Written by

Arne verdhye

8.22K Followers

AMD Vs. Intel: I Was Wrong About AMD

Dec. 23, 2021 11:19 PM ETAdvanced Micro Devices, Inc. (AMD), INTC311 Comments29 Likes

Summary

• I recently considered AMD’s server roadmap to be too weak. However, I made the mistake of only using Intel rumors and not AMD rumors.

• AMD’s 256-core Turin (Zen 5) could spoil Pat Gelsinger’s "unquestioned leadership" plans. Hence, I reverse to neutral on Intel: let Intel ship their chips first.

• In the long term, AMD’s success is dependent on TSMC. Conversely, process leadership is a requirement for Intel to be able to overtake AMD.

• Intel has a solid PC roadmap, but until it regains process leadership (in 2025+) and shows proof of being a "data center-first" company, Intel might at best trade sideways.

Andrei Berezovskii/iStock Editorial via Getty Images

Investment Thesis

I recently detailed AMD’s (AMD) data center announcements through 2023. I noted that given AMD’s sluggishness to move to TSMC's (TSM) 5nm, AMD was giving Intel (INTC) an opportunity to catch up, and perhaps overtake AMD even earlier than incoming CEO Pat Gelsinger promised investors upon his return to Intel (2024-2025).

However, upon second thought, I have reconsidered my stance. Although Intel (as per my original analysis) indeed seems to be moving quicker than AMD, that means that going forward Intel will simply be (much) less behind… but without evidence yet that Intel will actually reclaim performance leadership.

More specifically, in the second part of the original article, I considered some Intel rumors (to make the case Intel could not just catch up, but also overtake AMD), without also considering further AMD rumors. However, by also considering the AMD rumors, here I will revisit the previous analysis (which others have also done already).

In conclusion, AMD’s roadmap likely was stronger than I had anticipated in the previous article. So although I remain bullish on Pat Gelsinger (although I have consistently warned investors that this could take years to yield results), I have somewhat adjusted my timing expectations for the results to become visible (in the data center).

Source: Twitter

Background

Although frequent readers of Seeking Alpha may not assume this, in the past I have actually been quite critical of Intel’s data center execution. For example, I discussed this at length in Ice Lake-SP Tells A Cautionary Tale For Intel Investors.

One example I have used numerous times is that Intel delayed its Granite Rapids Xeon from Q1’22 by at least one year in the wake of the 7nm delay, despite that Intel said 7nm had been delayed by just 6 months. (For comparison, Intel’s other 7nm CPU, Meteor Lake for PCs, seems to have been delayed by just 3-6 months.) It is not clear why Intel would delay its data center CPUs much more.

If Granite Rapids had launched on schedule in Q1'22, then it would have been a clear leadership product, since Milan-X would have had vastly lower core count, a larger process node, and an inferior overall feature set (lacking PCIe 5.0, CXL, DDR5, etc.). Then, later in 2022, AMD only would have been able to reduce its loss of leadership (not reclaim leadership) with Genoa. This, indeed, confirms Pat Gelsinger's statement that Intel's manufacturing is a rising tide that lifts all boats when it delivers, but conversely also drowns all boats when it doesn't.

Still, this 12-month delay also means Intel’s subpar execution in the data center can’t be fully explained by the process technology delays alone. For a company that said (in 2017) it would become "data center-first", this execution indeed should be somewhat worrisome for Intel investors. Indeed, going back to at least 2013 (and up to November 2021), Intel has consistently touted the data center as its big growth driver, but I have consistently identified this segment to be its weakest in terms of product execution.

More philosophically speaking, however, an analyst/investor can't just look at Intel's past and extrapolate, but has to judge each company's roadmap. From that view, Intel remains stuck between its poor 10nm execution and a supposedly improved roadmap going forward. So as one of the bullet points said: let Intel first deliver a strong CPU on schedule again.

Turin: AMD’s 256-core monster CPU

The thesis of my previous coverage was that investors should regard the new Zen 4c core in the 128-core Bergamo as a competitor to Intel’s Sierra Forest CPU, which is rumored to contain at least 128 cores too (and potentially well over 200 cores). The Sierra Forest CPU would have a socket with over 7K pins, which I pointed out was an indication that Intel had big plans for this line of CPUs.

However, since Intel has not officially announced this CPU, some readers pointed out I was using mere rumors to base my conclusion on (although the rumors came from independent sources). So if I was going this route, then I realized I should also have considered the AMD rumors.

