NVDA - Intel DCAI Business Update Leads To Premature Optimism

2023-03-31 04:19:32 ET

Summary

• Intel DCAI Business Update meeting caused a strong runup in the stock, but a keen evaluation shows the optimism is premature.
• The roadmap presented does not catch up to the competition until 2025 even if the execution is flawless.
• Intel's roadmap success is more likely to depend on the chips Intel tapes out to TSMC than the chips Intel fabricates in-house.

Intel ( INTC ) gave its DCAI business update on Wednesday. The update was well received by investors and INTC stock moved up by about 9% in the last couple of days - although part of the run-up was macro-driven buoyed by Micron Technology's ( MU ) inventory commentary.

As has become common lately, Intel's pitches start with a TAM narrative. This is because growing TAM is important in a world where market share gains are unlikely. The TAM estimates in the image below look reasonable - especially considering the AI megatrend underway.

The more questionable part of Intel's presentation was its view of the AI silicon market (image below). Intel expects 60% of AI silicon opportunity to be in CPUs and 40% to be in accelerators. In terms of the compute workloads, we know that a lion's share of the training is being done on GPUs and accelerators. This leaves inference as the main battlefield between CPUs and GPUs. Intel claims that Xeon dominates in inference for small to midsize inference models up to 10 billion parameters. Another Intel argument is that CSPs may move to accelerators but enterprises will likely use CPUs for inference. Furthermore, Intel's claim is that most AI workloads, such as data processing and analysis, are general-purpose workloads that run best on CPUs for several technical and economic reasons that include the ubiquity of the x86 architecture.

These seem like reasonable arguments but the 60/40 CPU/Accelerator split shown in the image is unlikely unless Intel's claims hold and most of the inference is done on CPUs instead of GPUs or accelerators.

The main thrust of the business update was on the product front. Intel laid out its new CPU product roadmap through 2025 (image below). The key to note here is that Intel is showing progress and commitment to its two-pronged server CPU strategy using two different CPU cores: The performance cores or P-Cores are optimized for high single-thread performance for mainstream applications. These are under the "Rapids" family. The efficiency cores or E-Cores are optimized for high throughput applications that benefit from high core density (instead of peak single-thread performance). These chips are under the "Forest" product family.

The market and many analysts likely found this roadmap attractive because, for the first time in many years, Intel appears to have a somewhat credible competitive roadmap.

Probably the best news for investors in the short term is the relatively strong adoption of Sapphire Rapids. Despite being very late into the market and competing unfavorably with competition on most metrics, management commented that the Company is seeing broad adoption across the enterprise, cloud service providers, comms service providers, OEMs, and ODMs. Management noted that the top 10 CSPs are deploying Sapphire Rapids and the Company now has over 200 designs shipping currently from all major OEMs and ODMs. To be sure, this is a stronger adoption than expected for this troubled product and it is unclear how much of this is due to Sapphire Rapids being part of Nvidia ( NVDA ) Hopper reference design. On the flip side, the volume seems to be on the lighter side with management commenting on being on track to ship 1 million Sapphire Rapids units by midyear.

Looking past Sapphire Rapids, as discussed below, this roadmap is far less impressive than it may appear to uncritical investors. Let us start with each product on the timeline.

The first product, Intel's next-generation Xeon and a follow-on to Sapphire Rapids called Emerald Rapids, is now sampling and will start shipping in Q4 (image below).

Emerald Rapids is an incremental update over Sapphire Rapids. While being socket compatible, the main change is that what was previously 4 die solution is now revised into a 2 die solution. Note that this integration is going in the opposite direction of the trend towards smaller dies and chiplets. The most likely reasons for the integration are that Intel has found the EMIB packaging on Sapphire Rapids expensive and the Company does not have an equivalent of AMD's Infinity Fabric ready yet to go the chiplet solution route.

This compromised solution is expected to have 66 cores across both the dies - more cores than Sapphire Rapids but far less than Advanced Micro Devices' ( AMD ) Genoa's 96 cores. In other words, Emerald Rapids likely improves Intel's competitive position slightly but falls well short of AMD's equivalent.

