Commoditizing the Petaflop: AMD ($AMD) in the AI Arena
For years, NVIDIA's graphics chips dominated artificial intelligence workloads, thanks largely to superior software support. AMD's GPUs, while powerful on paper, often lagged behind in real-world AI tasks. The reason wasn't just hardware; it was software. Nvidia's CUDA platform and optimized libraries made its GPUs the go-to choice for training neural networks, leaving AMD struggling with less mature tools. Developers lamented AMD's flaky driver support and limited AI ecosystem. In fact, ROCm (AMD's answer to CUDA) only recently became usable on gaming cards after being fairly mature on AMD's pricey data-center MI-series accelerators.
This long-standing CUDA moat gave Nvidia a virtually unchallenged lead in AI. But now, a series of bleeding-edge efforts from maverick hackers to startup projects suggest AMD's fortunes could be turning. Notably, renowned hacker George Hotz (known for iPhone and PlayStation hacks) has thrown his weight behind making AMD GPUs work for AI. His latest project, Tinygrad, and its custom driver code are showing that with the right software, AMD silicon can shine. It's a development that mirrors another recent breakthrough: DeepSeek's low-level optimizations on Nvidia hardware that achieved unprecedented performance. Taken together, these advances hint that AMD might finally become a serious contender in AI infrastructure.
Why AMD GPUs Lagged Behind in AI
The gap between Nvidia and AMD in AI hasn't primarily been about raw hardware. AMD's recent GPUs have competitive specs with plenty of compute cores and fast memory, theoretically enough to handle large neural networks. The issue was that Nvidia nurtured an entire ecosystem (CUDA, cuDNN, TensorRT, etc.) that made it easy to extract that performance for machine learning. By contrast, AMD's software stack was fragmented and immature. Key AI frameworks like TensorFlow and PyTorch only added stable AMD support much later, and even then, performance often trailed Nvidia. Drivers were another headache; many developers found AMD's GPU drivers and libraries prone to crashes or simply not optimized for AI workloads. Essentially, Nvidia's secret sauce was in software, highly optimized drivers and hand-tuned kernels that pushed its GPUs to the limit, whereas AMD's GPUs were underutilized due to suboptimal code.
George Hotz's Tinygrad: Rewriting the Software Stack
Enter George Hotz, also known as geohot, and his open-source project Tinygrad. Tinygrad is a lean neural network framework, a minimalist alternative to PyTorch. Hotz's goal isn't just academic; he wants to commoditize the petaflop, making AI compute cheap and accessible to everyone by enabling it on more affordable hardware. Recently, this has meant getting AMD GPUs to run AI models fast and reliably. Hotz identified that AMD's hardware wasn't the problem; it was the software holding it back. In true hacker fashion, he and the Tinygrad team began writing their own AMD GPU driver and runtime from scratch. This custom driver, called "AM," lives in user-space and bypasses large swaths of AMD's official driver code. Tinygrad can now directly communicate with AMD GPUs, optimizing specifically for machine learning tasks.
Instead of relying on AMD's GPU driver, which was prone to crash under heavy ML workloads, Tinygrad's stack memory-maps the GPU registers and manages them itself. This bold approach effectively reverse-engineered AMD's GPU interface, something possible because AMD publishes low-level chip information and has fewer secret sauce lock-ins than Nvidia. The result is a mostly working AI stack on AMD that Tinygrad completely controls, from the GPU ISA on up.
The DeepSeek Parallel: Custom Code, Big Gains
A similar story unfolded with DeepSeek, a small AI company from Asia, which optimized Nvidia GPUs by writing custom low-level drivers, dramatically improving the efficiency of large AI model training. DeepSeek achieved remarkable results by treating Nvidia's GPU like a blank canvas and coding directly at the hardware level. This vividly demonstrated that customized software solutions can vastly outperform manufacturer-provided GPU drivers.
The parallel to what George Hotz is doing is clear. Just as DeepSeek wrote custom low-level code to optimize Nvidia's GPUs, Hotz is writing a custom stack to leverage AMD's GPUs fully. Both efforts share the philosophy of rejecting default limitations and optimizing directly at the hardware level.
Implications: AMD as a Serious AI Contender
These developments have big implications for AMD and the AI industry. If AMD GPUs become viable for large AI models, it breaks Nvidia's near-monopoly, providing cloud providers, startups, and research labs with a second source for high-performance AI hardware. This competition could drive down AI computing costs and accelerate innovation.
For AMD, the stakes are high. Nvidia's valuation soared largely due to AI demand. If AMD captures part of this demand, proving their GPUs can handle AI equally well, it could significantly boost AMD's revenue and market valuation. George Hotz himself believes AMD's hardware has the potential to match or even exceed Nvidia's offerings when properly optimized.
In conclusion, AMD's collaboration with innovative software solutions like Tinygrad, combined with precedents set by DeepSeek, suggests AMD could become a critical player in AI technology, significantly reshaping its market position.
Tags: AMD, AMD 0.00%↑
Disclaimer:
“All views expressed are my own and are provided solely for informational and educational purposes. This is not investment, legal, tax, or accounting advice, nor a recommendation to buy or sell any security. While I aim for accuracy, I cannot guarantee completeness or timeliness of information. The strategies and securities discussed may not suit every investor; past performance does not predict future results, and all investments carry risk, including loss of principal.
I may hold, or have held, positions in any mentioned securities. I receive no compensation for this content and do not intend to influence market prices. Opinions herein are subject to change without notice. This material reflects my personal views and does not represent those of any employer or affiliated organization. Please conduct your own research and consult a licensed professional before making any investment decisions.”
Bullish outlook