“Fortunately for AMD, the current shortage of Nvidia GPUs in the market has presented an opportunity and attracted numerous enthusiastic partners.
Many of my recent opinion pieces on MarketWatch have centered around artificial intelligence (AI) and the approaches of tech companies like AMD towards this field. At times, I receive inquiries questioning the longevity of the AI revolution.
Similar to the challenges faced by AMD from its investors: Is the growth in the AI sector genuinely sustainable and advantageous for shareholders?
During the commencement of the “Accelerate AI” event in San Jose, Calif. last Wednesday, CEO Lisa Su aimed to address these concerns. One of the initial slides in her presentation highlighted the expansion of the projected Total Addressable Market (TAM) for data center AI accelerators up to 2027, which has surged from \(150 billion to an estimated \)400 billion within the past year alone. This indicates a growth rate of over 70% for the next four years, supporting Su’s claim to the audience that “AI is undeniably the top priority at AMD.”
By observing the company’s strategic pivot, accelerated pace in AI software development, and collaborations with partners, one can find Su’s assertions compelling. AMD foresees that the MI300 product line will drive a revenue increase of at least $2 billion, a bold forecast for a company known for its conservative financial projections.
The AI-focused event shed light on why AMD is confident in its products and competitive stance. The company unveiled two new data center AI processors, the Instinct MI300X and the MI300A. The MI300X directly competes with traditional data center GPUs like the Nvidia H100, whereas the MI300A integrates CPU and GPU cores in a unified package, resembling Nvidia’s Grace Hopper and Intel’s delayed Falcon Shores project.
The spotlight was on the MI300X, currently available through various partners. AMD discussed the performance claims for the MI300X, stressing the necessity for external third-party validation, including comparable performance with the Nvidia H100 GPU in AI training workloads and a 40-60% enhancement in AI inference workloads. With 192GB of memory per GPU compared to the H100’s 80GB, this significant memory capacity likely contributes substantially to the performance of AMD’s new chip.
While Nvidia recently unveiled the H200 with up to 141GB of memory and improved overall performance, it is not yet accessible for testing, hindering direct comparisons with AMD’s offering.
Lucrative AI Ecosystem
Over the past year, AMD has placed considerable emphasis on the software ecosystem, particularly since Su introduced the MI300 chip at CES in January 2023, setting ambitious objectives for its AI strategy.
While Nvidia holds an edge with its CUDA development platform, AMD has made significant strides, supported by the industry’s shift towards more standardized models and frameworks in the AI ecosystem. The release of ROCm 6, a new iteration of its CUDA competitor, integrates numerous model optimizations and development library enhancements. OpenAI has pledged to support the MI300 in its standard releases, and representatives from AI companies such as databricks and Lamini have endorsed AMD’s software advancements, with one stating that they have surpassed CUDA.
Microsoft swiftly announced the availability of MI300X-based instances in its Azure cloud infrastructure, and Dell confirmed its readiness to accept orders.
These milestones are pivotal for AMD. Given Nvidia’s dominant market position, AMD depends on partners to advocate for its products. The present scarcity of Nvidia GPUs has created an opening for AMD, drawing in numerous partners eager to fulfill orders.
For Nvidia, this development was anticipated, with preparations underway since AMD’s initial announcement of the MI300 in 2022. The introduction of the H200 and recent performance enhancements from Nvidia were evidently strategic moves to mitigate the impact of AMD’s product launch on its perceived leadership. It is improbable that Nvidia’s chip sales will decline due to the MI300X; instead, AMD’s offering will likely address gaps in the production pipeline with key vendors like Microsoft, Dell, and Lenovo.
As the data center AI market expands, more opportunities will arise for AMD to leverage the demand for its MI300 chips in the foreseeable future.
Threats to Nvidia and Intel
Nevertheless, this poses a long-term risk for Nvidia. As more customers adopt the AMD MI300X and acknowledge its performance, comprehensive software support, and superior performance-to-cost ratio, it opens doors for AMD in future AI system integrations.
Once the market shortage diminishes, a broader array of cloud providers, system builders, and developers may consider AMD GPUs, whereas previously they might have hesitated to explore an alternative.
Intel faces more vulnerability than Nvidia from the MI300X family, positioning AMD as the de facto secondary source for AI computing in data centers, if it wasn’t already. Intel’s venture into the GPU space appears to have slowed down in recent quarters, with CEO Pat Gelsinger and the team primarily focusing on the Gaudi AI accelerators, which operate on a distinct architecture. Intel is hosting its AI event on December 14th, providing insights into its strategy for addressing the AI markets in both data centers and PCs.”
