AMD Achieves 4X AI Energy Efficiency Gain in Rack-Scale Solutions, Surpasses Mid-2026 Target

AMD claims its 2026 rack-scale AI solution is 4X more energy efficient than its 2024 AI platform -- company says it's pacing ahead of 20X efficiency by 2030

In addition to its normal product and technology roadmaps, AMD has effectively made its 'X-by-year' efficiency programs a recurring engineering roadmap. Back in 2025, the company unveiled its 20x2030 initiative that promised to increase energy efficiency of its rack-scale AI solutions 20 times by 2030. Back then, AMD envisioned that its 2026 rack-scale AI systems will be 3X more energy efficient than its 2024 machines. However, the new estimates published on Tuesday indicate that its latest solutions are 4X more efficient compared to the 2024 baseline. AMD estimates its progress at the rack level rather than on the CPU and AI accelerator levels, which obviously gives the company a lot of freedom in how to optimize performance efficiency. The company takes into account improvements in compute performance, process technology, memory bandwidth, data movement, interconnects, software, and system-level co-design. AMD says three major hardware characteristics determine AI system performance: compute capability, memory bandwidth, and interconnect bandwidth. New process technologies and architectures increase floating-point performance per watt, while improvements in memory integration and high-speed interconnects increase bandwidth available to processors. AMD expects these developments together to produce 20X higher AI performance per watt in 2030 compared to its 2024 baseline. Memory and interconnects play particularly important roles because modern AI systems tend to move enormous amounts of data between accelerators and systems. AMD says higher memory bandwidth, greater bandwidth density, improved bandwidth per watt, larger caches, and tighter integration of memory and compute can reduce wasted energy consumption and increase performance efficiency. Meanwhile, faster scale-up interconnects can improve communication between GPUs, CPUs, and other components, which again increases performance efficiency. AMD's software optimizations are another part of the effort as higher performance achieved with optimizations ultimately means lower power consumption required to achieve an expected result. What is a bit upsetting is that the 4X figure should be treated as an AMD estimate rather than a direct benchmark between two commercially available rack systems. AMD measures progress by comparing annual representative rack configurations with a 2024 baseline using the company's performance-per-watt methodology. In addition, its 2026 calculation combines measurements from actual products with modeled results in cases where final performance numbers were unavailable, which essentially means that AMD does not use its latest Instinct MI455X accelerators for its estimates. If AMD reaches its 2030 targets, the company estimates that around two AMD racks in 2030 could provide the same amount of compute as 570 racks based on the Instinct MI300X from 2024. This essentially means a 20X lower power consumption or 20X higher performance at the same power consumption by 2030. Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.

AMD inches closer to its goal of making AI suck less ... energy

AI's thirst for power remains an ongoing concern, but fear not: AMD says it's making steady progress towards its goal of boosting rack efficiency 20x by the end of the decade. In a blog post published this week, the House of Zen estimates that, as of 2026, its systems are already 4x more efficient than they were in 2024. Certainly, a lot has changed since then. As you may recall, AMD began volume production of the MI300X, its first true datacenter GPU designed for AI, that year. The 750-watt part boasted up to 2.6 petaFLOPS of dense FP8 performance, which at the time made it competitive on paper with Nvidia's Hopper generation of AI accelerators. Since then, AMD has pulled every lever and pushed every button at its disposal to squeeze more FLOPS per watt from its GPU systems, push its memory and scale-up fabrics harder, and optimize its software stack in order to catch up with its larger, more successful rival. This included adding support for 4-bit floating point data types, new memory technologies, increasing interconnect speeds, and transitioning from conventional GPU servers to fully-integrated rack-scale systems. "The counterintuitive thing here... is the bigger the device, the more efficient it is," AMD SVP and Fellow Sam Naffziger told El Reg last year when the chipmaker announced the initiative. Last month, AMD revealed the fruits of its labors with the launch of said rack-scale compute platform, codenamed Helios, which crams 72 MI455X GPUs into a single massive system. Compared to the MI300X, each MI455X boasts between 7.7x and 15.4x higher floating point performance, 2.25x more HBM, 4.4x faster memory, and 4x chip-to-chip interconnect bandwidth. Without question, the chip is faster, but it also requires more than 3x the power. Instead, the biggest performance gains come from just how efficiently AMD can scale AI workloads across the system's six dozen accelerators. As usual, AMD isn't exactly a pioneer here. Nvidia made the leap to rack-scale in late 2024, with the launch of its Grace Blackwell-based NVL72 systems that also pack 72 GPUs into a single rack-sized system. At the time, Nvidia CEO Jensen Huang boasted that compared to an equivalent number of Hopper GPUs, GB200 NVL72 racks delivered a 4x uplift in training and 30x improvement in inference performance. While the benefits of rack-scale architectures are clear, it's worth emphasizing AMD is using a very different methodology to calculate efficiency, by weighting max achieved FLOPS, memory, and interconnect bandwidth differently for training and inference, rather than basing their comparison on real-world application performance. It's also worth nothing that AMD's 4x claim is an estimate. The first Helios units should ship to customers this calendar quarter. We expect the first MLPerf and InferenceX benchmarks to follow not long after. However, assuming AMD can make good on its goals, it says two Helios racks will be able to do the same work that required 570 racks full of kit in 2024. Put more realistically, for the same power customers will be able to deploy 20x more compute, assuming the bubble hasn't already popped by then and taken demand with it. ®

