Large Language Models: AMD And Nvidia Analysis (NASDAQ:AMD)

The recent frenzy over OpenAI’s launch of ChatGPT, powered by GPT-3, has drawn mass market attention to the power and sprawling use cases of large language models (“LLM”) available today. While our recent series of coverage pertaining to LLMs has primarily focused on how OpenAI’s GPT-3 LLM could be both a threat and opportunity for competing developers counting Microsoft (MSFT), Twilio (TWLO) and Google (GOOG/GOOGL), the following analysis will dive further upstream to gauge the role of chipmakers in the said advancement in AI.

Specifically, AMD (NASDAQ:AMD) and Nvidia (NVDA) have rapidly gained reputation in recent years for their involvement in enabling key data center solutions critical to the development of AI advancements like LLMs today, from both the software and hardware aspect. As discussed in a recent coverage, Nvidia already demonstrates a direct benefit in its hardware business from the development of LLMs today, with its GPUs already found in the supercomputer developed for OpenAI’s GPT-3. Nvidia’s recent introduction of “NVIDIA NeMo LLM Service” also underscores its continued commitment to a full-stack hardware-software business model to further its leadership in AI developments, while enabling TAM expansion in related markets. Meanwhile, AMD continues to take an indirect approach, with its latest “AMD Instinct MI210” data center GPU and “ROCm 5” open software platform being key solutions offered to capture growing opportunities in the machine learning community. The following analysis will discuss the role of each data center processor market leader’s approach to opportunities stemming specifically from LLM development, and explore their implications on both companies’ longer-term growth prospects.

AMD – What is ROCm and AMD Instinct?

While AMD is most notably known for its rapid rise to market leadership in server processors – namely, its “EPYC” series – in recent years, it has also recently expanded its expertise into data center GPUs, Nvidia’s turf. The AMD Instinct MI200 series data center GPUs are currently the “world’s fastest HPC (high performance computing) and AI accelerator”, referring to chips designed specifically to increase performance, improve efficiency, and reduce latency in AI and ML applications. Powered by AMD’s “CDNA 2” architecture tailored for HPC and AI workloads, the MI200 accelerators complement the ROCm 5 open software platform to exhibit “exascale-class” performance, and are designed to address “growing demand for compute-accelerated data center workloads and reducing the time to insights and discovery” for all users.

With up to 2.3x better HPC performance leadership over rival GPUs like Nvidia’s “Ampere A100“, the MI200 series accelerators make a great option for supercomputers and servers used to train LLMs. More notably, the AMD Instinct accelerators are known for their integration alongside the EPYC server processors in more than 100 supercomputers in the most recent Top500 list, up from 73 in the prior year. The achievement underscores its performance competency in supporting complex workloads spanning “climate, biology, and medicine, new energies and materials”, inclusive of application-specific LLMs, to accelerate discovery.

Circling back to the ROCm open software platform, the solution aims at providing tools for developers to optimize AI/ML developments. ROCm is designed for compatibility with different vendors’ hardware – from AMD’s very own Instinct data center and Radeon workstation GPUs to Nvidia’s equivalents – as well as support for a variety of the “most popular ML frameworks”, including TensorFlow and PyTorch.

ROCm also allows portability of “computation to one or more CPUs or GPUs” across different devices without having to rewrite the code, facilitating maximum performance and optimized efficiency. For instance, Nvidia’s “CUDA” software development toolkit specific for “GPU-accelerated applications” is limited for use with Nvidia hardware and would require additional “porting solutions” to run CUDA code on non-Nvidia hardware. The porting solutions include HIP (Heterogeneous-Computing Interface for Portability) which is supported by ROCm. Essentially ROCm can automatically convert CUDA software to HIP, making its application universal on “different server processors and GPUs”. ROCm code can also be universally applied on non-AMD server processors and GPUs as well, making it an accessible and convenient tool in enabling development of next-generation AI workloads, including LLMs.

Nvidia – What is NeMo LLM Service?

