AMD 2024 Q2 Financials: It's all AI++
Analysis and Analyst Q&A
If you've been following AMD's financials the last few quarters, there have been a number of key trends in play. The biggest is the piece of the AI pie that the company is getting - launch of its MI300X series of GPU/AI accelerators has occured and ramping through from Q4 2023 into 2024. Throughout the last few quarters, AMD has been increasing its 2024 revenue estimate from the MI300 series from $1b to $3b to $4b and now $4.5b. It's a crazy number, given that AMD's entire revenue in 2015 was under $4b. But even with today's numbers, it showcases AMD has a strong growth trajectory. We'll go through the numbers and the trends in each business unit.
Headlines:
💵 Overall Revenue, $5.835b, up 9% YoY from $5.359b and up 7% QoQ
GAAP:
📈 Gross Margin at 49%, up 3% YoY and up 2% QoQ
💰 Op Income at $269m, up from $-50m YoY and up from $36m QoQ
🪙 EPS $0.16, up 700% YoY, up 129% QoQ
Non-GAAP
📈 Gross Margin (non-GAAP) at 53%, up 3% YoY and up 1% QoQ
💰 Op Income at $1.264b, up 18% YoY and up 12% QoQ
🪙 EPS $0.69, up 19% YoY and up 11% QoQ
Outlook is as follows:
💵 Q3 Revenue, $6.700b, +/- $300m
📈 Q3 Gross Margins (non-GAAP) of 53.5%
If we're looking on the face, AMD beat market expectations here slightly, but going historically, that Q2 2022 high is a hard one to reach. During that time demand was going wild (sometimes referred to as the 'Chip Shortage'), but also AMD had no sizeable AI income at that time. This is why going into the individual business unit numbers are going to be so interesting.
Also a word on that outlook. $6.7b in a quarter would be AMD's best quarter as a company, ever, in its history. It's slightly above market expectations for Q3 (GMs are about where market expects), but AMD is continuing to improve its prediction for AI hardware revenue as it executes through the year. That revenue is 2H heavy, and as you'll read, AMD hit $1b revenue for Instinct for the first time in Q2, so there's still a ramp through the second half in that segment.
This is where we stand with each business unit.
Data Center - EPYC, Instinct
➡️ Revenue $2.834b, up 115% YoY from $1.321b
➡️ Operating Income $743m, up from $147m YoY
➡️ Operating Margin 26% , up from 11% YoY
This is a record quarter for data center, and by sheer will it slightly misses out on hitting 50% of revenue (it's actually around 48.5%). In the past we've seen companies talk about 50% DC/embedded and 50% everything else, well AMD has hit that very easily and then some.
AMD spoke about hitting $1b revenue for Instinct for the first time, but also the growth is driven by the strong market resonance with both EPYC and Instinct being deployed at hyperscalers and enterprise/tier 2 clouds. Microsoft is showcasing that MI300X is cheaper in GPT inference than the competition, and has stated that a lot of co-pilot workloads today are running on MI300X. The DC segment is expected to grow through the rest of the year, and AMD says it sees opportunity to improve EPYC market share (>30% today) as well as the expected AI GPU growth.
In the last quarter, AMD announced the MI325X, a higher memory version of the MI300X set to come to lead customers end of Q4 but properly ramping through 2025, followed by a new CDNA4 architecture in the MI350 series in 2025. The big number here is AMD expecting a 35x inference performance increase from MI350, showcasing that they're promising big. In the Q&A after the financials, CEO Lisa Su said that MI350 will be a family of parts targeting different markets, but the aim is to drive adoption of the hardware/software stack given that a portion of MI350 is going to be platform compatible, enabling those who develop today to deploy when they get the hardware.
Client - Ryzen, Ryzen Mobile/AI
➡️Revenue $1.492b, up 49% YoY from $998m, small gain QoQ vs $1.368b
➡️Operating Income is $89m, up from $-69m YoY
➡️Operating Margin is 6%, up from -7% YoY
Even though this quarter shows a massive uptick in client revenue, next quarter's YoY numbers will be muted by comparison as from Q2 2023 to Q3 2023, AMD gained around 50% revenue in client through new product launches. Nonetheless, AMD is still seeing strong pickup around Ryzen 7000 and Ryzen 8000 series in desktops and notebooks respectively.
