2023-07-11 09:21:05 ET
Summary
- Advanced Micro Devices, Inc. is making strategic moves to capitalize on the growing demand in the AI and HPC markets, unveiling its AI/HPC native GPU solutions, the MI300X, and the OCP-based Instinct Platform.
- AMD's MI300X GPU variant and the OCP-based Instinct Platform are set to ramp up by the end of 4Q23, offering industry-leading memory configuration and bandwidth, potentially outperforming NVIDIA's solutions.
- Despite promising advancements, AMD faces challenges in competing with Nvidia Corporation's CUDA technology, which is the industry standard for GPU-accelerated computing.
Back in May, I wrote an article outlining potential opportunities for Advanced Micro Devices, Inc. ( AMD ) in the AI sphere. Now, after attending AMD's "Data Center and AI Technology Premiere" event and taking extensive notes, it's clear that AMD is making decisive moves to capitalize on the growing demand in the AI and HPC markets. With the unveiling of its AI/HPC native GPU solutions, specifically the MI300X and the OCP-based Instinct Platform, AMD is strategically positioning itself to leverage the opportunities presented by the rising demand for Artificial Intelligence General Compute ((AIGC)).
Event Takeaway: AMD Contests AI GPU Supremacy
AMD's recent "Data Center and AI Technology Premiere" event , which we attended, demonstrated a strong commitment to capitalizing on the expanding AI and High-Performance Computing ((HPC)) market. The unveiling of their AI/HPC native GPU solutions, specifically the MI300X and the OCP-based Instinct Platform, underscores AMD's strategic positioning to leverage the rising demand for Artificial Intelligence General Compute ((AIGC)).
AMD's MI300X GPU variant and the OCP-based Instinct Platform are set to ramp up by the end of 4Q23. This strategic launch corresponds with the increasing adoption of large language models (LLMs) like ChatGPT, which necessitate advanced GPU computation capabilities. AMD's solutions are built on the CDNA architecture foundation, with the latest CDNA 3 employing a new compute engine and the most advanced chiplet design on 5/6 nm process technology and packaging.
The introduction of the MI300X , a GPU-only variant, marks a significant advancement from the MI300A APU solutions launched at CES earlier in the year. The MI300X replaces the 3 Zen 4 Core chiplet with 2 more CDNA3 chiplets and includes more memory configuration. It provides an industry-leading 192GB HBM3 and 5.2TB/s bandwidth. This upgrade positions the MI300X to offer solutions that are more optimized for AI and LLMs.
Compared to NVIDIA's H100 solutions, AMD claims that the MI300X offers up to 2.4x more HBM density and up to 1.6x HBM bandwidth. This substantial increase allows LLM processing to be performed locally and directly on memory, which enhances compute efficiency per GPU and reduces the total number of GPUs needed. These changes speed up compute performance, particularly for inferencing, and help optimize Total Cost of Ownership for AMD's customers.
AMD's Instinct Platform, powered by 8 MI300X on the industry's standard OCP structure, enables faster time to market with minimal changes to existing infrastructure and lowers the overall barrier of adoption. AMD has already begun sampling the MI300A to its CSP and AI customer earlier this quarter, while MI300X and the Instinct Platform are set to kick off sampling from beginning-3Q for delivery both from 4Q23 onwards.
In terms of socket design, the MI300A adopts the SH5 socket, which is different from the Zen 4 Genoa on SP5, but it does not differ in terms of the LGA6096 pin form-factor adoption and similar ASP profile. AMD's choice of a traditional socket methodology allows it to maintain its focus on the sales and distribution of its chip solutions, in contrast to NVIDIA's hardware-related emphasis with its GH200 Grace Hopper solutions. Supply chain conversations suggest that the pricing for MI300A could be below $10,000, significantly lower than NVIDIA's A100/H100, thus enhancing AMD's competitive position in the AI market.
Overall, as showcased at the event, AMD's AI product offerings and strategies emphasize its ambition to become a major player in the AI and HPC markets. The company's focus on performance, efficiency, and competitive pricing, coupled with its strategic partnerships, positions it well to capitalize on the long-term opportunities in the AIGC space.
Shots Fired
The recent showdown between AMD's MI250 and the Nvidia Corporation ( NVDA ) A100 in the realm of Large Language Models (LLMs) has undeniably ignited a competitive spark in the AI industry. The experiment, conducted by MosaicML—a company championing more efficient AI operations—has provided an intriguing glimpse into the capabilities of these two GPUs.
MosaicML , guided by ex-Intel AI Chief Naveen Rao, utilized an out-of-the-box MI250 and made no code alterations to undertake a training run against NVIDIA's A100. The team used ROCm libraries to replace CUDA and leveraged PyTorch 2.0 for this experiment. Impressively, the AMD MI250 delivered performance levels at approximately 80% of the A100-40GB and roughly 73% of the A100-80GB. This result, though not victorious for AMD, is a clear signal of the competitive potential of the MI250.
While Nvidia remains the leader in this head-to-head comparison, the proximity of AMD's performance is noteworthy. The ability to replace CUDA with ROCm libraries and run a segment of a training run for a smaller LLM with zero code changes demonstrates the significant strides AMD has made in providing a user-friendly and efficient solution.
The results of this experiment also underline the potential for AMD's forthcoming MI300x as it prepares to go toe-to-toe with Nvidia's H100 in the next round. Given that AMD's MI250 could already achieve between 73% and 80% of the performance of Nvidia's A100, there is a real possibility that the MI300x could further narrow, or even close, this performance gap.
However, it is crucial to remember that the GPU's speed is not the only factor to consider when evaluating the AI stacks of Nvidia and AMD. Other significant elements—such as software support, ecosystem, power efficiency, and cost—must also be taken into account.
Risks
We have extensively covered CUDA, which we consider to be one of Nvidia's primary competitive advantages that will be hard to rival.
Nvidia's CUDA technology has long been the industry standard for GPU-accelerated computing. It has gained significant traction in various applications, especially in the field of artificial intelligence ("AI"), where it powers most of the current AI software stack. This creates a significant barrier for AMD and its GPUs as they attempt to secure a larger share of the market.
Large companies and AI researchers have heavily invested in CUDA-based hardware and software infrastructure. Moreover, AI researchers are in high demand and often operate under stringent time constraints, making them reluctant to invest time in rewriting code optimized for CUDA to be compatible with AMD's hardware.
AMD's ROCm libraries, designed to be a CUDA alternative, are a step in the right direction. However, they face an uphill battle in terms of ecosystem maturity, developer adoption, and overall performance. While they provide the functionality to run code on AMD GPUs, they currently lack the extensive suite of developer tools, pre-optimized libraries, and community support that Nvidia's CUDA enjoys. This difference can lead to additional development time and potential performance loss, making the transition less attractive to AI researchers.
For AMD to gain significant market share, it needs to overcome these barriers. This could be achieved through a focus on making the process of rewriting code for AMD AI chips as seamless as possible. Offering robust tooling, comprehensive documentation, and strong community support can help alleviate some of the burdens of transitioning from CUDA to ROCm. Additionally, AMD could collaborate with software vendors and the open-source community to build out a more comprehensive ecosystem around its GPUs and ROCm libraries.
Furthermore, AMD could find a niche market where it can outperform Nvidia's offerings. This could involve focusing on specific AI applications where AMD's GPUs have a distinct advantage, or targeting industries where the higher memory configuration and the lower price point of its GPUs provide a significant edge over Nvidia's solutions.
Financial & Valuation
Note: All historical data in this section comes from the company’s 10-K filings , and all consensus numbers come from FactSet.
In a clear divergence from its previous upward trajectory, AMD's Q1 FY earnings report, released on May 3, 2023, exhibited a 9.1% year-over-year (y/y) decline in revenue to $5.35 billion. Even though the figures were in line with consensus estimates, the market reacted negatively, resulting in a 9.2% drop in stock price. Despite maintaining a decent gross margin of 50%, the operating margin was notably down at 20.5% compared to 31.2% from a year ago. Moreover, the EPS for the quarter, although beating consensus by 6.5%, was down by a significant 47% y/y to $0.60.