The next product in the roadmap is a new E-Core server chip that competes with AMD's Bergamo (image below).

Sierra Forest is expected to be the first product on the Intel 3 process. That makes it an extremely high-risk product especially when one considers that Intel is only shipping the Intel 7 process now and has pointed to the high cost of Intel 7 as a reason for some of the recent margin crunch. Since the product is sampling, one could technically claim that the process works. But it is a long way from a prototype to production on a brand-new process.

Even assuming that process is not a problem, readers should be aware that Sierra Forest has 144 cores but does not support simultaneous-multi-threading, or SMT. AMD's Bergamo, on the other hand, has 128 cores but has SMT and, consequently, supports 256 threads. Bergamo is expected to be shipping in H2 2023 and has no process risk - given that TSMC ( TSM ) has been in production with its N5 process for about 3 years now.

E-Core that Intel used in Sierra Forest is a reduced version of the P-Core and thus less performant than the Zen 4c cores used in Bergamo. This makes it likely that Bergamo will beat Sierra Forest to market, support more threads, and be more performant in several applications.

In other words, Sierra Forest, arriving in H1 2024 will not help regain Intel's competitiveness. Although it could help Intel narrow the competitive product gap on the server side.

The next product on the roadmap, Granite Rapids, did not get much discussion but appears to be a mid-2024 or H2 2024 product. By this time, AMD would likely be shipping its next-generation Zen 5 product line. Achieving product leadership in that timeframe seems doubtful. Time will tell if Granite Rapids helps Intel narrow the gap with AMD.

This takes us to the final product on the roadmap - Clearwater Forest product in 2025. This product is being targeted for 18A process which Intel expects will be its leadership process. If Intel can deliver on this promise, is conceivable that, for the first time in many years, Intel could regain product leadership - this of course assumes that AMD fails to field a competitive product in that timeframe.

Not Much To Write Home With Non-CPU Roadmap

The rest of the roadmap, in GPUs, FPGAs, and accelerators does not show anything particularly impressive for 2023 or 2034 (image below).

Habana Gaudi seems to be a compelling product but mostly in comparison to Nvidia's A100 family. With Nvidia starting H100 shipments last quarter, it will take Gaudi 3 for Intel to have a chance at leadership. And Gaudi 3 is not expected to arrive until 2024. This, when coupled with the already known cancellation of some GPU programs, suggests that Intel will be relegated to the lower end of the accelerator market for the next couple of years and compete primarily against Nvidia's A100 and AMD's MI250. Intel's wins against H100 or MI-300 generation will be rare.

While Intel focused a good bit of discussion during the update on how its OneAPI effort which makes developing applications seamless across CPUs, GPUs, FPGAs, and accelerators, the challenge is that it will be difficult to win sockets against a strong incumbent like Nvidia without offering compelling GPU/accelerator performance. So, the software story is not likely to have much traction until Intel's hardware is competitive - which does not appear likely until 2025 or later.

Management stated that Intel is the volume leader in the IPU market with our FPGAs, Xeon D, and ethernet components being deployed in 6 out of the top 8 hyperscalers. While this may be true, when it comes to the latest generation DPU products, Nvidia's Bluefield and AMD's Pensando solutions seem to be winning a lot of sockets against Intel's Mount Evans alternative.

Summary

The biggest news, and the investor enthusiasm, has to do with the fact that after years of fielding non-competitive roadmaps, Intel has put forward a more competitive roadmap. But there is a lot of difference between showing a roadmap and delivering products. Even assuming Intel can execute per the disclosed roadmaps, leadership products from Intel will be absent until 2025 or later.

What could save Intel in this high-risk venture is whatever silicon it is outsourcing to TSMC. If the silicon is similar to what Intel is developing on its internal nodes, the risk of the products being in the market would be reduced, although these products would still be much higher cost compared to AMD products based on small die chiplets.

All things considered, the optimism that caused the stock run-up in the last few days is, at best, premature.

For further details see: