- After over two years of delays, Intel has launched its highly anticipated Sapphire Rapids Xeon, its biggest update in more than half a decade.
- As Pat Gelsinger had promised, without delays, it would have been a leadership product, obliterating AMD’s… last-gen Epyc Milan. AMD’s recent Genoa still reigns supreme.
- For Intel investors, this CPU is a stepping stone and proof point for what Intel could accomplish without process and design issues. Intel remains on track to the 2024-2025 turnaround.
- Intel finally announces dividend reduction. It is one of the many actions Intel has done to respond to the near-term demand environment. The technology progress remains on track.
The new Intel (NASDAQ:INTC) Sapphire Rapids Xeon is ultimately a tale of two stories. On one hand, it is genuinely a great product. It was meant to compete against AMD’s (AMD) Milan series, and it easily gets the job done, from time to time outright obliterating the part. On the other hand, due to the much-discussed execution issues with designing this Xeon (which have overshadowed even the prolonged process delays), it is late to market by at least two years. This means by now it is actually competing against AMD’s next-gen Genoa, which due to its superior 5nm process Sapphire Rapids is simply no match for.
As such, Sapphire Rapids is a CPU that for the first time in years gives investors a glimpse of what Intel leadership in the data center could (have) look(ed) like (when one ignores the new Genoa part in the reviews). As a CPU design that was conceived in 2015, two CEOs ago, this gives investors a decent baseline for what Intel could accomplish with its newer CPUs going forward. As Pat Gelsinger said in the dividend reduction call, Intel is starting to see the transformation unfold right in front of them, with many green shoots starting to emerge.
In other words, since Intel maintains that both the process and design execution issues have been fixed, Sapphire Rapids at this point should merely be seen as a steppingstone towards regaining leadership in 2024-2025, in line with what Pat Gelsinger promised when he became CEO.
Last month I covered Intel’s Q4 results: Intel Q4 Earnings Disappointing, But Long-Term Growth Story Intact (NASDAQ:INTC). As this article will discuss, the overall thesis remains valid, although obviously there will now be a lower dividend while the financial performance remains subpar.
Nevertheless, the long-term investment thesis is arguably unchanged, as investing during a downturn could lead to strong returns down the road, in this case in preparation for the next upturn in conjunction with Intel’s return to tech leadership.
Sites like Tom’s Hardware and Serve The Home have published deep dives on Sapphire Rapids, so I will refer to those for the technical overview and benchmarks. Let’s summarize some of the key points of why Sapphire Rapids is important:
- Intel’s first disaggregated (chiplet/tiles) CPU, as promised in 2017, bringing Intel’s advanced packaging tech to the data center, allowing Intel to break past what is called the reticle size limit of lithography tools, which limits a traditional monolithic chip to around 850mm2. For comparison, Sapphire Rapids is nearly twice that size as it consists of four 400mm2 tiles.
- Closing the core count gap (60 cores max) that existed since AMD went to 64 cores on 7nm in 2019 while Intel remained at 28 cores and moved to 40 cores in early 2021, although AMD recently increased its core count to 96 with Genoa.
- Closing the architecture gap with both Intel’s own PC parts as well as AMD, as Golden Cove in Sapphire Rapids already debuted in Alder Lake in late 2021.
- Moving to next-gen I/O with DDR5, PCIe 5.0 and the introduction of CXL.
- Besides these more traditional platform improvements, Sapphire Rapids introduces a large barrage of new and/or improved accelerators, either as part of the CPU architecture or as additional silicon besides the CPU cores.
Expanding a bit on the last point, by definition accelerators drastically improve performance and efficiency on select dedicated (data center) workloads. For example, the biggest triumph is arguably the new AMX instruction set, which can improve training and inference performance of AI by an order of magnitude gen-on-gen (10x). Note that while training has long moved to GPUs, the inference (production use) part of AI workloads remains a valid use case for CPUs.
While there continues to be debate about how to evaluate such accelerators, in the case of AMX for AI, unless one has lived under a rock it should be clear (Stable Diffusion, ChatGPT, etc.) that AI has become an important part and workload of computing, making AMX a real and valid selling point over any other CPU.