AMD on track for a 20X increase in AI rack-scale energy efficiency by 2030

With the exponential performance improvements we've been seeing from AI rack-scale solutions from companies like AMD, demand for power and cooling has increased. With the wider consensus that AI data centers are anything but efficient, AMD has confirmed it's on track to deliver a 20X increase in rack-scale efficiency by 2030. According to AMD's internal projections, two AMD racks in 2030 will deliver the same compute as 570 AMD Instinct MI300 Series-powered racks from 2024. This represents a 20X-fold increase in energy efficiency and a 28X reduction in "carbon intensity." Of course, this isn't to say that we're going to see a reduction in the amount of racks being deployed, so the probable result will end up being the same amount of energy being used with a 20X increase in compute performance measured in FLOPs. AMD adds that its goal is to "deliver substantially more compute performance without requiring energy consumption to grow at the same pace," so that's a plus for those concerned about the sheer amount of power currently being used up by AI data centers. In addition, AMD has confirmed it has achieved an estimated 4X increase in AI energy efficiency as of mid-2026, ahead of schedule. This would include the recent announcement of the new AMD Instinct MI455X GPU built on a 2nm process, a key part of the new AMD Helios rack. "The next wave of AI efficiency will depend on tighter co-optimization across compute silicon, memory, interconnects, software and rack-scale system design," Sam Naffziger, Senior Vice President and Corporate Fellow, AMD said. "Our estimated 4X improvement through 2026 reflects the strength of our approach and the progress AMD is making across the full system, putting us ahead of our projected pace toward the 2030 goal."

AMD Claims Just Two Of Its 2030 AI Racks Will Match 570 Of Its MI300X Racks, Slashing Electricity Use By 20x

AMD is making big claims today on the rapid efficiency gains that its rack-scale AI offerings continue to scale in pursuit of its lofty 2030 goals, which aim to squeeze the performance footprint of as many as 570 of AMD's MI300X-based racks within just two of its 2030-launching ones. AMD claims that its 2030-launching AI racks will be 20x more energy efficient relative to the MI300X-based racks that entered volume production in 2024 AMD has disclosed some very interesting statistics today, noting that it "has achieved an estimated 4x increase in AI energy efficiency as of mid-2026, ahead of its 3x projected target for this stage and more than double the historical trendline for the same point." What's more, the company says that it "is building momentum toward its 2030 goal to deliver a 20x increase in rack-scale energy efficiency for AI training and inference." AMD aims to achieve this feat by: Even so, in what is perhaps the most interesting statement in the entire press release, AMD now claims that "approximately two 2030 AMD racks are expected to deliver the same compute as 570 racks in 2024, enabling a reduction in use-phase electricity by 20x and carbon intensity by 28x." Let's parse this statement. The MI300X AI accelerators entered volume production in 2024. And so, we can presume that AMD has used its racks based on those GPUs as the reference point. If just two of its 2030-launching AI racks will offer the same performance as 570 of its MI300-based racks, then effective rack count for the same workload decreases by 285x (570/2). However, AMD is claiming that energy consumption will fall by only 20x. This suggests that each 2030-launching AI rack will consume 14.25x (285/20) the power load of each MI300X-based AI rack, which require anywhere between 40 and 125 KW of power. Now, to put all of this information together, this means that each of AMD's 2030-launching AI racks will consume between 570 KW and 1,781.25 KW of power. Of course, AMD has already taken a significant step towards the densification of its rack-scale solutions with the just-launched Helios rack, which is the company's first full-stack offering for AI workloads, and features: Follow Wccftech on Google to get more of our news coverage in your feeds.