The launch of the new Zen 5 generation of products has started recently. The Ryzen AI 300 series (Strix Point) for notebooks, with built-in AI and Microsoft co-pilot support, was literally launched at the weekend, and so revenue from those will start to be visible in Q3 but better in Q4 as a wider range of hardware comes to market. For the desktop, new Ryzen 9000 series are launching in a couple of weeks, also built on Zen 5. There is technically a 1-2 week delay from the initial announcement on these parts, however when speaking to AMD they said they expect the delay to be immaterial to the balance sheet even in the short term. Again, AMD is expecting good pickup on its desktop parts, as launching in Q3 according to the CEO has 'historically been good' for them.
Gaming - Radeon, Consoles
➡️Revenue $648m, down 59% YoY from $1.581b
➡️Operating Income is $77m, down from $225m YoY
➡️Operating Margin down 2pts to 12%
Numbers for gaming are down substantially. It might be best to represent this with a graph.
AMD is saying that this is primarily due to its gaming semi-custom sales - we're entering the fifth year of the console cycle, which shows continued softness as expected, however this feels more than that - compared to Q1 2023, there's over a billion dollars difference in those revenue numbers, and I don't think console sales in Y4/Y5 accounted for a billion dollars of revenue. I say this because the Radeon numbers were relatively flat QoQ.
It's worth noting that gaming (GPUs) is often a lower-than-average margin business for most companies, especially compared to AI (GPUs). AMD has been pushing out new models over the last 12 months, but none of the big companies have started talking about a new upgrade cycle in a while now, instead focusing on software updates, some of which are driven by AI.
Embedded - Xilinx, AMD Embedded
➡️Revenue $861m, up from $846 QoQ
➡️But down from $1.459 YoY, down 41%
➡️Op Income $345m, down from $757m YoY
➡️Op Margin 40%, down from 52% YoY
The embedded market, by comparison to the consumer market or a fast-cycle enterprise market, often lags a bit behind. It means whenever there's a boom or a bust, it takes a while for embedded to catch up. Right now, embedded is still in the 'bust' phase the rest of the semi industry had at the end of the pandemic. Even though AMD assumed 2024 Q1 was going to be the low point, Q2 is essentially flat to Q1. The company is still expecting embedded 2H revenues to rise single digits through the year, and gave indication in their notes that ordering patters have showed signs of returning towards normal. That's a small glimmer, but expected.
For those wondering why Embedded is still experiencing this glut in revenues (for example, Lattice has been hit as well this Q), it's because during COVID, when people wanted 10 chips, they ordered 50, expecting to get 5. As volumes shored up and the rest of the 45 chips were delivered, those markets stopped ordering more and over time digested the over-inventory they had purchased. That's still going on in the embedded space. For now.
Here's what all the business units revenues look like since AMD split it out into these categories in Q2 2021. Data Center has now 3x+ revenue in 3 years, while Client is a slight dip, gaming has halved, and Embedded is half its peak post-Xilinx acquisition. AMD's big wins here are in data center, and so the questions all come to whether there is still interest in AI build-out long term. AMD believes it is, and says they still have to due to the potential that the AI market can bring. Specifically, and I guess a lot of us have said this, longer term the revenue in the AI market is expected to be on the inference side of things - so while AMD says it is competitive in AI training, the future of its inference is aiming at being the market leader in power and power efficiency.
Watch this space, it seems.
Analyst Q&A
Here is a rough transcript of the Financial Analyst Q&A. It was written on the fly during the call, and is provided as-is. The only changes made are to my own spelling mistakes. All answers were given by Lisa Su unless marked JH, then the CFO Jean Hsu answered them.
Q: Ben Reitzes, Melius Research - MI300 - how you see it sequentially RoY. $2.8b for annual target left, so Q4? Also 2025, potential for rapid growth?
A: LS: MI300 and customer evolution - happy with MI300 progression. Started year with key point to get product into customer DCs, qualify workloads, prod capabilities. Now in 2H, great progress. In 2H, MI300 continue to ramp in Q3, Q4. Expanding current deployments and a large pipeline of customers getting familiar with arch, software. Overall very pleased with progress, continuing to what we expected from the capabilities from the product.
To 2025, we announced at Computex expanding roadmap. Feeling good, MI325 coming later this year, then MI350 in 2025 competitive with Blackwell. On our way to CDNA next. Bullish on market, market needs compute. AMD HW/SW are getting good traction and expanding that pipeline.
