- AMD is acquiring talent to close the instinct and the Blackwell GPU performance gap
- The Brium compiler experience could help AMD accelerate inference without specific hardware units
- The AI of Untether joins AMD, but existing customers run out of products support
The recent AMD movements in the AI sector have focused on strategic acquisitions aimed at strengthening their position in a market largely dominated by Nvidia.
These include the acquisitions of Brium, Silo AI, Nod.AI and the Untather AI engineering team, each aimed at reinforcing AMD AI software, inference optimization and chips design capabilities.
The objective is clear: it limits the performance and gap of the ecosystem between the AMD instinct GPU and the Blackwell line of Nvidia.
Acquisitions calculated in the middle of a competitive ecosystem
AMD described the acquisition of Brium as a key step to improve its AI software capabilities.
“Brium provides advanced software capabilities that strengthen our ability to offer highly optimized AI solutions throughout the stack,” said the company.
Brium strengths are found in compilers technology and end -to -end inference optimization, areas that could be crucial to achieve better performance ready to use and make the AMD software stack less dependent on specific hardware configurations.
While this makes it a solid technical case, it also suggests that AMD is still playing in the AI software ecosystem, instead of leading it.
Brium integration will affect several ongoing projects, including OpenAi Triton and Shark/Iree, which are considered fundamental to boost AMD inference and training capabilities.
The use of precision formats such as MX FP4 and FP6 points to a strategy of squeezing higher existing hardware performance. But the industry has already seen similar movements of NVIDIA, which continue to lead both in the unprocessed processing power and in the maturity of the software.
Another remarkable movement was the absorption of AMD of the entire Untather AI engineering team, a Canadian startup known for its energy efficiency inferences processors. AMD did not acquire the company, only talent, leaving the products of unterminated without support.
“AMD has signed a strategic agreement to acquire a talented team of Hardware Engineers and UNATHER AI AI software,” confirmed the company, highlighting an approach to the development of the compiler and the kernel along with the design of SOC.
This indicates a strong impulse in specific inference technologies, which are becoming increasingly critical as training revenues based on training face a possible decrease.
“The acquisition of AMD of the Untether engineering group is proof that GPU suppliers know that models training is over, and that a decrease in GPU revenues is just around the corner,” said Justin Kinsey, president of SBT Industries.
While that can exaggerate the situation, it reflects a growing feeling in the industry: energy efficiency and inference performance are the next borders, not simply building the fastest systems to train large models.
Despite AMD’s optimism and commitment to “an open and scalable artificial intelligence software platform” there are questions about its ability to match the close integration of NVIDIA between hardware and CUDA -based software.
Ultimately, while AMD is taking steps calculated to close the gap, Nvidia still has a considerable lead both in hardware efficiency and in the software ecosystem.
These acquisitions can approach AMD, but for now, Nvidia’s Blackwell remains the reference point of what is widely considered the best GPU for AI workloads.