This leads us to Turin. Not the Italian city, but the Zen 5-based data center CPU from AMD that will succeed Genoa and Bergamo. The title says everything one needs to know about this CPU:

AMD EPYC Turin Zen 5 CPUs Rumored To Feature Up To 256 Cores & 192 Core Configurations, Max 600W Configurable TDPs

Based on the information AMD provided at the recent event, we can now infer why there are two configurations for the core count: the regular Zen 5 will have 192 cores (2x Zen 4), and the Zen 5c core will have 256 cores (2x Zen 4c). What is noteworthy about this is that AMD will apparently manage to double the core count (2x), despite that TSMC is only expecting a 1.6x shrink from 3nm. (Obviously, this is possible given that AMD’s CPU consists of relatively small chiplets, which have room to become larger: this also explains the rise in TDP.)

If one steps back, just a few years ago, AMD took the data center by storm by announcing a 64-core CPU when the current maximum from Intel was 28 cores. Looking forward, in 2024, AMD will launch 192-core and 256-core CPUs, so clearly this continues a steady increase in core count roughly in line with Moore’s Law. From that view, I was previously too bearish about AMD’s data center roadmap.

To be sure, some AMD fans have proposed that AMD would launch Zen 5 in 2023, but this seems exceedingly unlikely given the Genoa and Bergamo launches in late 2022 and early 2023: if it takes AMD over three years to move from 7nm to 5nm, why would AMD suddenly be able to move to 3nm in just one year or less? Some recent (unconfirmed/contested) rumors indicate that AMD may use Samsung’s 3nm, but that wouldn’t change much.

In conclusion, to AMD’s credit, AMD has executed very well in the data center, in contrast to Intel. This is best illustrated as follows: AMD has made sure to always launch its data center CPUs within at most a few months of launching its desktop CPUs (in stark contrast to Intel’s 6-18 months). Nevertheless, Zen 5 should firmly be regarded as a 2024 product, which does mean it should be compared to Intel’s 2024 roadmap.

Intel roadmap

As noted above, Intel announced in 2017 that it would become a “data center-first” company. Indeed, one can find quotes from Intel executives where they promised to launch the 10nm++ and 7nm Xeons "first" (before any PC CPUs).

Instead, Alder Lake has already launched (ahead of Sapphire Rapids), and the 7nm delay caused Granite Rapids to be put behind Meteor Lake. Hence, to date the pattern still remains that Xeon tends to launch many quarters after the PC Core CPUs, despite that the data center is now a ~$1B per quarter business for AMD (growing at triple digits) and despite the data center explicitly being Intel’s biggest growth engine.

This leads to the core thesis of this article: if Intel’s execution in the data center remains subpar, then Intel is at risk of losing more market share to offset any general market growth. And if the data center isn’t growing, which other segment is going to move the needle for Intel’s growth? None.

With that said, investors should note that Intel went through a CEO transition in early 2021, and in mid-2021 the EVP of the data center (from 2017-2021), Navin Shenoy, was removed from Intel in a reorganization of the data center. From the discussion in this article, it is clear that a reorg was indeed required.

In any case, I will revisit my summary graph from previous article to make it more rigorous. I have modified two things. First, I removed Sierra Forest since the rumors are not unambiguous about its core count. Investors should just keep in mind that Intel seems to be working on its own ultra-high core count CPU, but will need to see more proof first before investors should jump on any bandwagons. Secondly, I extended the projection to 2024, now incorporating the Zen 5 rumors. Clearly, Intel Rapids suddenly doesn’t seem to be so rapid anymore.

Source: Own Work

Has Intel lost?

Of course, the above graph needs to revised further (with less official information), since it is known that Intel plans to launch a successor to the 112-core Granite Rapids in 2024; I already indicated Turin should be compared against Intel’s 2024 roadmap. Intel’s 2024 CPU is known as Diamond Rapids.

Source: Moore's Law Is Dead (YouTube)

Although nothing is currently known about Diamond Rapids, the rumor mill has already run wild about what kind of performance increases it could bring. I have included one such rumor above just as illustration. Furthermore (and perhaps more interesting), another leak (from an internal Intel analysis in late 2020) shows that either (1) the above rumor may be true, or (2) Intel may have severely underestimated AMD:

Intel is very confident about Diamond Rapids and believes it will be their most competitive product in years. At this stage, Intel likely believes that neither AMD or ARM based server CPU providers can compete.

In any case, without further credible (and quantitative) information, the only estimate that one can make is to observe that (in the most bullish scenario) Intel will move from Intel 4 (Granite Rapids) in 2023 to Intel 20A (Diamond Rapids) in 2024. This should allow Intel to double the core count, and leads to following further revised roadmap:

Source: Own Work

As such, the outlook in the data center for next few years is just what one wants to tell himself:

1. AMD bulls will say that AMD was clever to introduce the Zen c-core. Even if Intel quadruples its core count from 56 in 2022 (Sapphire Rapids) to 224 in 2024 (Diamond Rapids), AMD will still be ahead (256 cores).

2. Intel bulls may neglect the c-core version of Zen 5 and compare a potentially 224-core Diamond Rapids to the 192-core Turin and observe that Intel is on track to regain substantial competitiveness.

3. Alternatively, the independently confirmed Sierra Forest rumors indicate that Intel plans to introduce a high core count CPU in the data center based on its E-cores. Given how small Intel’s E-cores are, the sky is the limit (e.g. 1024 cores in 2024-2025?). Although this may be exaggerating, those E-cores have already contributed to Intel regaining leadership in the PC with Alder Lake, so it might be just a matter of Intel extending this strategy from the PC to the data center.

There is actually a lot of truth in the statement that the sky is the limit: given that both Intel and AMD have moved to chiplet-based designs, one could in principle just keep adding chiplets indefinitely to further increase the core count.

Upon review, I have to make one more comment though. I realized that it actually isn’t sure yet that Intel will move to 20A in 2024 (in the data center). If 20A misses the launch schedule for Diamond Rapids, then I assume that Intel will outsource this CPU to TSMC 3nm. In that case, both AMD and Intel might launch a 3nm CPU in 2024, further demonstrating that both companies will be quite closely matched. AMD bulls will see this as a sign that nothing seems to prevent AMD from continuing to gain market share.

Pat Gelsinger comments

Pat Gelsinger recently confirmed the thesis in this article. As he described it, the development cycles in the data center are longer, so he expects Intel and AMD to leapfrog each other with their consecutive product launches ("nip and tuck"). So overall, neither Intel nor AMD will be able to maintain a firm leadership position.

Ice Lake wasn't our greatest product, but it closed some of the gaps. Sapphire Rapids closes all the gaps. So we're now ahead with Sapphire Rapids coming in Q1 of next year. And on some of the benchmarks it's close to AMD, on other benchmarks, it's unquestionably the best product. And like the AI performance of Sapphire Rapids is now 50% of the highest end Nvidia specialty processor, right. It is dramatic, it's hundreds of percent better than AMD’s alternative at this point. So we're clearly having some very strong points of advantage with it.

AMD, we expect them to respond next year. So it's going to be a little bit nip and tuck over the next couple of years. And, as you suggests, it's going to be out a couple of years until we're back to unquestioned leadership in the data center. It takes a bit longer in the data center side, the development cycles are longer. And as our process technologies get better and better, our server roadmap takes a little bit longer to take advantage of that.

Although I would expect AMD bulls to question if even Sapphire Rapids will be competitive, for the AMD bull thesis, AMD actually doesn't need to have a huge advantage since AMD still has <15% market share.

Source: Twitter

Long-term implications

In this section, I will discuss (at a high level) the implications in the data center in the long term. At least for Intel investors, this may be instructive given the new management and reorganizations in 2021.

1. AMD is reliant on TSMC

The first observation is that in the data center, AMD’s success is basically solely dependent on TSMC (and its competitiveness against Intel). The reason for this is because data center CPUs can make use of an unlimited amount of transistors: this is primarily seen in the exponential increase in core count.

At first sight, this may seem trivial, and nothing for investors to worry about given TSMC’s execution in the last several years. Nevertheless, as I have discussed previously, this may be changing: TSMC Confirms 3nm Delay: What This Means For TSM Stock.

1. TSMC for the first time in many years has explicitly used the word "delay" to describe one of its nodes. Furthermore, TSMC used "2025" and "2nm" in the same sentence, whereas a two-year cadence would have indicated a (late) 2024 launch. Depending on when in 2025 exactly 2nm launches (early or late in the year), this could signal a second consecutive delay at TSMC.

2. Intel, for the first time in many years, has explicitly used the words "on track or ahead of schedule" to describe all of its nodes until 2025. Intel claimed that it would regain process leadership with 18A in 2025.

Obviously, such changes take many years. So AMD bulls could observe that although Intel will likely improve its competitiveness (at least if it succeeds in Gelsinger’s plan), being roughly on par going forward (a few months ahead or behind) will not dramatically impact AMD’s competitiveness, especially if by then AMD has already captured a significant amount of market share.