Q: Aaron Rakers, Wells Fargo - On DC, as we look forward FY, thinking about EPYC server CPU growth expectations as we go forward. Any updated thoughts on continue to gain share in server CPU in next 2Q?
A LS: Pleased with EPYC progress. Competitive, 4th gen EPYC is really doing very well. Broad adoption across cloud. Focused on enterprise and 3rd party cloud instances. Nice traction in enterprise with new customers and existing customers. 3rd party cloud is good pickup. Our EPYC portfolio has done well. We feel good for 2H - the market looks like it's improving. Return to spending in enterprise and cloud, positive trend. In process of launching Turin, started prod in 2Q, launch in 2H. See revenue from Turin in 2H. Overall server market and AMD ability to grow share in server is what we see.
Q: Tim Arcuri, UBS: DC GPU roadmap. 325 launching later this year. Does 4.5b revenue include MI325? Speak more about MI350 - there’s a shift towards rackscale for comp - is that what MI350 going to look like?
A LS: Looking at MI325X, on track. Small contribution in Q4 for revenue, but mostly MI300. MI325 will start in Q4, then ramp through 1H25. For 350 series, there will be multiple SKUs from air to liquid cooled, there are people who want rack level and we're doing more in system level integration, we'll invest more in that, but we also have customers who want current infrastructure. MI350 beauty will fit into same infra as MI300 series. Will lend itself to easy ramp. We see range of options, expanding roadmap that we're planning.
Q: Ross Seymore, Deutsche Bank: For client - talk about AI PC, how is AMD positioned? Any competitive intensity changing with AMD systems to 2H seasonality?
A: Pleased with client results. Strong roadmap. Zen5 products, notebook and desktop in mid year. Very positive feedback on products - we launched the first notebooks over weekend, the reviews are very positive. Our view is that AI PC is an important add to the category - for 2H we have better seasonality and we can do above typical seasonality based on timing. In 2025, see AI PC across larger set of price points for more opportunity. PC Market is a rev growth opp. Business is performing well, products are strong. Working with ecosystem and OEM partners for strong launches.
The PC market is a big market, we're underrepresented. We take all comp seriously, but our products are very well positioned.
Q: Matt Ramsey, Cowen - Draw parallel Instinct and what you did 5-6 years ago with EPYC. When Naples launched, there was a lot of reaction and sentiment around roadmap. If I remember, what was most important that was the insert into the market for leverage 3-4 gens forward. Is that a parallel with Instinct? Can you talk about engagement?
A: We're very pleased with Instinct roadmap. This is a long term play, it has a lot of parallels to EPYC. You gain more opportunity, broader workloads, better deployments gen on gen. Conversations with customers - in the near term, we had key milestones to pass in 2024, related to hardware volume in hyperscalers and large tier 2 customers - we've done that. Our software has matured and is in a lot of deployments - ROCm has features, functions, getting to perf with customers, we've got confidence and learned a lot. Building out system level and rack scale is where we're investing more in. Longer term conversations is also really important - all things have progressed well for MI300 but we have a lot more to do. Roadmap will help us do that to open up opportunity over next couple years.
Q: Vivek Aria, BOA Securities - Industry debate on monetization and ROI on CAPEX. Today they have options - GPUs from NV and software, or do custom chips, or AMD. How does this play out in 2025 - will customers have concern to consolidate CAPEX around others - how is your visibility given the debate and how will AMD will carve a debate?
A: You talk to a lot of the same people we do. Overall view on AI investment is that the industry has to invest given the potential being so large, impacting how enterprise operates. The investment cycle will continue to be strong. For the size of the market, I believe there will be multiple solutions - GPUs, custom chips, ASICs. In our case, we have demonstrated a strong roadmap to partner well with customers. From standpoint from hardware/software co-opt is important. For LLMs, GPUs are still arch of choice. Opp is large, and AMD piece of that is strong tech with strong partnerships with key market makers.
Q: Joe Moore, Morgan Stanley - MI300, talk about training vs inference. Initial focus was on inference, any traction on inference and split over time?
A: At launch, lots of good characteristics on it is memory bandwidth and memory capacity is leading the industry. Early deployments have been inf in most cases and fantastic perf. We have customers doing training, and from that standpoint the software has been optimized and we continue to ramp training over time. Going forward, inference > training from marketing, but from AMD, inference and training is growth for us.
