AMD issued a raft of reports at their Advancing AI 2025 occasion this week, an replace on the corporate’s response to NVIDIA’s 90-plus p.c market share dominance within the GPU and AI markets. And the corporate supplied a sneak peak at what to anticipate from their subsequent era of EPYC CPUs and Intuition GPUs.
Right here’s an outline of AMD’s main announcement:
AMD MI350 Collection GPUs
The headline announcement: AMD launched the Intuition MI350 Collection that they mentioned delivers as much as 4x generation-on-generation AI compute enchancment and as much as a 35x leap in inferencing efficiency.
They provide reminiscence capability of 288GB HBM3E and bandwidth of as much as 8TB/s together with air-cooled and direct liquid-cooled configurations.
And so they assist as much as 64 GPUs in an air-cooled rack and as much as 128 GPUs in a direct liquid-cooled racks delivering as much as 2.6 exaFLOPS of FP4/FP6 efficiency in an {industry} standards-based infrastructure.
AMD CEO Lisa Su
“With the MI350 sequence, we’re delivering the biggest generational efficiency leap within the historical past of Intuition, and we’re already deep in improvement of MI400 for 2026.” AMD CEO, Dr. Lisa Su, mentioned. “[The MI400] is admittedly designed from the bottom up as a rack-level resolution.”
On that entrance, AMD introduced its “Helios” rack scale structure, obtainable subsequent 12 months, that may combine a mixture of the subsequent era of AMD expertise together with:
- Subsequent-Gen AMD Intuition MI400 Collection GPUs, that are anticipated to supply as much as 432 GB of HBM4 reminiscence, 40 petaflops of FP4 efficiency and 300 gigabytes per second of scale-out bandwidth3.
- Helios efficiency scales throughout 72 GPUs utilizing the open commonplace UALink (Extremely Accelerator Hyperlink)to interconnect the GPUs and scale-out NICs. That is designed to let each GPU within the rack talk as one unified system.
- sixth Gen AMD EPYC “Venice” CPUs, which can make the most of the “Zen 6” structure and are anticipated to supply as much as 256 cores, as much as 1.7X the efficiency and 1.6 TBs of reminiscence bandwidth.
- AMD Pensando “Vulcano” AI NICs, which is UEC (Extremely Ethernet Consortium) 1.0 compliant and helps each PCIe and UALink interfaces for connectivity to CPUs and GPUs. It can additionally assist 800G community throughput and an anticipated 8x the scale-out bandwidth per GPU3 in comparison with the earlier era.
ROCm 7 and Developer Cloud
A serious space of benefit for NVIDIA is its dominance of the software program improvement enviornment – the overwhelming majority of AI utility builders use NVIDIA’s CUDA programming platform. Builders who turn into adept at utilizing CUDA are likely to proceed utilizing… CUDA. Customers of functions constructed on CUDA are likely to need… serves utilizing NVIDIA GPUs. Together with GPU efficiency, competing with NVIDIA on the AI software program entrance is a significant problem for anybody making an attempt to carve out a considerable share of the AI market.
This week, AMD launched AMD ROCm 7 and the AMD Developer Cloud below what the corporate known as a “builders first” mantra.
“Over the previous 12 months, we’ve shifted our focus to enhancing our inference and coaching capabilities throughout key fashions and frameworks and increasing our buyer base,” mentioned Anush Elangovan, VP of AI Software program, AMD, in an announcement weblog. “Main fashions like llama 4, gemma 3, and Deepseek are actually supported from day one, and our collaboration with the open-source neighborhood has by no means been stronger, underscoring our dedication to fostering an accessible and progressive AI ecosystem.”
Elangovan emphasised ROCm 7’s accessibility and scalability, together with “placing MI300X-class GPUs within the fingers of anybody with a GitHub ID…, putting in ROCm with a easy pip set up…, going from zero to Triton kernel pocket book in minutes.”
Usually obtainable in Q3 2025, ROCm 7 will ship greater than 3.5X the inference functionality and 3X the coaching energy in comparison with ROCm 6. This stems from advances in usability, efficiency, and assist for decrease precision information varieties like FP4 and FP6, Elangovan mentioned. ROCm 7 additionally affords “a sturdy strategy” to distributed inference, the results of collaboration with the open-source ecosystem, together with such frameworks as SGLang, vLLM and llm-d.
AMD’s ROCm Enterprise AI debuts as an MLOps platform designed for AI operations in enterprise settings and consists of instruments for mannequin tuning with industry-specific information and integration with structured and unstructured workflows. AMD mentioned that is facilitated by partnerships “inside our ecosystem for creating reference functions like chatbots and doc summarizations.”
Rack-Scale Power Effectivity
For the urgent drawback of AI vitality demand outstripping vitality provide, AMD mentioned it exceeded its “30×25” effectivity aim, reaching a 38x enhance in node-level vitality effectivity for AI-training and HPC, which the corporate mentioned equates to a 97 p.c discount in vitality for a similar efficiency in comparison with techniques from 5 years in the past.
The corporate additionally set a 2030 aim to ship a 20x enhance in rack-scale vitality effectivity from a 2024 base 12 months, enabling a typical AI mannequin that at the moment requires greater than 275 racks to be educated in below one rack by 2030, utilizing 95 p.c much less electrical energy.
Mixed with anticipated software program and algorithmic advances, AMD mentioned the brand new aim may allow as much as a 100x enchancment in general vitality effectivity.
Open Rack Scale AI Infrastructure
AMD introduced its rack structure for AI encompassing its fifth Gen EPYC CPUs, Intuition MI350 Collection GPUs, and scale-out networking options together with AMD Pensando Pollara AI NIC, built-in into an industry-standard Open Compute Challenge- and Extremely Ethernet Consortium-compliant design.
“By combining all our {hardware} parts right into a single rack resolution, we’re enabling a brand new class of differentiated, high-performance AI infrastructure in each liquid and air-cooled configurations,” AMD mentioned.