For example, a more material process technology advantage would be on the order of 1-2 years, but that means if Intel only reaches parity in 2024, then TSMC must stall completely for Intel to regain a meaningful process advantage even by 2026. Obviously, such far-out timelines are the reason for this article’s more neutral stance towards Intel vs. AMD.

Nevertheless, one more cautionary remark is required. Since AMD became a TSMC customer, the closest it has been to a leading edge customer was about three quarters after initial ramp (at 7nm), but this is increasing to two years at 5nm. So for example, if AMD would be 12-18 months behind TSMC to adopt 2nm, then it would launch a 2nm CPU in 2026, compared to Intel which might launch a 20A CPU (Diamond Rapids) in 2024.

In other words, if Diamond Rapids is a 20A CPU (as discussed, it might also be a 3nm CPU though), then AMD could suddenly be two years behind Intel, since it might not able to move to 2nm itself until 2026. This is exactly what I meant by AMD being reliant on TSMC as a major risk. Hence, AMD not being quick to adopt TSMC's latest tech could allow Intel regain a material process advantage faster than expected.

2. Bullish case for Intel: flexing its financial muscles

If Intel is still so far from anything that vaguely resembles leadership, why would one be bullish about Intel? It ultimately comes down to three main arguments:

1. Intel hasn’t had a highly proficient technical CEO for ages. Bob Swan and Paul Otellini were both CFOs. The latter missed the mobile revolution, while the former was too focused on buybacks and operating margin despite the process technology issues still going on, which resulted in underinvestments in technology. In between, BK successfully transformed Intel’s strategy beyond the PC, but internal politics and cost cutting (perhaps due to pressure from the board) likely harmed Intel, such as the major 2016-2017 reorg. In contrast, besides strong upper management changes since Pat Gelsinger’s re-entry, Intel is also gradually ramping up the R&D spending after many years of stagnation (and conversely the buybacks have been paused). Intel’s main issue has simply been that its roadmap always kept getting delayed, not that its products have been poorly engineered. For example, as Pat Gelsinger said, Intel has already added 6K engineers in just three quarters in 2021, which compares to AMD’s 15K total employee count.

2. As discussed above, Intel's improved technology execution. For example, the chance that TSMC will have 2nm in 2024 is slim, which means Intel could have RibbonFET (gate-all-around) ahead of TSMC. In other words, Intel regaining process leadership is more realistic than investors may currently assume. In addition, Intel is buying up big amounts of leading-edge capacity at TSMC.

3. The main idea in these kind of discussions is that supposedly core count is the only thing that matters. This is a very flawed assumption. Given various differences in architecture and feature set, the actual performance of AMD vs. Intel CPUs is highly variable and more nuanced than the superficial core count metric. This means core count is not a reliable predictor of actual leadership. One common example is the important class of workloads of AI, where Intel can deliver many times higher performance even with substantially lower core count: I have estimated Sapphire Rapids will be an astounding 240x faster than Milan-X. And before one would be tempted to dismiss such differences, one just has to look at how NVIDIA (NVDA) has amassed a ~$800B market cap (at its peak), which for a large part hinges on its performance in AI workloads in the data center.

The 12-month Granite Rapids delay has delayed Intel’s ability to regain unquestioned data center leadership by about 24 to 36 months. So overall, as I have stated previously, I would currently view Intel as AMD in 2013 or so: it is known that Intel has competitive technologies in development, like was known about AMD in 2013, but it still took several years before these products even launched, never mind to be reflected in the stock price.

Source: Twitter

Overall, the bullish scenario of Intel flexing its financial muscles to build more fabs and regain technology leadership remains valid (and as bonus, Intel will also buy a ton of TSMC capacity), but I currently don't see a lot of reasons (besides the low stock price) for investors to rush to buy the stock, since they may have a few more years the opportunity to do so.

Lastly, to come back on the third point above, I would refer to this recent analysis as one example.

something that a number of folks told me at the shows is that while the 3rd generation Intel Xeon “Ice Lake” may not have a core count advantage anymore, some of the accelerators are actually a big deal. The folks telling me this were either large hyper-scalers or vendors that sell to the hyper-scalers.

Risks

Although this article did a poor job in the following regard, as mentioned above, investors should be aware that there is much more to a (data center) CPU than just its core count. A CPU contains literally dozens of additional features, going from security to virtualization, interconnect, acceleration, and more. Additionally, the architecture also determines the performance that can be achieved per core.

