Argomenti trattati
Picture this: the world of AI is on fire, and yet, here comes AMD, teetering on the edge of irrelevance, struggling to keep its footing while Nvidia dances around like it owns the place. Now, AMD’s hardware might be packing some serious heat with chips like Ryzen AI Max and Threadripper, but when it comes to software? Well, let’s just say they’ve been MIA. But hold your horses, because AMD is claiming it’s finally getting its act together, with ROCm 7 poised to shake things up.
AMD’s big reveal and the ROCm platform
During their recent Advancing AI event, AMD put the spotlight on their enterprise-class GPUs, specifically the Instinct lineup. But let’s not overlook ROCm—an under-the-radar software platform that AMD seems to think is a game changer. With ROCm 7 rolling out, AMD is boasting a threefold increase in AI inferencing capabilities. And guess what? It’s finally making its way to Windows, stepping up to Nvidia’s CUDA, which has been holding the crown way too long.
Radeon Open Compute (ROCm) isn’t just a fancy name; it’s AMD’s entire open software stack designed for AI computing, complete with drivers and tools to actually run AI workloads. Remember the Nvidia 5060 fiasco? Yeah, without software, it was just a glorified paperweight. AMD has had its fair share of these blunders too, but instead of throwing cash like Nvidia, they opted to cater to big businesses. Ramine Roane, AMD’s corporate VP of AI solutions, admitted it was a sore point for them. “We focused ROCm on cloud GPUs, but it wasn’t always working on the endpoint — so we’re fixing that.” Well, isn’t that just adorable?
Lessons from the past
In the cut-throat world of tech, merely having a great product isn’t enough anymore. You need to charm the pants off your customers and partners to win their hearts. Just look at the history—when Sony embedded a Blu-ray drive into the PlayStation, movie studios jumped on board, giving Blu-ray the edge over HD-DVD. AMD’s Roane finally acknowledged that Windows is where AI developers hang out. “It was a decision to basically not use resources to port the software to Windows, but now we realize that, hey, developers actually really care about that,” he said. Well, no kidding! Took you long enough, didn’t it?
Support and projections
ROCm is set to gain traction with PyTorch in preview by Q3 2025 and ONNX-EP in July. This means AMD processors—be it a laptop with a Ryzen AI processor or a desktop with Radeon GPU—will finally see some love in AI applications. PyTorch, the darling of machine learning, is where many popular AI models are built. It’s about time AMD users get to actually harness the power of their hardware.
And let’s not forget about Linux users! ROCm will find its cozy home in various Linux distributions, including Red Hat and Ubuntu, in late 2025. Roane even laid out what these AMD platforms can handle, from a humble Ryzen AI 300 notebook to the monstrous Threadripper platform. The numbers are impressive: ROCm 7 boasts a 3.2X performance enhancement in Llama 3.1 70B, 3.4X in Qwen2-72B, and a whopping 3.8X in DeepSeek R1. For the uninitiated, the “B” stands for billions of parameters—more parameters typically equate to better quality outputs. In the AI race, being fast isn’t just a luxury; it’s a necessity.
Training versus inferencing
Now, let’s break it down: training generates the AI models you see in products like ChatGPT, while inferencing is when these models spring to action. It’s like teaching a dog to fetch; the real fun happens when you throw the ball. AMD claims that ROCm 7 not only enhances inferencing but also keeps training performance on par with previous generations—about three times better, to be precise. And let’s be real, it’s not a one-horse race. AMD’s MI355X, running the new ROCm software, is expected to outperform Nvidia’s B200 by 1.3X on DeepSeek R1 with 8-bit floating-point accuracy. Talk about a showdown!
Where’s the consumer application?
Despite the impressive performance improvements, one glaring hole remains in AMD’s strategy: they lack a consumer-focused application to lure users into the AI world. Intel has its AI Playground, Nvidia has its own third-party app, but AMD? They’re still figuring it out. Roane called AI Playground a “good idea,” but when asked about plans to create something similar, he seemed to shrug it off. “No specific plans right now, but it’s definitely a direction we would like to move.” Well, that’s reassuring, isn’t it? Maybe someday, AMD will get around to catering to the consumer crowd.
So here we sit, waiting to see if AMD can truly make a dent in the AI landscape. The stakes are high, and the clock is ticking. But for now, all we can do is watch and wait, with a healthy dose of skepticism. Can ROCm 7 really turn the tide for AMD? Time will tell, but don’t hold your breath.