Benchmarking RDS SQL Server on the New AMD EPYC (M8a/R8a) Instances

4–6 minutes
Header graphic for a blog post benchmarking AMD EPYC M8a and R8a instances on Amazon RDS for SQL Server

Who this is for: anyone running RDS for SQL Server on Intel-based instance classes who’s wondering whether the new AMD EPYC-powered M8a/R8a families are worth a resize, especially if SQL Server licensing costs are a real line item for you.

I moved one of our reporting replicas from a db.r6i.2xlarge to a db.r8a.2xlarge mostly out of curiosity, expecting a modest bump and nothing worth writing about. What actually surprised me wasn’t raw throughput, it was that I could drop down an instance size on the next test and still keep pace with the old r6i, which is the part that actually moves the needle on our Microsoft licensing bill.

Why This Matters

AWS added 5th Gen AMD EPYC-powered M8a and R8a instance classes to RDS for SQL Server in May 2026. AWS’s own numbers claim up to 70% higher throughput than comparable x86 Intel instances at the same size for common workloads, and AMD’s published benchmarks show R8a.xlarge delivering up to 63% more transactions per second than the equivalent R8i (Intel) instance. That’s a big enough number that it’s worth verifying yourself rather than taking it on faith, especially because SQL Server licensing on RDS is billed per vCPU, so anything that lets you run the same workload on fewer or smaller instances is a direct cost reduction, not just a performance nicety.

What we actually measured

M8a gives you a 1 vCPU to 4 GiB RAM ratio, R8a gives you 1 vCPU to 8 GiB, both on the same 5th Gen EPYC silicon, with each vCPU mapped to a physical core rather than a hyperthread. We ran our standard reporting workload (a mix of large aggregation queries and a nightly ETL batch) on a db.r6i.2xlarge (Intel, 8 vCPU / 64 GiB) and a db.r8a.2xlarge (AMD, same shape) for a week each, same storage class, same parameter group.

aws rds modify-db-instance \
--db-instance-identifier reporting-replica \
--db-instance-class db.r8a.2xlarge \
--apply-immediately
-- Query used to track batch completion time and CPU usage per run
SELECT
session_id,
DATEDIFF(SECOND, start_time, GETDATE()) AS duration_sec,
cpu_time,
total_elapsed_time
FROM sys.dm_exec_requests
WHERE session_id = @@SPID;

Our nightly ETL batch finished 34% faster on r8a.2xlarge with no other changes. That’s short of AWS’s headline 70% figure, which makes sense since our workload is a mix of CPU-bound aggregation and I/O-bound bulk loads, not a pure OLTP transaction benchmark. When we tried a smaller db.r8a.xlarge (4 vCPU / 32 GiB) against the same workload, batch time was still about 9% faster than the original r6i.2xlarge despite having half the vCPUs, which is the number that actually changed our instance sizing plan.

The licensing angle: Optimize CPU

M8a and R8a also support the RDS “optimize CPU” feature, which lets you dial down the number of vCPUs actually enabled on the instance without changing the instance class itself. Since SQL Server licensing on RDS is billed per enabled vCPU, this is the mechanism that turns a performance win into a cost win: if a workload hits its target throughput using fewer enabled vCPUs than the instance physically has, you can cap it there and stop paying SQL Server license fees for the difference.

aws rds modify-db-instance \
--db-instance-identifier reporting-replica \
--processor-features "Name=coreCount,Value=4" "Name=threadsPerCore,Value=1" \
--apply-immediately

On our r8a.2xlarge test, capping to 4 enabled cores still comfortably beat the old r6i.2xlarge’s batch times, which meant we could size down our SQL Server Standard Edition licensing footprint on that instance by half.

Where This Requires Care

M8a and R8a launched in a limited set of regions (N. Virginia, Ohio, and Oregon among the first). If your RDS footprint is in a region that isn’t supported yet, there’s nothing to test.

The throughput gains AWS and AMD publish are workload-dependent. Our 34% batch-time improvement is a real number for our workload, not a universal one; a pure OLTP transaction workload with small, latency-sensitive queries will see a different delta than a batch-heavy reporting workload like ours.

Resizing to a new instance family (rather than just a new size within the same family) triggers a reboot on single-AZ instances, or a failover on Multi-AZ. Plan the change during a maintenance window, not mid-day.

Before you touch the “optimize CPU” processor-features setting on a production instance, test it on a replica first. Under-provisioning enabled cores on a workload that occasionally spikes will show up as query timeouts, not a clean licensing win.

Quick Reference

  • M8a is 1 vCPU : 4 GiB RAM, R8a is 1 vCPU : 8 GiB, both on 5th Gen AMD EPYC with 1 vCPU per physical core.
  • Benchmark your own workload before trusting AWS’s or AMD’s headline throughput percentages; ours landed well under the published figures.
  • Use RDS “optimize CPU” (processor-features) to cap enabled vCPUs and directly reduce per-vCPU SQL Server licensing cost once you’ve confirmed the instance can hold throughput with fewer cores.
  • Check regional availability before planning a migration; M8a/R8a rolled out to a limited region list first.
  • Test any processor-features change on a non-production replica first; under-provisioning shows up as timeouts under load spikes.

My Take

The AMD EPYC instances are a genuine upgrade, not a marketing exercise, but the number that matters for most SQL Server shops isn’t the raw throughput percentage, it’s whether the “optimize CPU” feature lets you shrink your licensed core count while holding performance steady. That’s where the real savings showed up for us. I’d test this on a read replica or a non-critical reporting instance first, since the throughput delta is workload-dependent enough that you shouldn’t resize a primary based on a vendor benchmark alone.

Further Reading

Arvind Toorpu

Arvind Toorpu

I am a Database Admin/Architect with over 14+ years of experience managing multiple DBMS systems

Let’s connect

Archives