Advanced Micro Devices’ (AMD) investors know well that Zen 2 CPU core has brought the Company to the pole position of the x86 world – in servers, desktops, and laptops. AMD has been gaining market share rapidly since the introduction of Zen 2 family products. In addition to raw performance, one of the significant advantages of Zen 2 family products is a less known feature called PCIe 4.0.
PCI Express, or PCIe, is the technology that facilitates moving information between various components of PCs, workstations, and servers. PCIe specification, which was first introduced in 2003, has been stagnant at 3.0 level from 2010 to 2017 until PCIe 4.0 specification was ratified. As can be seen from the table below (image from Wikipedia article), PCIe 4.0 doubles the rate at which data can be moved between various components of PCs and servers.
As these standards go, there is typically a lag between when a specification becomes ratified and when products based on that specification proliferate the market. For a new standard like PCIe 4.0 to take off, a critical mass of suppliers in the PC ecosystem must start supplying devices based on the standard. Intel (INTC) typically leads the charge but has fallen behind when it comes to PCIe 4.0. Given Intel’s tardiness, even though PCIe 4.0 was ratified in 2017, most servers and PCs currently shipping are based on the older 3.0 specification. Intel’s delay gave AMD the opportunity to lead the march to PCIe 4 with a family of CPUs and GPUs with the new specification.
Nvidia A100 Has Already Begun Driving EPYC Adoption
In servers, where humongous amount of data is transferred between various systems, PCIe 4.0 delivers considerable value immediately. The need to move large amounts of data at high speeds is especially acute in AI applications which caused Nvidia (NVDA) to adopt PCIe 4.0 in the recent A100 generation of GPU products. Given Intel does not have compatible PCIe 4.0 devices and given the higher performance offered by EPYC, Nvidia picked AMD EPYC CPU to pair with its A100 GPUs for its high-end DGX A100 system.
Since Nvidia announced its A100, there has been a clear trend of AI hardware manufacturers announcing solutions that use AMD EPYC with Nvidia A100. Since Intel does not have a server CPU capable of supporting PCIe 4.0, it is increasingly not a choice for Nvidia A100 based systems. Recently, CRN put together a list of 10 Nvidia A100 based systems that are currently available or will soon be available. As can be seen below, EPYC dominates.
- Asus ESC4000A-E10: 1 EPYC 2 CPU, up to 4 A100 GPUs
- Atos BullSequana X2415: 2 EPYC 2/3 CPUs, 4 A100 GPUs, 4 Mellanox Infiniband adapters
- Gigabyte G492-Z51: 2 EPYC 2 CPUs, up to 10 A100 GPUs
- HPE Apollo 6500 Gen10: 2 Xeon CPUs, up to 8 A100 GPUs
- Inspur NF5488A5: 2 EPYC 2 CPUs, up to 8 A100 GPUs
- Lenovo ThinkSystem SR670: 2 Xeon CPUs, up to 4 A100 GPUs
- Nvidia DGX A100: 2 EPYC 2 CPUs, 8 A100 GPUs, 6 NV Switches, 9 Mellanox Infiniband adapters
- Penguin Computing Relion XE4118GT: 2 Xeon CPUs, up to 8 A100 GPUs
- QCT QuantaGrid D43KQ-2U: 2 EPYC 2 CPUs, up to 8 A100 GPUs
- SuperMicro AS-2124GQ-NART: 2 EPYC 2 CPUs, 4 A100 GPUs
That is a staggering 80% EPYC presence when the most recent AMD disclosed EPYC overall market share is only about 10%. Other than the dominance of EPYC over Xeon, two things to note here are that a) all but one system have two CPUs and b) each CPU is being paired with 2 to 5 GPUs although ratios of 1:2 or 1:4 seem most popular. Note that, “up to” in GPU counts suggests that all the GPUs may not be populated initially. Assuming that two EPYC 2 gets sold for every 4 to 8 A100s on average, the relative opportunity of each company per system at list prices is about $10,000 for AMD compared to $50,000 to $100,000 for Nvidia.
To be more competitive, Intel is expected to start offering PCIe 4 with Ice Lake CPUs in late 2020 and early 2021. However, AMD is likely to introduce its Zen 3 generation of CPUs prior to that. Ice Lake is likely to not fare well against Zen 3. Zen 3 EPYC will not only have higher performance for standard server workloads but is also expected to have strong support for AI applications through the popular bfloat16 format. As a result, AMD is likely to price Zen 3 CPUs at a premium and consequently AMD ASPs are likely to inch higher once Zen 3 based CPUs are introduced in Q4. It is possible that customers and AI hardware manufacturers will move their A100 based systems to Zen 3 server chips when they become available, thus increasing the size of AMD opportunity.
So, how much could this A100 attach rate be worth to AMD?
Per the recent Nvidia earnings call, Nvidia A100 revenues for the July quarter were less than a fourth of overall datacenter sales. That suggests A100 sales could be some number around $400M. Applying the list price ratio above, this could mean $40M to $80M incremental revenue for AMD in a comparable time period (note that AMD’s and Nvidia’s quarters are staggered). Going forward, as A100 ramps, AMD will benefit proportionally. Zen 3 adoption will likely further accelerate revenues.
Since this business has already become a run rate business, it is unclear how much of this upside is already built into the guidance, but it is unlikely that AMD has guided based on the dominance we are seeing.
