In the world of AI, there’s a desperate search for an alternative to Nvidia’s GPUs, and AMD is stepping up to the plate. AMD detailed its updated GPU roadmap, with new hardware coming in 2024, 2025 and 2026. AMD previously released GPUs on a roughly 1.5- to two-year cycle, but overwhelming demand for AI hardware is weighing on the company.
For the first time, AMD announced the MI400 GPUs, which the company will release in 2026. It is positioned as the Nvidia GPU killer for 2026. AMD’s plans for 2024 include the MI325X GPU, followed by the 3-nanometer MI350 next year in 2025, with both based on HBM3E memory.
AMD didn’t share any more details on the MI400 GPUs other than saying it would be for inferencing and training and based on CDNA-Next architecture.
AMD Roadmap: Finally Clear
AMD’s GPU sales are growing nicely, but the company hasn’t provided a clear roadmap until this week. Analysts were uncomfortable about a lack of information on AMD’s GPU plans in 2025 and 2026, considering Nvidia outlined annual GPU plans in 2025 and 2026 last year.
Lisa Su, AMD’s CEO, said the company’s roadmap was driven by customer feedback.
“They’re… giving us significant feedback on the roadmap and what we need to meet their needs. Our chiplet architecture is actually very flexible, and so that allows us to actually make changes to the roadmap as necessary,” Su said on an April earnings call, according to a transcript on The Motley Fool.
AMD’s approach aligns with the fast-changing computing requirements of AI workloads. Training AI models drove initial GPU sales, and there’s a wider focus on answering user queries against LLMs.
Intel is also relying on customer feedback for its GPU roadmap. The company last scrapped its Falcon Shores CPU-GPU integrated chip after a lack of customer interest. The chip did not meet customer requirements, and Intel is now rearchitecting the chip as a discrete GPU-only product due next year.
Still Far Behind Nvidia
Su also said on the earnings call that AMD’s MI300 sales are exploding and totaled $1 billion in less than two quarters.
“We now expect data center GPU revenue to exceed $4 billion in 2024, up from the $3.5 billion we guided in January,” Su said.
AMD’s data center revenue in the most recent quarter was $2.3 billion, compared to Nvidia’s $22.6 billion.
AMD is still trailing Nvidia in market share but is seeing wider customer adoption.
Microsoft helped AMD develop MI300X GPUs and announced last month that it was offering the accelerator in its ND MI300X v5 VM on Azure.
“It offers the best price-performance on GPT-4 for inference,” said Satya Nadella, Microsoft’s CEO, during a keynote speech at last month’s Microsoft’s Build conference.
A block of eight AMD MI300X GPUs includes 1.5 TB of HBM and 5.3 terabits per second of HBM bandwidth. Microsoft offers a wide range of GPUs and plans to bring Nvidia’s Blackwell GPUs to Azure.
The Upcoming GPUs
AMD’s 2024 GPU includes the MI325X accelerator, which will be out by the end of this year. It will be an upgrade to the current MI300 series and will be available by the end of the year.
The 325X will include 288GB of HBM3E memory and six terabytes per second of memory bandwidth. Nvidia has put HBM3E memory in its H200 GPU. The MI325X has backward compatibility, suggesting it is only a memory upgrade.
AMD’s 2025 GPU is the MI350 series, which the company claimed will deliver a 35x boost in inferencing compared to the MI300 series GPUs. Beyond the claim, AMD is trying to highlight the architectural improvements for inferencing, which is becoming more important than training.
The MI350 will be made using the 3-nanometer process.
Inferencing allows low-power AI processing to answer questions on trained models, and the claim is an indirect way of AMD boasting improvements for lower-precision FP4 or FP6 data types. Nvidia has put the FP4 and FP6 data types in its Blackwell GPU, which was announced in March.
AMD’s CPU Improvements
AMD said its 5th Gen AMD processors, code-named “Turin“ and based on the Zen 5 x86 architecture, will ship within the next six months. The chips, with up to 192 cores and 384 threads, will compete with Intel’s Granite Rapids CPUs.
The company also announced the Embedded+ chip design, which pairs its Ryzen CPU with the Versal AI chip. The chip will go into embedded products, such as robots, which need AI capabilities on the edge. Typically, AI chips such as FPGAs or ASICs can’t boot up devices on their own and rely on CPUs like AMD’s Ryzen for instructions and offload capabilities.
The Embedded+ chip architecture is a result of AMD uniting its CPU technology with Xilinx’s ASIC/FPGA designs.
ROCm IT
AMD said it continued to mature its software development tools, ROCm, to support the new GPUs.
Lawrence Livermore National Laboratory announced in 2020 that El Capitan software development was based on ROCm. ROCm is based on open-source tools and includes libraries and development tools that allow customers to tune their software tools to AMD’s GPUs.
AMD has acknowledged that it has a lot of work to do on developer support for AI. There’s no coordination on software releases with support for tensor cores in the new GPUs, though the MI325 GPU backward compatibility suggests no environmental changes.
Nvidia coordinates the releases of its CUDA versions with the release of new GPUs. Nvidia GPUs typically have newly architected tensor cores and GPU improvements, which involve new libraries that developers can use to fine-tune their applications for the new GPUs.
However, CUDA is proprietary, which means Nvidia locks customers into buying their software development tools, such as AI Enterprise, to take advantage of its latest GPUs.
Major chip makers, including Intel and AMD, launched the UXL Foundation last year in a bid to bring some kind of competition to CUDA. The foundation’s base is Intel’s OneAPI, which has a SYCLomatic tool that strips out CUDA code so AI loads on a wide range of GPUs and accelerators. Argonne National Laboratory relied on Intel’s OneAPI to develop software for the Aurora supercomputer.
Topics:AI, Edge, Hardware, Processors
Sectors:, Community, Energy, Entertainment, Financial Services, Government, Life Sciences, Manufacturing, , Retail, semiconductor, ,
Tags:CUDA, GPT-4, GPU, HBM, MI300X, MI325X, MI350, OneAPI, ROCm, Turin, Versal