AMD Achieves 4x AI Energy Efficiency Gain, Surpasses Mid-2026 Target

This progress shows AMD is building momentum toward its 2030 goal to deliver a 20x increase in rack-scale energy efficiency for AI training and inference. It also reflects the work being done by AMD across silicon, systems and software to help customers scale AI workloads more efficiently. AMD has achieved an estimated 4x increase in AI energy efficiency as of mid-2026, ahead of its 3x projected target for this stage and more than double the historical trendline for the same point. This progress shows AMD is building momentum toward its 2030 goal to deliver a 20x increase in rack-scale energy efficiency for AI training and inference. It also reflects the work being done by AMD across silicon, systems and software to help customers scale AI workloads more efficiently. AI compute is accelerating rapidly, raising the bar for what data center infrastructure needs to deliver. Meeting that demand requires innovation across the whole stack: CPUs, GPUs, networking, software, power and cooling all need to work together efficiently at the rack and data center level. Improving energy efficiency can help deliver more AI performance without increasing power, improve total cost of ownership and help customers scale faster. "The next wave of AI efficiency will depend on tighter co-optimization across compute silicon, memory, interconnects, software and rack-scale system design. Our estimated 4x improvement through 2026 reflects the strength of our approach and the progress AMD is making across the full system, putting us ahead of our projected pace toward the 2030 goal." -- Sam Naffziger, senior vice president and Corporate Fellow, AMD Based on a representative AI training workload, AMD projected rack-scale energy-efficiency gains are expected to produce one of two related benefits by 2030:3 AMD is working to increase efficiency across the full AI stack through advances in compute architecture, process technology, memory bandwidth, data movement, interconnects, software and system-level co-design. The goal is to deliver substantially more compute performance without requiring energy consumption to grow at the same pace. As AI infrastructure scales from individual nodes to full racks, efficiency increasingly depends on how well CPUs, GPUs, memory, networking, storage and software work together. AMD is applying system-level co-design, with teams working across product categories, to help reduce bottlenecks, move data more efficiently and improve performance per watt across the platform. Beyond software optimization, three primary factors drive AI system performance: compute capability, memory bandwidth and interconnect bandwidth. Advanced process technology and architectural improvements help increase floating-point compute performance per watt. Advanced architectures and memory integration improve bandwidth, while high-speed interconnects and tighter integration increase the network bandwidths. Together, AMD expects these innovations to deliver 20x more AI performance per watt by 2030 compared with 2024. Modern AI workloads depend on moving large amounts of data efficiently. Improving memory bandwidth, bandwidth density and bandwidth per watt helps keep compute engines fed while reducing wasted power. High-bandwidth memory, larger caches and tighter integration between memory and compute can reduce unnecessary data movement, which costs energy. Interconnect technology has also become an important part of efficient AI infrastructure as larger models and workloads increasingly depend on how efficiently GPUs, CPUs and other system components work together. High-speed scale-up interconnects help larger systems reduce bottlenecks and share data more efficiently. Hardware provides the foundation, but software helps unlock continued efficiency gains. The AMD ROCm™ software stack, open standards and close customer collaboration help developers and organizations deploy AI workloads more efficiently on AMD platforms. Open ecosystems also support broader optimization across AI and high-performance computing (HPC) deployments. These optimizations can help improve both large-scale AI training and inference workloads, where improving throughput and reducing energy consumed per generated token are increasingly important as AI applications scale. Customers need more useful compute for the energy available. The projected rack-scale efficiency gains AMD has set forth can enable AI and HPC customers to achieve data center-level gains in power and cooling infrastructure, which can be limiting factors to increasing compute performance. Better energy efficiency can help reduce operational electricity use, improve total cost of ownership, and support more sustainable AI and HPC growth.