Lisa Spelman, Intel VP: AVX-512 is such a great feature. It has tremendous value for our customers that use it, and sometimes I marvel when I read your audience’s comments about the raucous debate they will have about the value of AVX-512. But I tell you, and no joke, that a week and a half or so before launch I was looking at a list of the deal wins in one of our geographies. 70% of those deal wins, the reason listed by our salesforce for that win was AVX-512. Optimization is real.

Secondly, Intel still has the advantage of being able to offer customers a much more comprehensive overall portfolio: CPUs, GPUs, NPUs, FPGAs, DPUs/SmartNICs/IPUs, Ethernet, Ethernet switches, Wi-Fi, silicon photonics and Optane.

Thirdly, as I have pointed out in other Intel coverage, Intel (and perhaps also AMD) is already valued on the stock market based on the expectation that Intel will fail to hit all of its roadmap targets. This removes a lot of further INTC downside risk, but potentially could lead to increased returns if Pat Gelsinger does start to deliver "unquestioned leadership".

Fourth, obviously both Intel and AMD could experience delays from their roadmaps. In particular, many AMD bulls seem to expect that Zen 5 will launch in 2023, for which there is no evidence whatsoever.

Summary

As I see it, both Intel and AMD will be competitive going forward in the medium term; it just depends on the use-case one looks at. Does one just need a ton of cores for simple throughput tasks, then there is actually a company that already offers even more cores than AMD: Arm-based Ampere (led by a former Intel President). Does one need high SIMD throughput for HPC or AI, then Intel AVX-512 and AMX remain unmatched. Does one need security features like SGX or high crypto/compression throughput, then Xeon will also offer much higher performance than Epyc despite the lower core count.

However, as an Intel shareholder, I would not consider it enough for Intel to just be "competitive" in some workloads given that for years the data center has been considered (including by Intel itself) to be Intel’s highest growth market – and despite that, it is clear that Intel’s execution in the data center has been its worst among all its segments.

So given the Zen 5 roadmap discussed, unless Intel comes up with some unexpected products (Sierra Forest?), then AMD shareholders will likely assume Intel could lose quite a bit more market share. Ultimately, Intel shareholders are banking on the Intel Accelerated roadmap (regain process leadership) as well as Pat Gelsinger to make sure Intel solves its longstanding execution issues, but this will likely take several more years.

Investor Takeaway

In this article, I extended the data center discussion to 2024. This has shown that AMD retains a strong roadmap even several years from now, as it plans to launch a 256-core CPU. Intel in likely the best-case would be at 224 cores (2x Granite Rapids, although not supported even by rumors yet).

So although my previous observation that Intel was moving faster (5.6x increase in core count) than AMD (4x increase in core count) seems to remain valid, this still may not be enough to have any chance of meaningfully surpassing AMD. This leads to a few takeaways:

1. The bottom line is that Intel and AMD seem to be fairly closely matched (in core count) going forward in the data center. I highlighted that both AMD and Intel have advantages depending on specific workloads.

2. It is not clear yet how Pat Gelsinger intends to achieve "unquestioned leadership" in 2024-2025. It is on Intel to counter the AMD roadmap and show proof of this in its financial results.

3. Although not further discussed, I previously suggested the Intel Sierra Forest CPU line as Intel’s solution, just like how Intel has finally caught up with AMD on the desktop due to its hybrid architecture (P- and E-cores) with Alder Lake.

4. Ultimately, the core count race in the data center will be determined by Moore’s Law, for which AMD is reliant on TSMC and Intel on its own fabs. This was already clear in the last few years, when Intel was stuck with just 28 cores at 14nm, and more recently had to delay Granite Rapids due to the 7nm delay. Indeed, Intel has only promised to achieve parity in 2024 (and leadership in 2025). This suggests Intel’s earliest chance to achieve a meaningful gap over AMD might be in the second half of the decade. I assume AMD investors will consider this to be a sufficient time gap.

So given the current trends in process technology, in the mid term ultimately it will likely be other factors (such as pricing, sales and marketing, differentiated features) than just the core count that will determine who will have highest market share going forward and the relative market share movements. Intel should be better positioned in that regard, but the Zen 5 rumors are more bullish than AMD’s execution has been on Zen 4, although under the caveat that Zen 5's precise launch window remains unknown. Conversely, AMD bulls should not disregard Intel’s "new" roadmap under Pat Gelsinger too early either.