therefore, I think the notion that Intel fabs must compete with TSMC and Samsung for future designs is a great idea and doesn’t negate the strategic benefit of Intel’s IDM strategy. I would love to see Ice Lake on TSMC’s 5nm process, wouldn’t you? 

Products

Let’s talk products. I left the conference more confident with the client computing roadmap through 2022 that was shared for reasons I cannot divulge under NDA. Gregory Bryant aka “GB” answered all my questions and I think he had a great grasp of the market situation and answers for most. Let me just say that GB and the client group isn’t backing down. At all.

Innovate Through Platforms
 INTEL

I had a good 1:1 chat with Lisa Spelman, who leads the Xeon product line . I am quite intrigued by the upcoming Xeon roadmap, particularly in 2023. I think strategically, given where Intel is, what it wants to accomplish, and the competitive set, it’s the right direction. Spelman also briefed us all on an exciting new category of datacenter products but unfortunately, I cannot dive into it. There is a market need for these new products and I am excited about them. I expect we’ll hear more about this in the new year.

3rd Gen Intel Xeon Scalable Processor
 INTEL

AI training

One of the biggest things investors and CSPs are clamoring from Intel was a competitive accelerated ML and DL training solution. Nervana didn’t pan out as planned, GPUs weren’t coming to the rescue, so Intel intelligently acquired Habana Labs. The great news was that Amazon AWS announced at its 2020 re:Invent that same week that it was creating instances based on Habana Gaudi. I wrote about this here. Promised in 2021, AWS set a goal of delivering up to 40% better single-node price-performance compared to its current GPU-based EC2 instances supporting ML. This is a really big deal as NVIDIA hasn’t any competition in training from a tier 1 silicon vendor. I know the EC2 folks at AWS pretty well and don’t think for a second that AWS would productize anything that it didn’t deem competitive, no matter the price. While one CSP make not a market, if you’re going to start with anyone, you want it to be AWS to pull through all the other CSPs, tier 2s and enterprises.

AWS CEO Andy Jassy announces Intel Habana Gaudi Instances 
INTEL

Summary

Overall, I felt much better exiting Intel’s industry analyst day than I did entering it and I appreciate everybody at Intel taking the time to do it. I liked almost everything I saw on the slides and in the conversations, but in the end, it really does come down to Intel’s execution.

I think Intel is going to coming storming back. This does not mean I think the company will achieve the same market share in its traditional 95% PC and server markets- I don’t. I think the cat is already of the bag on that. 

do think Intel can grow as it has widened the aperture considerably when I look at its increased TAM and SAM. The “old” Intel did CPUs and chipsets with monolithic designs for datacenter servers and PCs. The “new” Intel does a family of XPUs (CPU/GPU/NPU/FPGA) with disaggregated designs for datacenter server, storage and networking, the edge, carrier and autonomous cars. The new Intel will use the fab or foundry that is best suited for its disaggregated designs that best meets its customers needs by decoupling design from the state of the fab or its technologies. If that’s the Intel fab-great. If not, TSMC or Samsung. Customers don’t care.

If I were a customer, I would prefer the new Intel over the old. Now it’s up to Intel to prove everybody wrong and execute. 

To get up to speed on my opinions of Intel’s products and strategies, please find these resources below:

Note: Moor Insights & Strategy writers and editors may have contributed to this article.