
Introduction
Picking an EBS volume for an EC2 workload usually comes down to two options. gp3 gives you flexible general-purpose SSD performance. io2 gives you provisioned IOPS, higher durability, and lower latency.
The wrong call costs real money or real performance. Overbuy io2 and you pay a premium for headroom nobody uses. Underbuy and a transactional database stalls at peak load.
Teams also run into instance limits, account quotas, and per-volume billing quirks that change the math. This guide compares the two on pricing, IOPS, throughput, latency, durability, Multi-Attach, and workload fit. Prices are us-east-1 USD as of November 18, 2026. Confirm current figures in AWS documentation and the AWS Pricing Calculator before you commit.
Key Takeaways
- gp3 is general-purpose SSD with 3,000 IOPS included; io2 is provisioned-IOPS SSD built for critical databases.
- Start with gp3 when measured IOPS, throughput, and latency fit its limits.
- io2 earns its premium for 99.999% durability, sub-millisecond latency, Multi-Attach, or IOPS beyond 80,000 per volume.
- Check peak IOPS, queue length, latency, and EC2 limits before choosing.
AWS EBS gp3 vs io2: Quick Comparison
| Attribute | gp3 | io2 (all io2 now runs on Block Express) |
|---|---|---|
| Volume size | 1 GiB-64 TiB | 4 GiB-64 TiB |
| Max IOPS per volume | 80,000 | 256,000 |
| Max throughput per volume | 2,000 MiB/s | 4,000 MiB/s |
| Designed durability | 99.8%-99.9% | 99.999% |
| Annual failure rate | 0.1%-0.2% | 0.001% |
| Latency | Single-digit milliseconds | Average under 500 microseconds for 16 KiB I/O on Nitro |
| Multi-Attach | No | Up to 16 Nitro instances in one Availability Zone |
| Elastic Volumes | Supported | Supported (Multi-Attach restricts type changes) |
| Encryption | Supported | Supported |
One point of confusion: AWS moved all new and existing io2 volumes to Block Express on April 30, 2025. You no longer choose between standard io2 and Block Express.
Full performance requires Nitro-based instances. Other instance types can attach io2 provisioned up to 64,000 IOPS, but they deliver only up to 32,000 IOPS.
Cost, performance, and fit at a glance
- Cost: gp3 bills storage plus any IOPS and throughput above its baseline. io2 bills storage plus every provisioned IOPS, with tiered rates.
- Performance: Per-volume maximums are ceilings. The attached EC2 instance and your account quotas set what you can actually consume.
- Fit: gp3 suits boot volumes, application servers, and general-purpose databases. io2 suits mission-critical transactional databases and latency-sensitive workloads.

Specifications and prices vary by region and change over time. Verify every number against the AWS EBS documentation before you size a production volume.
What is gp3?
gp3 is AWS's general-purpose SSD volume. Unlike gp2, it decouples capacity from performance, so you can buy 100 GiB of storage with 16,000 IOPS without paying for 5 TiB of disk.
Performance model
- Every volume includes 3,000 IOPS and 125 MiB/s.
- Extra IOPS can be provisioned up to 500 IOPS per GiB. Reaching 80,000 IOPS requires at least 160 GiB.
- Throughput scales at 0.25 MiB/s per provisioned IOPS. The 2,000 MiB/s maximum needs at least 8,000 IOPS and 16 GiB.
Cost model
us-east-1 rates, per AWS's rate card:
| Charge | Rate |
|---|---|
| Storage | $0.08/GB-month |
| Additional IOPS (above 3,000) | $0.005/IOPS-month |
| Additional throughput (above 125 MiB/s) | $0.04/MiB/s-month |
Worked example: a 1,000 GB volume with 10,000 IOPS and 500 MiB/s.
- Storage: 1,000 × $0.08 = $80
- Extra IOPS: 7,000 × $0.005 = $35
- Extra throughput: 375 × $0.04 = $15
- Total: about $130/month

Confirm this total in the Pricing Calculator before you budget.
gp3 storage also costs 20% less per GiB than gp2's $0.10/GB-month, and gp2 ties baseline performance to size at 3 IOPS per GiB.
Limits and caveats
- Volume ceilings don't override EC2 limits. An r6i.32xlarge tops out at 160,000 IOPS and 5,000 MB/s across all attached volumes.
