AMD's Instinct MI355X accelerator is a pivotal next-generation product in the company's AI portfolio. Recently, tech giant Oracle announced it as a significant customer. According to Oracle's latest financial report, the company has acquired 30,000 MI355X accelerators to construct a large-scale AI computing cluster. Oracle's co-founder, Larry Ellison, revealed, "We entered a multibillion-dollar partnership with AMD in the third quarter, aiming to implement 30,000 of the latest MI355X GPUs to bolster powerful AI computing capabilities."

The MI355X represents AMD's data center GPU based on the new CDNA4 architecture, utilizing TSMC's advanced 3nm process. Designed specifically for AI tasks, this accelerator offers outstanding performance. It delivers 2.3 petaflops of FP16 computational power, while its FP8 performance rises to 4.6 petaflops, marking an approximate 77% improvement over the previous MI300X series. Furthermore, the MI355X introduces support for FP4 and FP6 low-precision numerical formats, elevating its FP4 computation ability to 9.2 petaflops. This substantially enhances computational efficiency while reducing memory usage, making it perfect for contemporary large-scale AI model training and inference.

Regarding memory, the MI355X is outfitted with 288 GB of HBM3E memory across eight stacks, offering a total bandwidth of 8 TB/s. This marks a significant increase from 256 GB and 6 TB/s in the previous MI325X, directly competing with Nvidia's Blackwell series regarding memory capacity and bandwidth. The incorporation of HBM3E technology ensures that data-intensive AI applications can operate effectively, providing robust support for ultra-large-scale computing.

Oracle intends to leverage the cluster to advance its cloud services and AI offerings. Ellison mentioned that four major cloud security companies — CrowdStrike, Cybereason, Newfold Digital, and Palo Alto Networks — have chosen to transition to Oracle Cloud. Moreover, Oracle introduced a new product named the AI Data Platform, aiding its vast database clientele by analyzing data using the latest AI models. The platform integrates state-of-the-art AI technologies such as OpenAI, xAI, and Meta while utilizing the vector capabilities of Oracle's Release 23 database to swiftly convert existing customer data into AI-compatible formats. This functionality allows organizations to unlock their data's value rapidly, ensuring data privacy and security.

The production and market deployment plans for the MI355X are explicit with AMD slated to commence shipments in the latter half of 2025. This timeline aligns with AMD's annual data center GPU refresh cycle, allowing ample preparation time for Oracle's data center expansion. Reportedly, Oracle is intensifying its cloud infrastructure, planning to double data center capacity by the end of 2025 to meet climbing AI and cloud computing demand. Thus, the 30,000 MI355X GPUs cluster will become the central powerhouse of Oracle Cloud, enhancing its AI workload competitiveness.

AMD's momentum in the AI accelerator arena is gathering pace. Alongside the MI355X, AMD plans to introduce the MI400 series, based on the CDNA "Next" architecture in 2026. The MI355X's chipset design continues AMD's multi-chip packaging strategy, enabling multiple compute chips to collaboratively achieve high performance. This approach not only boosts manufacturing efficiency but also reduces costs and offers adaptability against Nvidia.

It's noteworthy that the MI355X's performance metrics closely rival Nvidia's Blackwell B200, which also provides 9 petaflops of FP4 computing power. However, the MI355X's 288 GB of HBM3E memory surpasses the B200's 192 GB by 50%, affording greater space for memory-intensive applications. Additionally, the MI355X's 8 TB/s bandwidth matches the B200, ensuring proficient data transfers. This hardware specification comparison demonstrates AMD's strategy to capture a larger share of the AI accelerator market by challenging Nvidia with more memory and comparable bandwidth.

Oracle's acquisition decision underscores industry trends: the enterprise demand for AI computing power is escalating, and cloud service providers are capitalizing on the market through hardware advancements. The MI355X deployment will enable Oracle to accommodate larger-scale AI models, like language models and multimodal systems encompassing hundreds of billions or trillions of parameters. This capability is crucial for customers in sectors such as finance, healthcare, and security, who require real-time data processing and insight generation.

With the MI355X's general availability slated for the second half of 2025, the AMD-Oracle partnership is poised to invigorate AI technology development. This accelerator not only signifies a major leap in hardware performance but also heralds the future trend of deep integration between cloud computing and AI. For tech enthusiasts, this hardware-driven AI evolution is one to watch closely.