Nvidia’s Ampere Gaming Cards Will Drive Ryzen Adoption On The Desktop
On the PC and workstation front, data movement needs are not as high as they are on the server side. Nevertheless, these needs jump a bit with the new Nvidia high-end gaming GPUs. AMD is also set to drive the bandwidth needs up with its high-end GPU code named “Big Navi”.
The new Nvidia cards have PCIe 4.0 support in Ampere graphics and gaming solutions. Big Navi is certain to have PCIe 4.0 given that the lower-end Navi which is currently shipping already supports the specification.
Whether one needs PCIe 4.0 to get the best gaming or workstation performance out of Ampere or Big Navi is debatable and depends on the system configuration and applications – higher end desktops and workstations are more likely to benefit. However, with the pervasiveness of high performance PCIe 4.0 NVME memory devices and other peripherals, gamers and professional workstation procurers will be loath to buy systems not capable of PCIe 4.0. There is not much point to buy a top of the line graphics card for several hundred of dollars (some cases over $1000) and high-priced storage and then handicapping them with a weak IO bus.
With Intel lacking PCIe 4.0 desktop CPUs, Ampere buyers are likely to predominantly choose AMD CPU platforms. Big Navi buyers will also primarily purchase AMD Ryzen options. This is not only true for new system purchasers but also for customers who are only looking to upgrade to the top of the line gaming or graphics card. These customers could choose to upgrade from their current Intel based systems even though a GPU upgrade is all they might need. Intel’s disadvantage is likely to get more acute once AMD introduces its newer generation Zen3 CPUs with improved gaming performance on Q4. Similar to the server side, AMD is likely to increase CPU ASPs with the introduction of higher performance Zen 3 Ryzen CPUs.
In terms of opportunity, each high-end desktop system is likely to have a CPU in the range of $500 (high-end Ryzen) to $3,000 (high-end Ryzen Threadripper). The average high end GPU price is likely to be north of $500. In other words, on the desktop, the opportunity for CPU and GPU is likely to be more evenly split compared to the AI server side where Nvidia GPU opportunity dominates.
For this holiday season, it is likely that gamers and high-end workstation users will default to AMD CPUs paired with Nvidia Ampere or AMD Big Navi. Due to this dynamic, AMD will see considerable jump in high-end Ryzen desktop CPU demand as customer adoption of next generation Nvidia and AMD GPU solutions ramp. Ryzen high-end CPUs are likely to fly off the shelves once Ampere and Big Navi start shipping.
What is the incremental impact of PCIe 4 based GPU upgrades on AMD?
It is extremely difficult to put hard numbers to characterize this phenomenon, but a first order approximation is that AMD will enjoy between $200M to $500M in incremental revenues in Q3 and Q4. With Nvidia Ampere launch slated for September 1st, most of this benefit is likely to fall into Q4 – about 20% of the upside may fall in Q3 and the remaining in Q4.
Interestingly, AMD Q3 of year guidance does not seem to reflect the size of this potential upside. Currently, AMD’s guided growth in Q3 is ~$620M and Q4 is expected to be approximately flat with Q3. However, vast majority of this growth can be explained by Sony (SNE) PlayStation and Microsoft (MSFT) Xbox ramps. EPYC and laptop ramp should explain for the rest. It does not appear that AMD’s guidance covers any of the upside addressed here. Consequently, it is possible that the Company will deliver a $50-$100M beat on these dynamics alone in Q3 and $150M to $400M beat in Q4. With Q3 consensus estimates at $2.56B and Q4 consensus estimate at $2.62B, the upside can be hugely material.
Just about the only limitation on what AMD can deliver seems to be wafer capacity at TSMC (TSM). Given the already known tightness of 7nm capacity, it is unclear how much product AMD can get from TSMC to supply demand above guidance.
If AMD cannot deliver additional product, then the chances are high that we will see a nice jump in gross margins as AMD optimizes product mix to prioritize higher ASP products.
Availability of high-end PCIe 4.0 based GPUs from Nvidia and AMD is about to shift CPU market share in some key server and desktop segments from Intel to AMD. The market share shift will move into a higher gear with the arrival of Zen 3 family devices in Q4.
Given the incremental opportunity, AMD’s current Q3 and 2020 guidance appears conservative. Assuming capacity is available from TSMC, AMD is highly likely to outperform guidance in Q3 and Q4 by a very good margin. Considering the COVID-19 related trends, blowout results are possible.
If AMD is capacity limited, we are likely to see smaller revenue upside but a higher gross margin upside.
Our View On AMD: BUY
Subscribers to Beyond The Hype have access to all the linked articles that may otherwise be inaccessible. For timely, cutting-edge insights, analysis and investing ideas in technology, semiconductor, solar, battery, autonomous vehicles, and other emerging technology stocks, check out Beyond the Hype. This Marketplace service gives you early access to my best investing ideas, along with event driven and arbitrage opportunities when they are most edgy and actionable.
Act before prices increase on 10/1: https://seekingalpha.com/instablog/9110881-enertuition/5498111-join-beyond-hype-price-increase
Disclosure: I am/we are long AMDN,INTC, TSM. 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.
Additional disclosure: Short NVDA
Any positions disclosed by the Author are part of a diversified portfolio. Author considers lack of diversification to be one of the gravest errors made by many investors.
Shorting is a sophisticated investment strategy that requires superior investment skills and must be avoided by all but experienced investors with the appropriate skill and wherewithal.