Nevertheless, admittedly this does provide some difficulty when comparing CPUs. For example, in the case a reviewer decides to ignore all those accelerators and only use traditional benchmarks, then clearly not all of the Sapphire Rapids silicon on the package is used, as those accelerators remain idle. On the other hand, since those accelerators provide asymmetrical performance improvements (measured in triple digit or higher percentages of performance and power improvements), then incorporating those benchmarks in a summary performance graph may similarly give a misleading picture, since the average would comprise both regular and accelerated scores.
With that out of the way, the above overview (average of a large set of benchmarks), which indeed includes a set of benchmarks using the AMX accelerator, shows that Sapphire Rapids on average outright obliterates Milan, the CPU it was intended to compete against. This means Intel has finally re-established itself as the unquestioned leader in the data center. Job done.
Job not done
Except that, as mentioned, AMD has since moved past Milan. That is why Sapphire Rapids is a tale of two stories, which warrants a further discussion of these competitive dynamics.
On the day of the Sapphire Rapids launch, the New York Times of all publications released an interesting article that dived a bit deeper into the issues that Intel encountered. It says that Sapphire Rapids pathfinding started in 2015 (when AMD was at the brink of bankruptcy). The CPU would eventually be “completed” in 2019, with first silicon in early 2020. Except that it wasn’t complete, as a barrage of bugs were found, requiring many more revisions, pushing the timeline back from early 2021 to late 2021 to spring 2022, and then to early 2023 with the latest bug found last year.
In an explanation reminiscent of the cause of the 10nm delays, Intel said that it tried to pack too much innovation in one new product, which seems plausible given the summary provided above.
It turns out that taking so much risk in a market that demands a 1-2 year cadence will get punished, which is what Genoa has done and continues to do after the Sapphire Rapids launch.
While obviously it isn’t true, let’s assume for the sake of argument that Sapphire Rapids had launched on schedule, what kind of articles would have been written then? Well, early 2021 is when AMD launched Milan, and as mentioned above the reviews show that, when ignoring Genoa, at worst Sapphire Rapids is roughly at parity with Milan, while in the best-case the dozen or so accelerators in Sapphire Rapids can be utilized to result in unbeatable performance. Hence, Intel would have been the unquestioned leader in the data center for the last two years.
Of course, making use of TSMC’s (TSM) process leadership with 5nm (which has been in production since 2020 already, so AMD hasn’t even made fully use of it), AMD recently moved to Genoa, which in turn obliterates Sapphire Rapids, which people and reviews have rightfully pointed out.
Nevertheless, the reason I have focused here on Milan vs. Sapphire Rapids is because Intel never intended Sapphire to compete against Genoa, as instead it had Granite Rapids on 7nm (renamed to Intel 4) in the pipeline to have a level playing field in terms of process technology, a CPU that has also been delayed by two years.
So for investors, arguably it is more instructive to observe that Sapphire Rapids is a leadership product against the CPU it was intended to compete against. It doesn’t compete against that product because of design and manufacturing execution mishaps, but that is exactly why Intel two years ago already replaced Bob Swan (a CFO who mostly just continued the strategy of his predecessor BK) with Pat Gelsinger.
In other words, the turnaround has been in progress for two years now, although it is only now that one of the products that necessitated the turnaround in the first place is coming to market. The product is solid (just as 10nm on the process side), but the non-timely execution was not.
Under the assumption that all issues are truly fixed now, and at least on the process side Intel has maintained for the last two years that it has remained on schedule, then going forward data center leadership should roughly be a function of two things:
- The process leadership of the foundries that produce the CPUs;
- The time gap between initial process introduction and data center CPU product launch.
In this regard, Intel has a compelling roadmap on both points. First, Intel will be less than a year behind TSMC’s 3nm with its own 3 process, before leapfrogging TSMC in 2024 with 20A and 18A. Secondly, Sierra Forest and Granite Rapids are in fact the lead products for Intel 3, significantly reducing the time gap that traditionally existed for the data center group. Although not confirmed yet, if Diamond Rapids in 2025 is produced on 18A, it could also be one of the lead products for that node (although Intel’s recent disclosures indicate the actual lead product could be Lunar Lake, the time gap between Lunar and Diamond Rapids would still be quite small if it launches in 2025).
For AMD, it could in principle have exploited Intel’s process leadership deficit significantly more if it had moved to 5nm much quicker. Similarly, it seems it will take AMD again at least a year or more to move to N3, although perhaps not two years as with N5.
Hence, referring to the two points above, going into 2025, Intel will have reclaimed process leadership from TSMC, and it also seems that Intel will move its data center portfolio quicker than AMD to its leadership nodes, further increasing the effective advantage.
