Track Hyper | AMD's continuous acquisitions strengthen AI layout

Author: Zhou Yuan / Wall Street News

On June 5th and the day before, AMD announced the completion of two strategic acquisitions: the acquisition of the open-source software company Brium under undisclosed terms, and the acquisition of the core engineering team from AI inference chip developer Untether AI.

These two transactions are AMD's latest moves following its acquisitions of Mipsology, Nod.ai in 2023, and Silo AI in 2024, marking a new phase in AMD's strategy to build a full-stack AI solution.

AMD's two transactions aim to build a competitive edge in the "chip-software-system" integration.

Untether AI's chip design capabilities enhance AMD's hardware strength in edge computing and data centers, while Brium's software optimization capabilities fill AMD's ecological gap in the inference stage.

This "hardware-software synergy" strategy aligns with AMD's previous acquisitions of Mipsology (compiler), Nod.ai (framework adaptation), and others, all aimed at forming a complete AI technology stack.

Brium's core technology focuses on optimizing the operational efficiency of AI inference software across different hardware architectures.

The company's more significant technological route is primarily centered on the cloud AI software ecosystem, with tools integrated into the official PyTorch plugin library, allowing developers to achieve AMD hardware acceleration through simple configurations. This open-source strategy helps attract more developers to use the AMD platform, gradually breaking NVIDIA's CUDA monopoly.

In other words, the tools developed by Brium can adapt AI models originally designed for NVIDIA GPUs to platforms like AMD Instinct GPUs, addressing the industry pain point of "workloads often being extensively adjusted based on NVIDIA GPUs."

This technological capability holds strategic value for AMD: currently, over 80% of AI software globally is developed based on the NVIDIA CUDA ecosystem, while AMD Instinct MI300X and other hardware, although close to NVIDIA H100 in computational parameters, release less than 50% of their performance in actual deployment.

The acquisition of Brium is a crucial part of AMD's effort to build an open AI software ecosystem.

Since 2023, AMD has formed a complete software stack covering compilers, framework adaptation, and model training through the acquisitions of Mipsology (Zebra compiler), Nod.ai (Shark abstraction layer), and Silo AI (large model development team).

With Brium joining AMD's ranks, it will further strengthen AMD's optimization capabilities in the inference stage when designing AI accelerator cards, helping developers achieve "one-time development, multi-platform deployment," thereby reducing dependence on NVIDIA hardware.

It is worth noting that Brium's technological route can complement AMD's recently launched ROCm 6.0 framework. The issue is that ROCm 6.0 introduces new automatic tuning tools but still has limitations in cross-hardware compatibility.

Brium's solutions can reduce inference latency by over 30% while maintaining model accuracy, especially in scenarios with high real-time requirements such as medical image analysis and financial risk control The core value of Untether AI lies in its low-power, high-performance AI inference chip technology.

The chips developed by the company for automakers such as General Motors and Mercedes-Benz achieve technical performance that is twice as fast and 40% more energy-efficient compared to competitors in autonomous driving perception systems.

Through dynamic resource allocation, the unique Spatial Architecture can efficiently run complex visual Transformer models on edge devices while supporting high-density deployment in enterprise data centers.

The engineering team acquired by AMD will directly enhance AMD's two technological capabilities: first, AI compiler development; Untether AI's experience in dynamic compilation and kernel optimization can improve the software adaptation efficiency of AMD hardware; second, SoC design; the acquired team previously led the development of automotive-grade chips with integrated AI accelerators, which will help AMD expand into the autonomous driving and industrial IoT markets.

The continuous acquisitions reflect AMD's transformation from a "hardware supplier" to an "AI solutions provider."

By integrating Brium's software optimization capabilities and Untether's engineering team's chip design experience, AMD is building a "chip-software-system" triad of competitiveness.

This strategy was already evident during its 2024 acquisition of ZT Systems (a data center infrastructure company): by incorporating hardware, networking, and system integration capabilities into its own framework, AMD can directly compete with NVIDIA's DGX systems.

On the technical level, AMD's full-stack layout focuses on two major directions: first, attracting developers through an open-source ecosystem; for example, Brium's tools have been integrated into the official PyTorch plugin library, allowing developers to achieve AMD hardware acceleration with simple configurations; second, optimizing edge-cloud collaboration; Untether's chips support hybrid deployment with AMD MI300X GPUs, significantly reducing overall inference latency in smart factory scenarios.

Currently, the AI chip market is still dominated by NVIDIA, which has a 75% market share in data center GPUs, and its new product B200, based on the Blackwell architecture, leads AMD MI325X in inference performance by more than two times.

AMD's response strategy is "differentiated competition": in the high-end market, it aims to capture large model single-node deployment scenarios with MI300X's 1.5TB HBM capacity; in the edge market, it quickly enters fields such as autonomous driving and industrial robotics by leveraging Untether's low-power technology.

Building an open-source ecosystem is key to AMD's breakthrough.

Through acquisitions of companies like Brium, AMD has increased the number of developers for the ROCm framework to 200,000, a 300% growth compared to 2023. However, this number is still significantly smaller compared to NVIDIA's base of 2 million CUDA developers.

Therefore, AMD needs to continue investing in toolchain maturity, documentation support, and other areas. For instance, Brium mentioned in its blog that its tools still require manual adjustments to environment variables to achieve optimal performance, which is an issue that needs urgent resolution Mergers and acquisitions by giants often lead to integration issues, as seen with Intel, which has failed to deliver the expected positive feedback despite several large-scale acquisitions. Therefore, continuous acquisitions pose a test for AMD's integration capabilities.

In 2022, after acquiring Xilinx, AMD took 18 months to complete product line synergy, while the teams from Brium and Untether are scattered across multiple countries, and cultural differences and technical integration may affect organizational and business progress.

Additionally, AMD's R&D expenditure reached $1.73 billion in the first quarter, with about 30% allocated to software ecosystem development, but there is a significant gap compared to NVIDIA's $5.8 billion R&D investment during the same period.

Apart from organizational and cultural integration, achieving the expected market validation is another major challenge facing AMD.

Although the AMD MI325X leads in hardware specifications, its actual deployment has resulted in performance reaching only 60% of the claimed value due to software issues, leading customers to prefer NVIDIA's B200.

Whether the technologies from Brium and Untether can significantly improve this situation by the end of 2025 will determine if AMD can achieve a breakthrough in market share in the AI chip market