A closer look at the financial trends over the past three fiscal years shows a compounded annual growth rate of 51.9% in AMD's revenue. Yet, the sell-side consensus forecasts a contraction of 2.2% in revenues this fiscal year, taking it to $23.1 billion. However, they predict an optimistic growth of 18.6% in the following fiscal year, taking it to $27.4 billion. This shows a mixed future outlook, with potential short-term struggles but a longer-term recovery in revenue growth.
Furthermore, over the past three fiscal years, AMD's EBIT margin has impressively increased by 14.4% points, from 12.5% to 26.8%. Despite this, the consensus forecasts a contraction of 459 basis points this fiscal year to 22.2%, before a predicted expansion of 576 basis points the following fiscal year to 28.0%. This fluctuation in the EBIT margin could potentially indicate unstable profitability in the short term.
In terms of capital management, over the past three years, AMD spent 3.5% of its revenue on share-based compensation ((SBC)). This has resulted in the diluted outstanding common shares increasing by 27.3%. Despite these dynamics, AMD's EPS has outpaced its revenue growth, growing at a CAGR of 76.2% over the past three fiscal years.
Looking at free cash flow ((FCF)), consensus estimates forecast it will reach $4,032 million this fiscal year, marking a strong 17.5% FCF margin. This is a considerable jump from four fiscal years ago when the FCF was a mere $282 million, with a 4.2% FCF margin. Over the past four fiscal years, the company has generated an average FCF of $1,856 million, which is an impressive average FCF margin of 12.6%.
Capital expenditure as a percentage of revenue has averaged at 2.8%, suggesting that AMD operates a capital-light business model. In terms of financial health, AMD boasts a robust balance sheet , with net cash of $3,472 million. However, AMD does not pay a dividend, which could be a point of concern for income-oriented investors, especially when compared to the S&P 500's dividend yield of 1.5%.
As of now, AMD trades at $113.58 per share, with a market value of $182.9 billion and an enterprise value of $179.4 billion. Relative to the S&P 500, AMD is trading at substantial premiums across all key valuation multiples: an EV/Sales premium of 179.5%, an EV/EBIT premium of 40.0%, a P/E premium of 50.8%, and an FCF premium of 51.7%. Despite these premiums, AMD's FY2 PEG ratio is currently at a 54.2% discount to the S&P 500's PEG ratio, indicating potentially undervalued growth prospects.
In terms of historical valuations, AMD is currently trading at a forward 12-month P/E of 31.8, which is slightly below its 5-year mean of 35.7. It is, however, within its 2-standard deviation range of 15.3 to 56.1, indicating a medium valuation relative to its 5-year range.
When compared to its peers, AMD's valuation appears relatively modest. When considering forward 12-month P/Es, NVDA and Marvell Technology, Inc. (MRVL) are trading at 47.0 and 31.6, respectively. This puts AMD at a relatively modest value in comparison.
Conclusion
AMD's recent strides in the AI and HPC markets are undeniably bold and ambitious. The company's AI product offerings and strategies, as showcased at the event, emphasize its ambition to become a major player in these markets. The MI300X and the OCP-based Instinct Platform represent AMD's latest efforts to innovate and compete in a sector dominated by Nvidia's CUDA.
However, the path ahead is not without challenges. Nvidia's strong CUDA ecosystem presents a significant hurdle for AMD. Overcoming this will require AMD to make the process of rewriting code for its AI chips as seamless as possible and to build a robust ecosystem around their products. Despite these challenges, the company's focus on performance, efficiency, and competitive pricing, coupled with its strategic partnerships, positions it well to capitalize on the long-term opportunities in the AIGC space.
Ultimately, while AMD's recent moves in the AI space are promising, its success in gaining significant market share will depend on how well it can navigate these challenges. Its ability to deliver on its promises, continue innovating, and win over AI researchers and businesses will be critical factors in determining whether AMD can truly become a formidable player in the AI and HPC markets. In the meantime, the AI industry will be watching AMD's progress closely.
For further details see:
AMD Strides Into The AI Battlefield: An Opportunity Or A Challenge?