- Default account quota is 50 TiB of gp3 storage in us-east-1, and it's adjustable.
- Striping several gp3 volumes in RAID 0 adds throughput. It also means losing one volume loses the array, and backups need coordinated multi-volume snapshots.
Use Cases of gp3
gp3 fits most of what runs on EC2:
- Boot volumes and web or application servers
- Development and test environments
- Virtual desktops
- Medium single-instance databases whose measured needs fit under 80,000 IOPS
SAP's guidance is a useful benchmark. AWS's SAP HANA documentation lists gp3 alongside io2 Block Express for new deployments, with data volumes needing 7,000 IOPS and 400 MB/s and log volumes needing 3,000 IOPS and 250 MB/s.
Those targets sit inside gp3's range. Where you need consistent sub-millisecond latency, AWS points to io2 Block Express instead.
What is io2?
io2 is the provisioned-IOPS SSD volume for workloads that can't tolerate variance. You pay for every IOPS you provision, and you get higher durability, lower latency, and Multi-Attach in return.
Pricing and configuration
us-east-1 rates:
| Component | Rate |
|---|---|
| Storage | $0.125/GB-month |
| First 32,000 IOPS | $0.065/IOPS-month |
| IOPS 32,001-64,000 | $0.0455/IOPS-month |
| IOPS above 64,000 | $0.03185/IOPS-month |
Tiering applies per volume, not across your account. Ratios allow 1,000 IOPS per GiB, so 256,000 IOPS needs at least 256 GiB. Throughput scales at 0.256 MiB/s per IOPS, and the 4,000 MiB/s ceiling needs 16,000 IOPS.
Same 1,000 GB, 10,000 IOPS volume on io2: $125 storage plus $650 IOPS, roughly $775/month, versus about $130 on gp3. The gap is the price of the guarantees.

What justifies the premium
- 99.999% durability, versus 99.8%-99.9% for gp3
- Average latency under 500 microseconds on Nitro
- Multi-Attach for clustered applications
- IOPS beyond gp3's 80,000 ceiling
- Default account quota is lower (20 TiB and 100,000 aggregate IOPS), so plan quota requests early
A higher published ceiling doesn't guarantee your application can drive it. Instance limits and queue depth still apply.
Use Cases of io2
io2 is a strong fit for:
- Large transactional databases and SAP or other enterprise database platforms
- Latency-sensitive applications
- Clustered storage using Multi-Attach
- Workloads that exceed gp3's validated per-volume capability
Multi-Attach has strings attached. XFS and EXT4 aren't designed for simultaneous multi-server access. AWS calls for clustered file systems, application write coordination, and I/O fencing. Attaching the volume gives you none of that.
AWS's PostgreSQL benchmark (May 19, 2025) shows the benefit in a controlled setting:
- PostgreSQL 16.5 on r5b.24xlarge, HammerDB, roughly 1 TB dataset, 100 warehouses, 10 virtual users
- gp3: 8 volumes in RAID 0. io2: 2 volumes in RAID 0. Both provisioned 80,000 IOPS and 8,000 MiB/s total
- io2 delivered 2.08x transaction throughput, 3.85x lower OLTP latency, and 2.78x lower OLAP latency

The published results omit region, per-volume capacity, read/write mix, and run-to-run variance. Treat it as a synthetic test, not a production migration guide.
gp3 vs io2: What is better?
"Better" depends on measured requirements, not advertised maximums. A volume that can do 256,000 IOPS is wasted on a workload peaking at 6,000.
Decision checklist
- Peak and sustained IOPS and throughput
- Latency target (single-digit ms or sub-ms)
- Durability and replication requirements
- Multi-Attach need
- Volume size and quota headroom
- EC2 instance EBS limits
- Total monthly cost, including provisioned performance
Recommendations
- Choose gp3 when its limits and latency meet the workload and cost matters.
- Choose io2 when durability, sub-millisecond latency, Multi-Attach, or per-volume IOPS above 80,000 are genuine requirements.
