AMD Unveils MI325X AI Chip, Plans MI350 Series to Compete with Nvidia's AI Dominance

Image Source: AMD

Advanced Micro Devices (AMD) has announced plans to commence mass production of its latest artificial intelligence processor, the MI325X, in the fourth quarter of 2024. Unveiled at a recent event in San Francisco, CEO Lisa Su detailed the company's strategic move to enhance its foothold in the AI chip market, which has been predominantly led by Nvidia. The MI325X is poised to deliver enhanced performance through its innovative architecture and increased memory capacity, marking a significant upgrade from its predecessors, the MI300X and MI250X.

Next-Generation MI350 Series on the Horizon

Looking ahead, AMD revealed its roadmap for the next-generation MI350 series chips, slated for release in the latter half of 2025. These upcoming chips are expected to feature substantial memory enhancements and a revamped architecture designed to offer substantial performance boosts. This progression underscores AMD's commitment to staying at the forefront of AI technology advancements.

MI325X Technical Specifications

The MI325X AI chip is built on AMD’s cutting-edge CDNA3 GPU architecture, utilizing TSMC’s advanced 5nm and 6nm FinFET lithography processes. Key specifications include:

• Stream Processors: 19,456

• Matrix Cores: 1,216

• Compute Units: 304

• Peak Engine Clock: 2,100 MHz

• Peak Performance:

  • FP8 (E5M2, E4M3): 2.61 PFLOPs

  • FP8 with Structured Sparsity: 5.22 PFLOPs

  • FP16: 1.3 PFLOPs

  • FP16 with Structured Sparsity: 2.61 PFLOPs

  • TF32 Matrix (FP32): 653.7 TFLOPs

  • TF32 with Structured Sparsity: 1.3 PFLOPs

  • FP32 Matrix: 163.4 TFLOPs

  • FP64 Matrix: 163.4 TFLOPs

  • FP32: 163.4 TFLOPs

  • FP64: 81.7 TFLOPs

  • INT8: 2.6 POPs

  • INT8 with Structured Sparsity: 5.22 POPs

  • bfloat16: 1.3 PFLOPs

  • bfloat16 with Structured Sparsity: 2.61 PFLOPs

• Transistor Count: 153 Billion

These specifications highlight the MI325X’s robust processing capabilities, making it a formidable competitor in the AI chip market.

Market Response: Mixed Reactions from Investors

Despite the anticipation surrounding AMD’s announcements, the market reaction was lukewarm. Shares of AMD dipped by nearly 5% following the news, a decline attributed by analysts to the lack of announcements regarding major new cloud-computing partnerships. This sentiment contrasts with Nvidia’s shares, which saw a modest increase, and Intel’s, which experienced a slight drop. The absence of high-profile customers may have tempered investor enthusiasm despite the promising technological developments.

Competitive Dynamics: Nvidia and Intel in Focus

AMD’s strategic push into AI chips intensifies the competition with industry giants Nvidia and Intel. While Nvidia continues to lead with its Blackwell architecture, AMD’s MI325X aims to offer a formidable alternative. Intel, on the other hand, faces challenges in formulating a cohesive AI chip strategy, projecting AI chip sales of over $500 million for 2024. The heightened competition is set to drive innovation and potentially reshape the AI chip landscape.

[See our previous report: AI Chip Wars: Nvidia’s Rivals Gear Up for a Slice of the Market]

Manufacturing Strategy: Reliance on TSMC

AMD reaffirmed its commitment to utilizing Taiwan Semiconductor Manufacturing Company (TSMC) for advanced chip manufacturing processes essential for producing high-speed AI processors. CEO Lisa Su emphasized AMD’s aggressive use of TSMC’s Arizona facility, highlighting the company's preference to maintain production capabilities within Taiwan while expressing interest in expanding manufacturing capacity beyond the region in the future.

[See our previous report: TSMC's 2nm Breakthrough Powers the Next Wave of AI and Mobile Tech]

Expanded Chip Portfolio for Data Centers and PCs

In addition to AI processors, AMD unveiled a suite of new networking chips designed to enhance data movement within data centers. The company also introduced an updated server CPU line, codenamed Turin, featuring versions optimized to support GPU data feeds for accelerated AI processing. Furthermore, AMD announced three new Zen 5-based PC chips tailored for laptops, optimized for AI applications and compatible with Microsoft’s Copilot+ software. These developments reflect AMD’s comprehensive approach to catering to both enterprise and consumer markets.

[See our previous report: AMD's Ryzen AI 9 HX 370: The New AI Powerhouse in Asus Zenbook S 16]

Financial Projections and Market Valuation

AMD has revised its AI chip revenue forecast upward to $4.5 billion for the year, up from the previous estimate of $4 billion, driven by robust demand for its MI300X chips amid the surge in generative AI applications. Analysts project AMD’s data center revenue to reach $12.83 billion, though this pales in comparison to Nvidia’s expected $110.36 billion. Despite the strong performance, both AMD and Nvidia maintain high valuation multiples, trading at over 33 times their 12-month forward earnings estimates, reflecting high investor expectations in the rapidly evolving AI sector.

[See our previous report: AI Takes Center Stage at Computex 2024: A Tech Revolution in the Making]

Source: Reuters