In addition to an indirect approach like AMD on facilitating development of LLMs critical for next-generation applications via various supporting hardware/software solutions, Nvidia has also recently introduced a cloud service specific for LLM development and application. As discussed in our previous coverage, LLMs power more than just chatbots, and can also be used to facilitate complex use cases like “code development, as well as protein structure and biomolecular property predictions”, and simpler use cases like “lightning-fast semantic search”. The NeMo LLM Service and BioNeMo LLM Service are almost like a cloud-based hub of LLMs that developers can access and customize with ease and efficiency, reducing development costs and time to deployment by reducing the need to build and train models from scratch:

• NeMo Large Language Model Service – The NeMo LLM cloud provides developers with access to a “number of pre-trained foundation models” that can be further customize-trained using “prompt-learning“. Specifically, foundation models in NeMo LLM Service range from “3 billion parameters up to Megatron 530B, one of the world’s largest LLMs” trained to date, trailing only slightly behind Google’s PaLM in terms of parameters and performance. Developers can build customized solutions on top of the foundation models by only adding a few hundred additional data points for prompt training, which can be completed within “minutes to hours compared with weeks or months required to train a model from scratch”. NeMo LLM Service also offers a “no-code option”, which allows accessibility for mainstream users and enables effective and efficient development of industry-specific LLMs in the future. With more than 40% of the workforce noting that the use of low-code techniques would be critical in creating value in the data-driven era, NeMo LLM Service is well-positioned to capitalize on the growing role of LLMs in “transforming every industry”, big and small.
• BioNeMo Large Language Model Service – The BioNeMo LLM Service is an API (application programming interface) that extends LLM applications “beyond language” and into scientific use cases specific to pharmaceuticals and biotech. Specifically, the BioNeMo LLM Service houses two LLMs specific for chemistry and biology use cases, supporting “protein, DNA and biochemical data to help researchers discover patterns and insights to biological sequences”. The BioNeMo LLMs are illustrative of how LLMs can be used in complex non-language-specific workloads (e.g., chatbots, text summary, content generation, etc.) as discussed in the earlier section, including the storage, processing, and training on “information about the structure of proteins, evolutionary relationships between genes, and even generate novel biomolecules for therapeutic applications”.

Implications for AMD and Nvidia

AI workloads are accelerating and becoming increasingly complex. LLMs are only one of many “state-of-the-art” (“SOTA”) algorithms, underscoring the robust demand environment for related data center hardware/software solutions like those offered by both Nvidia and AMD in the coming years.

LLM applications are actually ubiquitous today, including “BERT” (Bidirectional Encoder Representations from Transformers) which is used in powering Google Search, a simple and common tool accessed on the daily. Although more complex LLMs are currently more commonly deployed in industry-specific applications like pharmaceutical and biotech as discussed earlier, they are bound to become more available for mass market applications in the coming years, given ongoing digital transformation trends and demands for convenient and efficient access to and usage of growing troves of data ensuing from the advent of connectivity. These mass market use cases span across chatbots to even gaming – for instance, OpenAI’s GPT-3 LLM is already being used by Fable Studio, an AI-focused story-telling and interactive solutions firm, in “creating a new genre of in-game interactive stories“.

Demand for natural language processing (“NLP”) use cases – such commonly used smart assistants like Alexa (AMZN) and Siri (AAPL) – is forecast to expand at a 23% CAGR in the next 10 years. With LLMs being critical to enabling different NLP applications, this corner of AI advancement just underscores one of a myriad of robust demand environments for AMD and Nvidia’s data center and AI solutions – both LLM- and non-LLM-specific hardware and software offerings.

For both Nvidia and AMD, data center has become a core part of their respective businesses, with the former leading in AI/GPU processors, and the latter in server CPU processors. While both companies have experienced meaningful declines across more inflation- and recession-prone corners within their respective businesses – particularly in gaming and workstation hardware – their growing prowess in the provision of data center solutions has been a fortress, backing resilience ahead of the looming cyclical downturn within the industry and across broader markets. Specifically, the critical role that data center solutions provided by AMD and Nvidia play in propelling key next-generation digital growth trends is likely to turn the chipmakers’ businesses from “cyclical to secular“, underpinning sustained longer-term upsides.

Final Thoughts

LLM is expected to become one of the key next-generation innovations to be integrated deeper into our daily settings, making strong longer-term tailwinds for both AMD and Nvidia. From a valuations perspective, we believe the current macroeconomic backdrop, which has been unforgiving on the semiconductor sector, has created a compelling entry opportunity for AMD. Meanwhile for Nvidia, although its longer-term growth trajectory remains intact given its continued market leadership and critical role in enabling key next-generation technologies, we believe its still-lofty valuation premium could be a cause for further vulnerability to broader market volatility ahead of the looming downturn, which would create better entry opportunities heading into the new year instead.