Validate with CloudWatch
Pull these over peak periods:
- VolumeReadOps and VolumeWriteOps for real IOPS
- VolumeReadBytes and VolumeWriteBytes for throughput
- VolumeQueueLength: AWS's guideline is an average queue depth of about one per 1,000 provisioned IOPS
- VolumeAvgReadLatency on Nitro instances
- VolumeThroughputPercentage for provisioned-IOPS volumes (not available with Multi-Attach)
BurstBalance applies to gp2, st1, and sc1, not gp3 or io2. Add database-level metrics like query latency and wait events.
Migration and testing
Elastic Volumes can change type or performance without detaching, on supported instances.
- Snapshot first for rollback.
- Modify during a representative window. The volume moves through modifying, optimizing, then completed; performance sits between source and target while optimizing.
- Remember the limit of four modifications per rolling 24 hours. You can't shrink a volume, and a Multi-Attach io2 volume can't switch directly to gp3.
- Load test and confirm the instance limit isn't your new bottleneck.
Real world examples/case studies
AWS's Oracle test (October 1, 2021) is archival and controlled, not a production migration. Its prices are historical.
Setup:
- Region and instance: us-east-1a, r5b.24xlarge
- Database: Oracle 12.1.0.2 with ASM, Swingbench
- Workload: about 500 GB OLTP data, 150 sessions, 30 minutes
- gp3: 8 × 1 TB volumes at 10,000 IOPS each
- io2 Block Express: 2 × 2 TB at 40,000 IOPS each
The test does not document the read/write mix.
| Result | gp3 | io2 Block Express |
|---|---|---|
| Transactions/minute | 405,966 | 603,980 |
| Latency at 40,000 IOPS | 0.478 ms | 0.25 ms |
| Historical monthly total | $1,375.36 | $5,408 |
The challenge was database throughput and latency at high IOPS. io2 won on both, and gp3 won on cost by roughly 4x in the published model. The capacities were unequal, so treat the comparison as directional.
Takeaway: gp3 favored cost efficiency and moderate IOPS. io2 favored latency-sensitive, high-IOPS databases where transaction speed has business value.
Before making fleet-wide changes, find out which volumes are over-provisioned. Lucidity monitors every volume across AWS, Azure, and Google Cloud with real-time visibility into workload behavior. Its free Assessment scans an environment in about 5 minutes to surface disk utilization and underprovisioned disks.
Conclusion
gp3 is usually the right starting point when measured performance, latency, and durability needs fit its limits. io2 is justified when its higher guarantees or features directly support a business-critical workload.
Before you commit:
- Measure real workload behavior
- Check EC2 and account limits
- Calculate region-specific total cost
- Test under representative load
Revisit the choice as requirements change. Elastic Volumes makes adjusting less painful than it used to be.
Frequently Asked Questions
What are the key differences between AWS EBS io2 and gp3 volumes?
gp3 is general-purpose SSD with 3,000 baseline IOPS, up to 80,000 IOPS, and single-digit millisecond latency. io2 offers up to 256,000 IOPS, 99.999% durability, sub-millisecond latency, and Multi-Attach, at a higher price.
What is the IOPS limit for gp3 volumes?
Up to 80,000 IOPS and 2,000 MiB/s per volume, which requires at least 160 GiB. The attached EC2 instance's EBS limits can cap what you actually get.
Which Amazon EBS volume type can be expanded to 64,000 IOPS and 1,000 MB/s throughput and offers a tiered pricing structure?
That pair matches older standard io2 (AWS listed 1,000 MiB/s). Current io2 on Block Express reaches 256,000 IOPS and 4,000 MiB/s, with tiered IOPS pricing still in place.
What are the differences in pricing between AWS gp2 and gp3 volumes?
gp2 ties performance to size at 3 IOPS per GiB, while gp3 separates capacity from provisioned performance. In us-east-1, gp3 storage is $0.08/GB-month versus $0.10 for gp2.
What is gp3 volume in AWS?
gp3 is a general-purpose SSD EBS volume with independently configurable IOPS and throughput. It suits boot volumes, dev/test, and medium databases. Verify current limits in AWS documentation.
What are the different EBS volume types?
General-purpose SSD (gp2, gp3) for boot and dev/test, provisioned-IOPS SSD (io1, io2) for critical databases, throughput-optimized HDD (st1) for big data, and cold HDD (sc1) for infrequently accessed data. This article focuses on gp3 and io2.