The three-way CPU race: how AMD flipped the script on EC2

Published September 20, 2025

For years, AWS capacity planning followed a predictable mental model:

1. Intel was the default baseline.
2. AMD was the budget drop-in replacement, priced at a flat 10% discount for roughly equal performance.
3. Graviton was the 20% discount tier for teams willing to recompile for Arm.

That rule held true through the 5th and 6th EC2 generations. If you ran an Intel m6i instance and swapped it for an AMD m6a instance, your hourly bill dropped by 10% without altering a line of application code.

With the 7th generation, AWS broke that formula. AMD EPYC “Genoa” instances are not cheaper than Intel. They cost 15% more per hour. Yet for many CPU-heavy workloads, paying that premium reduces total monthly cloud spend.

Here is the data behind the shift, what the benchmarks show across generations, and how to choose the right processor for your workload.

The 6th generation: when AMD was the budget brand

To understand how the pricing model changed, look at the 6th generation general-purpose instances in us-east-1 (Linux on-demand):

In the 6th generation, the AMD m6a.xlarge delivered 20.6% more CoreMark throughput than the Intel m6i.xlarge while costing 10% less ($0.1728 vs. $0.1920 per hour). It was an easy win: higher performance at a lower price on the same x86_64 instruction set.

Meanwhile, m6g.xlarge delivered 500,013 CoreMark per dollar, establishing Graviton as the efficiency leader for Arm-compatible workloads.

The 7th generation inversion: AMD Genoa takes the crown

When AWS introduced 7th generation instances based on AMD’s Zen 4 architecture (EPYC 9R14 “Genoa”), the pricing hierarchy flipped.

An m7a.xlarge costs $0.23184 per hour, which is 15.0% more expensive than an m7i.xlarge ($0.20160).

AWS priced it higher because the performance delta is massive. At 131,219 CoreMark, m7a.xlarge generates 47.2% more throughput than m7i.xlarge (89,171) and 27.3% more than m7g.xlarge (103,066). Per vCPU, AMD delivers 32,805 CoreMark compared to Intel’s 22,293.

The compute-optimized 7th generation family follows the exact same pattern:

On compute-optimized nodes, c7a.xlarge carries a 15.0% higher sticker price than c7i.xlarge, but provides 36.3% higher CoreMark throughput. Even on pure work-per-dollar metrics, c7a.xlarge (648,295 CoreMark/$) easily beats c7i.xlarge (547,070 CoreMark/$).

The math of downsizing: how a more expensive instance cuts costs

Paying 15% more per instance hour sounds counterintuitive for cost optimization. However, the 47% per-vCPU throughput gain changes cluster sizing math in two specific scenarios.

1. Downsizing instance sizes

If your application is CPU-bound and constrained by single-thread speed or core throughput, instance sizing is often dictated by compute capacity rather than RAM.

Consider a service running on an older 8-vCPU Intel instance or struggling near saturation on a 4-vCPU m7i.xlarge:

• Option A (Intel 7th Gen): m7i.2xlarge (8 vCPUs, 32 GiB RAM) costs $0.4032/hr and delivers ~178,342 CoreMark.
• Option B (AMD 7th Gen): m7a.xlarge (4 vCPUs, 16 GiB RAM) costs $0.2318/hr and delivers ~131,219 CoreMark.

If your workload requires ~120,000 CoreMark of compute capacity, a 4-vCPU AMD Genoa instance can handle the load that would otherwise force you into an 8-vCPU Intel instance. That drops your hourly cost from $0.4032 to $0.2318, a 42.5% net savings.

2. Shrinking Auto Scaling Groups

For horizontally scaled web fleets that scale on CPU utilization targets (such as 70% target utilization):

• A cluster of 10 m7i.xlarge instances costs $2.016/hr and provides ~891,710 total CoreMark.
• The equivalent compute capacity can be delivered by 7 m7a.xlarge instances, providing ~918,533 total CoreMark at $1.623/hr.

By switching to AMD Genoa, the fleet runs fewer nodes, reducing total hourly spend by 19.5% while maintaining identical headroom.

Where Graviton wins: total efficiency at scale

While AMD took the performance-per-core crown in x86, AWS Graviton remains the cost efficiency winner for applications that run on Arm.

Look at the CoreMark delivered per dollar spent across 7th and 8th generation instances:

• m7i.xlarge (Intel): 442,316 CoreMark / $
• m7a.xlarge (AMD): 565,989 CoreMark / $
• m7g.xlarge (Graviton3): 631,532 CoreMark / $
• m8g.xlarge (Graviton4): 694,513 CoreMark / $

Graviton3 (m7g.xlarge) costs 29.6% less per hour than m7a.xlarge ($0.1632 vs. $0.2318) while delivering roughly 78.5% of the total throughput. For scale-out microservices, queuing consumers, batch workers, and stateless API servers where work distributes evenly across many threads, Graviton yields the lowest cost per unit of work.

The 8th generation preview: Granite Rapids vs. Graviton4

The 8th generation widens these trends further. Comparing the initial 8th generation general-purpose offerings in us-east-1:

Intel’s Granite Rapids in m8i.xlarge advertises a 3.9 GHz clock speed, but its overall benchmark throughput on standard compute benchmarks sits at 88,968 CoreMark at $0.21168/hr.

By contrast, Graviton4 on m8g.xlarge delivers 124,679 CoreMark (31,170 per vCPU) at $0.17952/hr. That is 40.1% higher benchmark throughput at a 15.2% lower hourly cost, producing 694,513 CoreMark per dollar.

Why Intel is squeezed from above and below

Intel faces structural pressure from both directions on EC2:

1. Squeezed from below by Graviton: For standard Linux workloads running Java, Python, Go, Node.js, or containerized services, Graviton4 delivers substantially better performance at a 15% to 20% lower hourly price.
2. Squeezed from above by AMD: For workloads that require the x86 architecture, AMD Genoa (m7a, c7a) delivers roughly 47% higher throughput per vCPU. Even with a 15% price premium, AMD offers better work-per-dollar and enables node downsizing that Intel cannot match.

Unless a workload relies on proprietary Intel extensions (such as Intel AMX for specialized matrix operations or AVX-512 routines specifically tuned for Intel pipelines) or legacy binary compatibility that cannot run on AMD, Intel instances are rarely the optimal choice on price-performance metrics.

How to choose for your workload

When selecting instance types for your next deployment:

• Default to Graviton (m7g, m8g): If your stack compiles to arm64 (most containerized web applications, background workers, and open-source data stores), Graviton gives you the highest total compute throughput per dollar on AWS.
• Choose AMD Genoa (m7a, c7a): If you must stay on x86_64, have single-thread latency requirements, run CPU-bound workloads, or pay for commercial software licensed per-core. The higher per-core performance often allows you to downsize instance tiers or run smaller auto-scaling groups.
• Choose Intel (m7i, m8i): If you have software certified exclusively for Intel Xeon hardware or require Intel-specific instruction sets like AMX.

To compare exact pricing and benchmark numbers across all sizes and regions, explore the instance explorer or evaluate specific instance pairs side-by-side in the comparison tool.