https://seekingalpha.com/article/...losion?source=feed_tag_long_ideas
Besonders interessant finde ich folgende Abschnitte des Artikels:
"AMD Well-Positioned to Capture Inference Wave
AMD is entering this new AI-centric compute wave from a position of strength and strategic focus. While Nvidia (NVDA) has traditionally been the undisputed leader on both the training and inference hardware front, AMD has been steadily narrowing the gap on the inference front. AMD's latest Instinct MI300 series is purpose-built for the most demanding AI inference workloads, which is exactly what vibe coding calls for. In fact, the Instinct MI300x has 192 GB of HBM3 memory to hold the largest models on chip, which is double the memory capacity of Nvidia's flagship H100.
This memory advantage is key for inference when serving giant GPT-style models. Early indications show that AMD's design is highly competitive. While it may not deliver up to the 1.6x performance advantage over Nvidia's H100's as is claimed by AMD, the fact that it is even competitive with the H100 speaks volumes. In other words, AMD is not just a second source, but in some cases a superior solution for inference. Equally important is the fact that AMD has been building up its open-source ROCm ecosystem to fill its largest traditional weakness on the software front, which has been dominated by Nvidia's CUDA.
ROCm, Enterprise Partnerships, and MI300X Traction
AMD has wisely played to its strengths by championing openness and partnerships, rather than competing as a closed ecosystem, in which it would have little to no chance against Nvidia's CUDA. AMD's ROCm software stack is open-source, which means that researcher, cloud providers, and enterprises can collaborate and customize it to their own specific needs. AMD has heavily leaned into this open avenue by working closely with major machine learning frameworks to optimize performance of ROCm. In fact, frameworks like PyTorch now have strong ROCm backends, and with each iteration, it reaches closer and closer to parity with Nvidia's CUDA. The company late 2024 ROCm 6.3 specifically brings large boosts in inference throughout, which comes at a perfect time of the vibe coding boom.
If vibe coding and generative AI are to permeate every corner of software development, Nvidia simply will not have the capability to supply all the inference needs. Underlying models will be running everywhere, and AMD's commitment to an open AI ecosystem positions it well to soak up excess inference demand and perhaps even take market share from Nvidia. AMD is not just selling chips, but rather a platform that others can build freely on. This resonates more with the collaborative ethos of the AI developer community, as many in the community are already looking to solutions outside of CUDA.
AI Momentum Starting to be Reflected in Financials
AMD has not exactly had the greatest quarterly performances over the past year. Up until now, frontier AI models were not yet capable of sparking the kind of explosive inference demand we are starting to see. While these models were still a huge paradigm shift from anything the industry has seen prior, they were still just a step behind the capabilities truly needed to dramatically improve developer capabilities. The releases of truly capable reasoning models like Anthropic's Claude 3.7 thinking and OpenAI's o1 completely changes this. However, this vibe coding phenomenon has just started, so it is only natural that the inference explosion has not yet been reflected in AMD's financials.
Regardless of AMD's relatively weak performance over the past few quarters, the company is starting to see signs that its AI focus is starting to pay off. In its recent Q4 report, the company reported a record Data Center segment of $3.9 billion. This represents a 69% YoY increase, and reflects the increasing demand for its AMD Instinct GPU and EPYC CPUs. In fact, the company delivered more than $5 billion in AMD Instinct accelerators in 2024 alone. The success of AMD's Data Center segment also helped the company push gross margins to 54% in Q4. The rise of vibe coding will likely only push its revenue and gross margin figures even higher.
|
Thread abonnieren
