AWS just released its Graviton5 processor to general availability. The custom Arm chip powers new M9g and M9gd instances that promise better performance for databases, web applications and agentic AI tasks. Yet independent tests published today reveal a more nuanced picture. Graviton5 beats Intel’s latest Xeon but falls short of AMD’s EPYC Turin across most workloads.
The Phoronix review tested comparable 16 vCPU instances. It used m9g.4xlarge for Graviton5, m8a.4xlarge for AMD EPYC Turin and m8i.4xlarge for Intel Xeon 6 Granite Rapids. All ran with 64GB of memory under Ubuntu 26.04. On-demand hourly prices came in at $0.78 for the Graviton5 instance, $0.97 for AMD and $0.85 for Intel. Those numbers matter. They frame every performance-per-dollar calculation that cloud buyers actually use.
But first the silicon itself. AWS designed Graviton5 with up to 192 cores based on Arm Neoverse V3 architecture. It carries 180MB of L3 cache. That’s five times larger than the previous generation. Memory support hits DDR5-8800. Inter-core latency drops by 33 percent. AWS News Blog claims up to 25 percent better compute performance than Graviton4. Web apps and machine learning inference can see gains as high as 35 percent. Databases improve around 30 percent.
Meta already deploys Graviton5 at scale for agentic AI workloads. The company runs real-time reasoning, code generation and multi-step orchestration. Those tasks love large caches and low latency. They punish slow data movement between cores. So the hardware bets make sense. Still, raw benchmark results tell a different story when x86 competitors enter the ring.
Phoronix ran more than 140 tests. The geometric mean across all that ran successfully on the three platforms showed AMD’s m8a instance delivering 1.6 times the performance of Intel’s m8i. It reached 2.26 times the Graviton5 result. AMD won 91.7 percent of the individual tests. Not close in many cases. Graviton5 did outperform the Intel Xeon option. It also improved roughly 10 to 15 percent over Graviton4. Small victories. They don’t change the overall ranking.
Look at specific workloads. In quantitative finance tests using LIBXSMM and OpenFOAM, Graviton5 beat the Xeon but trailed Turin. JSON parsing and OCUDU followed the same pattern. Cryptography, video encoding and 7-Zip compression favored Graviton5 over Intel yet still lost to AMD. Code compilation, regular expressions, DSP and ClickHouse queries produced similar outcomes. Database tests gave Graviton5 an edge in some scenarios. Python benchmarks, ncnn, ASTC, GROMACS, sysbench and Nginx all saw the Arm chip lead the Intel part while AMD pulled ahead.
Power efficiency receives less concrete data in the review. AWS promotes its chip as the most energy efficient the company has built. The Tom’s Hardware article from December highlighted the 192-core design and massive cache as direct challenges to high-end EPYC and Xeon processors in AWS data centers. Replacement potential exists. Actual measured watts per workload remain sparse in public tests so far.
And pricing tells its own tale. The lower hourly rate for m9g helps close the gap on performance-per-dollar. In some earlier Graviton generations that advantage proved decisive. Here AMD’s raw speed often overcomes its higher instance cost. Cloud buyers must run their own workloads. No single geometric mean captures every application. Some shops will still pick Graviton5 for the efficiency and ecosystem lock-in.
Red Hat already validated Enterprise Linux on the new instances. Red Hat blog noted the preview availability back in March. Support matters for enterprises wary of Arm compatibility. Most modern software runs fine. Legacy x86 binaries or specific libraries can create friction. Those edge cases explain why Intel and AMD still win broad benchmark suites.
Agentic AI changes the equation. These workloads involve continuous loops of planning, tool use, verification and state management. They generate heavy CPU demand separate from GPU inference. Low tail latency becomes critical. Large on-chip cache reduces trips to main memory. Graviton5’s 180MB L3 and reduced inter-core latency target exactly those patterns. Meta’s deployment signals confidence. Other large customers have reportedly queued up capacity.
Availability rolled out first in US East, US West and Europe Frankfurt. Both M9g for general purpose and M9gd with local NVMe storage now sit in general availability. Customers tested them extensively since the re:Invent 2025 preview. Early feedback focused on the performance jump for databases and AI agents. Sustainability gains from better energy efficiency also factor into procurement decisions at organizations with aggressive carbon targets.
Still the Phoronix numbers hit hard. AMD’s EPYC Turin simply delivered more throughput in the majority of tests. Its performance-per-dollar lead across nearly every workload category surprised some observers. Intel lagged in most areas except those that benefit from its AMX instructions. Graviton5 carved out a respectable middle ground. Faster than Xeon. Cheaper than both. Not the fastest absolute performer.
That middle ground may prove exactly what many buyers want. Cloud economics rarely optimize for peak benchmark scores. Total cost of ownership includes software licensing, operational simplicity, power draw and carbon accounting. Graviton5 improves on all those dimensions even when raw speed trails. The 25 percent generational gain over Graviton4 compounds across large fleets.
Future tests will expand. Larger instance sizes with the full 192 cores should change the picture. Memory bandwidth and cache behavior scale differently at higher core counts. The current Phoronix comparison used modest 16 vCPU configurations to keep pricing and test times manageable. Real production deployments often run much bigger.
So AWS continues its custom silicon march. Graviton5 represents the latest step in a strategy that began years ago with early Arm designs. Each generation narrows the gap with merchant x86 chips while offering AWS unique control over cost and efficiency. AMD and Intel respond with their own dense core designs and specialized instructions. The competition benefits customers. It forces every vendor to improve.
Buyers should test. Run their actual code on m9g, m8a and m8i. Measure not just throughput but also latency distributions, power consumption and monthly bills. The benchmark averages provide direction. They never replace workload-specific data. In cloud infrastructure the right answer almost always depends on the details.
Graviton5 won’t displace EPYC Turin for every workload. It probably shouldn’t. The Arm chip does deliver enough advantages in cost, efficiency and targeted AI performance to claim a growing share of AWS compute. That shift matters for the broader industry. Custom silicon from hyperscalers now sets the pace in multiple categories. Traditional CPU makers must match both performance and total ownership costs to stay relevant.
WebProNews is an iEntry Publication