Intel’s new data center executive, Sandra Rivera, has talked a few times about “roadmap simplification”: apparently there were too many products in development. However, recently she also talked about Intel’s plan to ramp Emerald Rapids already in the second half of the year, followed by Granite Rapids in 2024. This means there will the three Rapids in two years, in addition to the new Sierra Forest roadmap in 2024. This is a cadence of product introduction that may still worry some investors, although it is obviously required to catch up.
In addition, there are some questions about what the competition will do. For a long time, it has/had been expected from rumors that Zen 5 Turin on N3 would double the core count to 196 (and 256 for the Zen 5c version), which obviously is an aggressive move that could make it difficult for Granite Rapids to compete against. However, from AMD’s disclosures it has since become clear that there will also be Zen 5 CPUs based on N4 (which is basically a 5nm-class node). It is not known if this will apply to Turin: in the worst-case it has received a spec downgrade to reduce its time to market, which may or may not be detrimental or beneficial to Intel (if for example this would push the “true” 196-core Turin to 2025).
As a last note, some close followers of Intel will likely have seen some recent rumors regarding Sierra Forest and its core count. The first rumor stated that Sierra Forest would have 334 cores, with a version with 512 cores also being in development. However, the same leaker then followed this up with another rumor that those version had been cancelled, apparently because a key customer was threatening to switch to AMD if it didn’t launch by Q2 2024.
Overall, this information lines up with a leaked slide that I coincidentally found via Google Images a while ago. This slide revealed that Intel had (originally) both AP versions of Granite Rapids and Diamond Rapids in development, and only an AP version of Sierra Forest. AP stands for Advanced Performance, and is a higher performance product line originally introduced with Cascade Lake in 2019 (Cascade Lake-AP consisted of two glued-together Cascade Lake-SP chips).
Instead, after the roadmap simplification, according to the rumor just discussed, while there are still AP versions of Sierra Forest and Granite Rapids in development, the AP version of Sierra Forest would consist of two glued-together Sierra Forest CPUs. Obviously, this is large competitive change from a purported 512-core Sierra Forest CPU, although the difference compared to the 344-core version would be smaller.
Over the last several weeks and months, Intel’s dividend has been a much-discussed topic on Seeking Alpha and elsewhere. Intel has now finally caved in and reduced the dividend by 66%.
My own opinion on the matter had been two-fold:
- When looking at Intel’s historical cash from operations performance, there should be no question that Intel would have any trouble paying the dividend; a single or even a few years of negative FCF during an investment period will be well compensated for once Intel’s returns to more historical financial performance (in other words, one should look at the average financial performance). In the worst-case Intel could fund the dividend by taking on some debt (which Intel has actually done recently).
- Nevertheless, given the very broad measures Intel has taken in response to the current macro environment (exiting non-critical businesses, lay-offs, capex reductions, employee compensation reductions), a reduction in dividend would certainly not be unwarranted, especially since the current financial decline coincides with the investment period (in order to return to tech leadership). As Pat Gelsinger said, Intel shouldn’t be both a top dividend payer and a top capex spender.
Overall, this fits into what had already been the Intel investment thesis for a few years, which is that it is a turnaround stock. Obviously though, initially the turnaround seemed merely to be technological (return to leadership in 2024-2025), but the macro downturn has also added a financial component.
However, for prospective investors, INTC stock’s decline over the last year means that the current downturn is already baked in, so there is no fundamental change due to the dividend reduction. As the demand (and inventory burn) environment returns to more historical levels, never mind if some of the new businesses like foundry start to scale, then there should be no reason why Intel couldn’t restore the dividend over time. In addition, such a recovery in financial performance should also cause a recovery in stock price (although partly offset by higher spending such as in R&D).
This was in fact exactly the thesis I already laid out last year, which is that at the current stock price, Intel isn’t even a turnaround stock anymore, but simply a bet on a financial recovery (“cyclical”): Intel Stock Bottomed And There Is Upside Potential (NASDAQ:INTC). In that sense, the expected turnaround only increases the potential stock returns. (In the article just mentioned, I made the distinction between a base case of $50-60, and bull/turnaround case of $90+.)
Intel held an investor call in conjunction with the dividend update, which has been summarized in the tweet below.
