Analysis On Intel's "AI Everywhere" Event

Here's my net-net on Intel’s "AI Everywhere" event, Intel held Thursday in New York City.  This proceeded with a 1:1 conversation I had with Intel CEO Pat Gelsinger on Wednesday. Here we go:

1/ Intel played its “AI Everywhere“ position well.

I see “everywhere” in the context of Intel as datacenter, the datacenter edge (ie, retail, factory), aka Web 4.0, and Windows PC. This does not include smartphones, smaller IoT devices, or automotive. I believe this is the right position for Intel to take as it can define the market it wants to “win” in, measure it, and go execute.     

2/ Intel is bringing more awareness to the “AI PC” category with “Core Ultra” code-named “Meteor Lake”, which is good for the overall growth of this market segment.

It is smart to stress scale related to the ISV community as the scale does matter to ISVs. In the end, new, bold, black-and-white, and demonstrable use cases that show the AI magic will matter most, along with marketing dollars, to make a difference at the point of sale. We can’t forget that Intel has had AI features accelerated by SIMD and GPU for many generations with little market impact. What I think is different is the on-device NPU at 11 TOPS and it has a more orchestrated effort with “100 ISVs” to leverage Ultra’s CPU, GPU, and NPU.

I will note that it is much harder for ISVs to characterize apps across CPU-GPU-NPU versus just a fixed TOPs NPU. I must note that Intel has the ISV scale advantage versus AMD, but this doesn't guarantee success else we wouldn't see AMD gain so much share historically. We will see whether Qualcomm, which has a giant and successful Android ISV program whose ISVs also have Windows apps (ie, Meta and Zoom), is aligned with Microsoft, and whose CEO is on Adobe’s BoD, and what that brings to Windows.

It is a positive sign for Intel to crank out its first distributed architecture SoC, using chiplets versus a monolithic design. Unlike Sapphire Rapids on the server side, plagued by delays, Meteor Lake shipped when it was expected, leveraging the learnings it had on Sapphire Rapids. Intel 4 with EUV and Foveros 3D Packaging all showed their faces with Core Ultra.

Intel will face a heated perceptual battle in 2024 with Qualcomm as more Elite X platform details and use cases emerge. Elite X has 45 NPU TOPS and a radical processor architecture that is racking up some early PPW numbers on DVTs.

Intel kicked off phase 1 of what I think is a 3 phase AI PC evolution in 2024.

3/ Intel delivered on its promise with 5th Gen Xeon aka Emerald Rapids.

While it was fun for pundits to kick Intel when it was down on the server side, those same pundits, if they are looking at the facts, concede that Intel delivered on Emerald Rapids. I don’t see how this takes market share from AMD’s Epyc, but quite frankly, I view Emerald Rapids as a pipe cleaner for its upcoming 2024-2025 roadmap that, if on time, will likely dent AMD's share. Intel’s execution has rapidly improved, but there are still many doubters amongst its customer base which is understandable as AMD gained a double-digit share and each CSP (AWS, Azure, GCP, OCO) has its own Arm-based SoC.

Intel did a decent job showing how AI workloads are accelerated on Xeon. As we saw with many ML workloads like recommendation engines, it’s more efficient and cheaper to do AI on Xeon with a tiny accelerator versus segmenting the app (AI and the app) and using a GPU or ASIC accelerator over PCIe, swapping memory back and forth. I’d like to see this on-chip capability more clearly communicated by Intel. The fact is, NVIDIA GPUs have accelerators (i.e., Tensor Cores), and so does Xeon. I think Intel holds back because they don’t want to look defensive for its lack of share on datacenter AI related to GPU and Habana.  

Net-net: Intel is demonstrating that it is executing what it said it would do. When you look at where the company was 2-3 years ago, that should say a lot.