Q: Toshi Ahari, Goldman Sachs - MI300. If you're shipping to demand, or updated forecast is supply constrained? Last Q, comments on HBM and CoWoS. On profitability, on MI300, talked about accretive and improving over time as you work through kinks. Has that view changed given competitive intensity and need to invest (organic or acquisition), or confident profit margins will continue to expand.
A: On supply side, AMD made great progress in 2Q. Supply a lot, >$1b in Q. See line of sight to increasing supply. The chain is tight, and remain tight through 2025. We have great partnerships across the chain, building cap there, and continue to ramp through 2024. Continue to ramp supply and growth opportunists, and accelerate customer adoption. See how it plays out 2H.
A: JH: Team has done good job to ramp MI300, it's a complex product. The team start to implement operational optimization to see GM improvement. Will be accretive to corp average. We are ramping revenue. Business model can leverage it, because rev ramp has been significant, op margin rise has been significant.
Q: Stacy Rasgon, Bernstein, Guidance for Q3 - $1b growth revenue across DC, client, embedded. How is that split across? 70% to DC, 20% to client?
A: The gaming business down double digit - think of it DC is the largest piece, then client, then in embedded, single digit sequential growth.
Q: How does the DC split? Instinct, EPYC?
A: We will see Instinct growth and nice growth on server side.
Q: Harsh Kumar, Piper Sands: Difference to competitor is rack scale product - when is UALink going to close that gap?
A: Overall, talking about system evolving, there's no question the systems are getting more complex. With large clusters. Our customers need help putting them together. Includes Infinity Fabric solutions, and general rack level integration. We're very pleased with all the UALink partners, but we have all the pieces have already under AMD umbrella with IF, with Pensando, and you'll see us invest more in this area. This is how we get customers to market faster. The CPU, GPU, networking, and system level solutions.
Q: Blaine Curtis, Jeffries: MI300. Can you give makeup of 1H customers - is there still a government contingency. On Software, the ramp of software, challenge of ramping customers. What's the biggest hurdle for next wave?
A: Supercomputing piece - was mainly Q4, and a bit in Q1. In Q2, that's almost all AI. That's MI300X, for large AI hyperscalers and OEM customers going to enterprise/tier 2 DCs.
In terms of the pieces we're doing - on memory, there's noise in the system. I wouldn't pay attention to the noise in the system. This has been an incredible ramp. The most impressive ramp we've ever done. $1b in 2Q, and ramping through Q3 and Q4. For memory, we have multiple suppliers qualified on HBM3. It's a tricky business, we've done it very well. We're also qualifying HBM3E with multiple suppliers.
The exciting part of this, the ROCm capability has gotten better because customers have been using it. Out of box perf, get a customer up and running, depending on the software companies are using like Pytorch, we can be out of box running well in small number of weeks. That's expanding the portfolio. We'll continue to invest in software, hence Silo.ai acquisition. These are 300 engineers that have experience with AMD hardware, and will get customers up and running on AMD hardware. Talent like Silo.ai, nod.ai (compiler talent), and we're hiring organically. We see leverage in the model, but we'll continue to invest. We have the pieces, this is building out scale.
Q: Tom O'Malley, Barclays - Client in 2H. Above seasonal, sept-dec. Launching new products, but also AI PC. Where are you seeing above seasonal trends - ASP, volume, AI PC?
A: Launching Zen 5 desktop and notebooks ramping in Q3, that's the primary reason. AI PC is one element, but the overall refresh, and desktop launches in Q3 are usually good for us. The products are well positioned.
Q: Chris Danley, Citi - GMs. incremental GM guidance for Q3 is dropping, is there a reason. Have you changed MI300 GMs as accretion point moved out?
A JH: Made a lot of progress to expand GM from 2023 at 50% to guided 53.5Q for Q3 2024. The faster DC growth is the main driver. Now it's close to 50% - that fast expansion helped with GM. For 2H, we expect DC to be the main line growth. Some other puts and takes as well - PC business will do better in 2H, and typically consumer focused. Also embedded business will be up each Q, but recovery is more gradual. So balance, that's why we see pace of GM change a little bit, but see continued expansion. For MI300, confident it will be accretive to corporate average. DC business will be driver of GM expansion.
Is AMD stock expensive or not? I checked the historical prices from 2020 to 2024. It has a very small growth and it’s turn over is low either. Compared to NV stock , it is still expensive with E/P almost 200. How do you think about the price?
Very nice summary of financials and breakdown of revenue per business unit quarter over quarter 👍