One of most important points from the call was the distinction between the financial performance (which is mostly dependent on the demand environment, and to lesser extent on market share shifts) on one hand, and the actual technological turnaround on the other hand. It is the latter where Intel said it is starting to see “green shoots” emerging. Simply put, Intel yet again reaffirmed that the process roadmap remains on track. Pat Gelsinger also in no ambiguous terms denied the recent Arrow Lake rumors regarding the delay of the TSMC N3 tile, saying that Intel’s execution will prove those rumors false.
Another interesting tidbit was that Intel said 40-50% of this year’s $20B+ gross capex goes towards the shell buildouts. This essentially confirms the whole discussion and investment thesis (as discussed here, in the previously mentioned article and at the investor meeting in February 2022), which is that Intel is investing for both a return to technology leadership, but that this will also serve to position Intel for long-term revenue growth.
Essentially, if Intel was only investing for its current business needs (i.e. it wasn’t investing in fab shells to anticipate future growth from areas such as its foundry business), then its 2023 capex could literally be ~$10B less (and that number likely could have been even be higher still if Intel hadn’t been required to do the 5 nodes in 4 years accelerated cadence to catch up from the 7nm delay). Obviously, a $10B reduction in capex would more than pay for the dividend, which more or less “proves” the point above that Intel in principle (on average) should be able to pay a ~$6B annual dividend.
To illustrate the point about long-term growth, the current 2023 revenue estimate for Intel is just above $50B (this is indeed just an estimate as Intel has not provided a full-year outlook), while Intel has been aspiring to reach $100B+ in revenue over time. Obviously, $100B revenue at a 20% FCF yield (according to Intel’s long-term model from the investor meeting), which compares to a current market cap not much above $100B, suggests a favorable valuation, confirming the investment thesis.
To be sure, the comparison of Milan(-X) vs. Sapphire Rapids, while interesting, was purely academic/counterfactual. Due to its ultimately self-inflicted wound, Intel remains in a position where it is meaningfully behind in most general workloads (although in some relevant cases it has a compelling proposition). This was clear from the absence of any Intel-provided competitive benchmarks, and confirmed by third-party reviews. This means AMD maintains the data center leadership that it has had since 2019.
Roughly speaking, with Genoa AMD increased the core count (and in fact also its pricing) by 50%. Since Intel did the same with Sapphire Rapids, the competitive landscape has not significantly changed compared to what it has been for the last two years (which admittedly implies Intel will continue to lose market share). As discussed, the picture gets murkier when considering the many accelerators introduced with Sapphire Rapids. For example, in AI inference clearly Sapphire Rapids is the only sensible option, vastly outperforming even the 96-core Genoa.
Still, note that until just a few months ago the 64-core Milan-X (on 7nm) with a gazillion megabytes of cache was AMD’s flagship CPU, and the 60-core Sapphire Rapids (on Intel 7) has dealt with it with relative ease. This means if not for the execution timeliness (the hundreds of bugs) Sapphire Rapids would have been a rock solid product. So the more general point is that with Sapphire Rapids finally launched and the execution issues (roadmap) largely de-risked (given the Granite Rapids among other milestones Intel has discussed in recent quarters), this should be a stepping stone towards retaking leadership in 2024-2025.
Put differently, it was only in the last year that we really got to know how awful the health of Sapphire Rapids truly was, but with this product out of the way, going forward Intel should again be bottlenecked by its process tech roadmap instead, which still looks like a leadership one going into 2024 and beyond, paving the way for data center leadership once again.
The dividend reduction should be seen as a symptom from the current macro/demand environment, not a cause. Due to this environment and the resulting decline in stock price, Intel has actually become a potential investment based on the cyclicality of semiconductor demand, rather than (or in addition to) being purely a turnaround stock.
In the former case, a recovery implies the stock price could double, whilst in the latter case the stock could potentially triple or more.
For reference, Pat Gelsinger’s unofficial target at the February 2022 investor meeting when the stock price was still close to $50 was to double both the earnings and the P/E multiple, suggesting a $200 stock price, with Dave Zinsner saying that the financial model would at least accomplish the doubling in earnings. Although that financial model also implied a level a revenue (likely well over $100B) that seems quite ambitious currently (to say the least), this expected recovery and growth is exactly why Intel keeps going forward with building out the shells, although in some cases (such as the German fab) at a lower pace.
Disclosure: I/we have a beneficial long position in the shares of INTC either through stock ownership, options, or other derivatives. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.