AMD Leverages Samsung HBM3E for Instinct MI350 AI GPUs

AMD unveils the Instinct MI350 GPUs using Samsung's HBM3E memory to enhance AI computing power.

The race to dominate the artificial intelligence hardware space is heating up, and AMD’s latest move is a clear signal they intend to lead. On June 12, 2025, AMD officially unveiled its Instinct MI350 and MI355X GPUs, flagship AI accelerators built for the most demanding data center workloads—and packed with some of the industry’s most advanced memory technology. What caught the attention of analysts and tech enthusiasts alike was AMD’s confirmation that these new GPUs are using Samsung’s latest HBM3E 12-Hi memory, a crucial detail that gives these chips a formidable edge in the AI arms race[1][4].

Let’s be honest: memory bandwidth and capacity are often the unsung heroes of AI acceleration. While most headlines focus on raw compute power, anyone who’s spent time inside a data center knows that feeding those hungry AI models with data is just as critical as crunching the numbers. That’s exactly why AMD’s choice of Samsung’s HBM3E is so significant—it’s not just about speed, but about delivering the right kind of memory for tomorrow’s AI workloads.


The AI Hardware Arms Race: Why Memory Matters

AI training and inference are memory-intensive tasks. Large language models, computer vision pipelines, and generative AI systems rely on vast amounts of data, and the speed at which this data can be accessed directly impacts performance. Traditional GDDR memory simply can’t keep up with the bandwidth and capacity demands of modern AI models.

HBM (High Bandwidth Memory) was developed specifically to address this bottleneck. HBM3E, the latest iteration, stacks memory chips vertically (hence the “12-Hi” designation), significantly increasing both capacity and bandwidth while keeping power consumption manageable. Samsung’s 12-Hi HBM3E is among the densest and fastest memory solutions available today, making it a natural fit for AMD’s Instinct MI350 series[1][4].


Inside AMD’s Instinct MI350 and MI355X: Powerhouse Specs

AMD’s new Instinct MI350X and MI355X GPUs are built on the CDNA 4 architecture, a major leap forward in GPU design for AI and high-performance computing. The MI350X, for example, features 288 GB of HBM3E memory—delivered via Samsung’s cutting-edge 12-Hi stacks—and a staggering 8 TB/s memory bandwidth. The MI355X pushes the envelope even further, with similar memory specs but additional compute resources for even greater performance[1][2][4].

Here’s a quick breakdown of the key specs:

Feature Instinct MI325X Instinct MI350X Instinct MI355X
Architecture CDNA 3 CDNA 4 CDNA 4
Memory 256 GB HBM3E 288 GB HBM3E 288 GB HBM3E
Memory Bandwidth 6 TB/s 8 TB/s 8 TB/s
FP64 Performance 72 TFLOPs 78.6 TFLOPs
FP16 Performance 2.61 PFLOPS 4.6 PFLOPS 5 PFLOPS

These numbers aren’t just for show. They translate directly into real-world performance gains for AI training and inference, with AMD claiming up to a 4x generational leap in some benchmarks and a jaw-dropping 35x faster inference performance in certain workloads compared to previous generations[1].


Chiplet Design: The Secret Sauce

AMD’s Instinct MI350 series isn’t just about brute force. The company has doubled down on its chiplet-based design philosophy, which allows for greater flexibility, scalability, and efficiency. Each MI350X GPU is made up of two I/O dies (IODs), built on TSMC’s 6 nm process, and each IOD can host up to four Accelerator Compute Die (XCD) tiles, built on TSMC’s 3 nm node. The result is a total of 288 compute units per package, with each IOD controlling four HBM3E stacks for a combined 288 GB of memory[4].

This modular approach not only improves performance but also makes manufacturing more cost-effective and scalable. It’s a strategy that has served AMD well in the CPU space, and now it’s paying dividends in the AI accelerator market.


Samsung’s HBM3E: The Memory Behind the Magic

Samsung’s HBM3E 12-Hi memory is a game-changer. By stacking memory chips vertically, Samsung has managed to cram more memory into a smaller footprint, while also boosting bandwidth and reducing power consumption. This is critical for data center operators who are constantly balancing performance, power, and space constraints.

AMD’s decision to use Samsung’s latest memory technology is a clear signal that they’re serious about competing head-to-head with NVIDIA’s Blackwell B200 series, which also relies on advanced memory solutions. In fact, AMD’s top-spec MI355X is being positioned as a direct competitor to NVIDIA’s B200, with both companies vying for dominance in the AI hardware market[1][4].


Real-World Impact: Where the Rubber Meets the Road

So, what does all this mean for businesses and researchers? For starters, faster memory and greater bandwidth mean that AI models can be trained and deployed more quickly, reducing the time-to-insight for everything from drug discovery to financial modeling. Data centers can handle larger models and more concurrent workloads, which is essential as AI adoption continues to accelerate across industries.

Consider the implications for generative AI, where large language models like GPT-4 and beyond require massive amounts of memory and bandwidth to function efficiently. With AMD’s Instinct MI350 series, organizations can push the boundaries of what’s possible, whether they’re developing the next generation of AI-powered applications or running complex simulations for scientific research[1][2].


The Competitive Landscape: AMD vs. NVIDIA

It’s impossible to talk about AI hardware without mentioning NVIDIA, the current market leader. NVIDIA’s Blackwell B200 GPUs are also built for massive AI workloads, and like AMD, the company is investing heavily in advanced memory technologies. The competition between these two giants is driving rapid innovation, with each company pushing the other to deliver faster, more efficient, and more scalable solutions.

AMD’s Instinct MI350 series, with its chiplet design and Samsung HBM3E memory, represents a serious challenge to NVIDIA’s dominance. The battle for AI hardware supremacy is far from over, but one thing is clear: the pace of innovation is accelerating, and the stakes have never been higher[1][4].


Looking Ahead: The Future of AI Hardware

As someone who’s followed AI for years, I’m excited to see how these advances will shape the future of artificial intelligence. Memory technology is just one piece of the puzzle, but it’s a critical one. With AMD and NVIDIA both pushing the boundaries of what’s possible, we can expect to see even more powerful and efficient AI accelerators in the coming years.

The adoption of advanced memory solutions like Samsung’s HBM3E is just the beginning. As AI models grow larger and more complex, the demand for faster, denser, and more efficient memory will only increase. This is a trend that will benefit not just tech companies, but anyone who relies on AI to drive innovation and solve real-world problems.


Excerpt for Article Preview:
AMD’s new Instinct MI350 AI GPUs, powered by Samsung’s latest HBM3E 12-Hi memory, deliver unprecedented AI performance and memory bandwidth, setting a new standard for data center acceleration[1][4].


Conclusion

AMD’s confirmation that its Instinct MI350 AI GPUs leverage Samsung’s HBM3E 12-Hi memory is more than a technical detail—it’s a strategic move that positions the company at the forefront of the AI hardware revolution. With industry-leading memory bandwidth, advanced chiplet design, and a clear focus on real-world AI workloads, AMD is setting the stage for the next generation of data center innovation. As the competition with NVIDIA intensifies, the real winners are the organizations and researchers who will benefit from faster, more efficient, and more scalable AI solutions. The future of AI hardware is here, and it’s built on memory that’s as fast as the ideas it enables.


TAGS:
amd-instinct, hbm3e, ai-accelerator, samsung-memory, nvidia-blackwell, data-center, generative-ai, machine-learning

CATEGORY:
artificial-intelligence

Share this article: