Solidigm™ D7-PS1010 SSD Performance With 5th Gen AMD EPYC™ Processors
Introduction
The D7-PS1010 is the first Solidigm Gen5 SSD that is purpose built to accelerate modern workloads and power the demanding AI storage needs. This drive is available in E3.S and U.2 form factors with capacities of 1.92TB, 3.84TB, 7.68TB, and 15.36TB. The D7-PS1010 offers superb performance for today’s data center market while providing 1 drive-writes-per-day (DWPD) for a 5-year period.1 This paper highlights Solidigm D7-PS1010 performance when used in a server powered by dual 5th Gen 96-core AMD EPYC™ processors. 5th Gen AMD EPYC processors are built on the AMD “Zen 5” architecture and deliver impressive performance and energy efficiency for AI workloads. This performance brief includes both detailed system configurations and test results for both single drives and when scaling up to 24 drives.
System Configuration
The testing described in this performance brief used the following system configuration:
| Item | Description |
|---|---|
| CPU | 2 x 5th Gen 96-core AMD EPYC2 |
| RAM | 768 GB (24 x 32 GB DDR5-4800) |
| Storage | 1 x Solidigm D7-PS1010 PCIe® 5.0 x4 (U.2 15mm)
|
| BIOS | RVOT1000F |
| Non-default BIOS Settings | 1. Deterministic Control -> Enabled 2. Deterministic Enable -> Power |
| OS | Ubuntu 22.04, kernel 5.15 |
| Test Suite | FIO v3.28 |
Table 1. System configuration used for testing
Key Considerations:
- Solidigm expects up to 10% throughput variation between drive-to-drive runs.
- All drives must be purged and pre-conditioned before measuring performance.
- For 100% read workloads, the drive should be pre-conditioned with the equivalent 100% write workload.
- Any change in system and/or drive configuration may impact drive performance.
- Random performance is measured using queue depth 512 (QD64 x 8), except that Scaling 8k Random Read uses queue depth 256 (QD32 x 8).
- Sequential performance is measured with queue depth 128 (1 Worker).
- Runtime for all workloads is 60 seconds.
- Random workload performance is measured in throughput (IOPS).
- Sequential workload performance is measured in bandwidth (MBps).
Test Results
Single Drive
Tables 2 through 5 showcase single-drive Solidigm D7-PS1010 performance results on the 5th Gen AMD EPYC server described in Table 1 with a single Solidigm D7-PS1010 SSD on the AMD EPYC Volcano Platform. We were able to accomplish more than 100% of what was expected for all workloads across different capacities (each workload was executed using FIO for 60 seconds).
| Drive Capacity: 1.92TB | |||
|---|---|---|---|
| Workload | Spec | Actual | % of Spec |
| 4KB Random Read | 2350000 | 2409894 | 102.55 |
| 4KB Random Write | 150000 | 157885 | 105.26 |
| 8KB Random Read | 1200000 | 126011 | 105.01 |
| 8KB Random Write | 75000 | 76938 | 102.58 |
| 128KB Sequential Read | 14500 | 14506 | 100.04 |
| 128KB Sequential Write | 4100 | 4275 | 104.27 |
| 4KB Random Mixed Read Write 70/30 | 335000 | 372825 | 111.29 |
| 8KB Random Mixed Read Write 70/30 | 175000 | 182692 | 104.40 |
Table 2. 1.92TB single-drive results
| Drive Capacity: 3.84TB | |||
|---|---|---|---|
| Workload | Spec | Actual | % of Spec |
| 4KB Random Read | 3100000 | 3254340 | 104.98 |
| 4KB Random Write | 315000 | 317250 | 100.71 |
| 8KB Random Read | 1550000 | 1667217 | 107.56 |
| 8KB Random Write | 157000 | 179899 | 114.59 |
| 128KB Sequential Read | 14500 | 14514 | 100.10 |
| 128KB Sequential Write | 8200 | 8234 | 100.41 |
| 4KB Random Mixed Read Write 70/30 | 729000 | 820722 | 112.58 |
| 8KB Random Mixed Read Write 70/30 | 364000 | 468291 | 128.65 |
Table 3. 3.84TB single-drive results
| Drive Capacity: 7.68TB | |||
|---|---|---|---|
| Workload | Spec | Actual | % of Spec |
| 4KB Random Read | 2800000 | 3331032 | 118.97 |
| 4KB Random Write | 400000 | 430232 | 107.56 |
| 8KB Random Read | 1400000 | 1710534 | 122.18 |
| 8KB Random Write | 180000 | 204255 | 113.48 |
| 128KB Sequential Read | 14500 | 14817 | 102.19 |
| 128KB Sequential Write | 9300 | 9319 | 100.20 |
| 4KB Random Mixed Read Write 70/30 | 880000 | 896613 | 101.89 |
| 8KB Random Mixed Read Write 70/30 | 440000 | 509610 | 115.82 |
Table 4. 7.68TB single-drive results
| Drive Capacity: 15.36TB | |||
|---|---|---|---|
| Workload | Spec | Actual | % of Spec |
| 4KB Random Read | 2750000 | 2779914 | 101.09 |
| 4KB Random Write | 380000 | 433604 | 114.11 |
| 8KB Random Read | 1350000 | 1391166 | 103.05 |
| 8KB Random Write | 180000 | 192499 | 106.94 |
| 128KB Sequential Read | 14500 | 14710 | 101.45 |
| 128KB Sequential Write | 9300 | 9607 | 103.30 |
| 4KB Random Mixed Read Write 70/30 | 880000 | 927267 | 105.37 |
| 8KB Random Mixed Read Write 70/30 | 440000 | 475360 | 108.04 |
Table 5. 15.36TB single-drive
Multiple drives
The 5th Gen AMD EPYC server used for testing includes 24 NVMe drive bay slots. All slots were populated with 1.92TB Solidigm D7-PS1010 SSD and the FIO workloads executed to evaluate scaling performance. Each drive was mapped to specific CPU cores for optimal topology using the cpus_allowed FIO parameter. The results reveal that performance scales linearly across all workloads tested.
4KB Random Workloads
Figure 1. 4KB random read
The 4KB Random Read workload with QD512 (64x8) achieved ~57.4 million IOPS across 24 drives, which exceeds the expected linear scaling value of 56.4 million IOPS for this workload.
Figure 2. 4KB random write
The 4KB Random Write workload with QD512 (64x8) achieved ~3.6 million IOPS across 24 drives, which aligns with the expected linear scaling value of 3.6 million IOPS.
Figure 3. 4KB random mixed read/write 70/30
The 4KB Random Mixed Read Write 70/30 workload with QD512 (64x8) achieved ~8.2 million IOPS across 24 drives, which exceeds the expected linear scaling value of 8 million IOPS.
8KB Random Workloads
Figure 4. 8KB random read
The 8KB Random Read workload with QD256 (32x8) achieved ~27.3 million IOPS across 24 drives.
Figure 5. 8KB random write
The 8KB Random Write workload with QD512 (64x8) achieved ~1.8 million IOPS across 24 drives, which aligned with the expected linear scaling value of 1.8 million IOPS.
Figure 6. 8KB random mixed read/write 70/30
The 8KB Random Mixed Read Write 70/30 workload with QD512 (64x8) achieved ~4.3 million IOPS across 24 drives, which exceeds the expected linear scaling value of 4.2 million IOPS.
128KB Sequential Workloads
Figure 7. 128KB sequential read
The 128KB Sequential Read workload with QD128(128x1) achieved ~346 GBps across 24 drives was achieved, which aligns with the expected linear scaling value of 348 GBps.
Figure 8. 128KB sequential write
The 128KB Sequential Write workload with QD128(128x1) achieved ~102 GBps across 24 drives was achieved, which exceeds the expected linear scaling value of 98.4 GBps.
Conclusion
The test results featured in this performance brief showcase stellar single- and multi-drive Solidigm D7-PS1010 performance in a server powered by dual 5th Gen 96-core AMD EPYC processor. These results demonstrate real-world performance that exceeds the data sheet specifications for Solidigm D7-PS1010 SSD SKUs. Scaling tests using multiple Solidigm D7-PS1010 1.92TB SSDs achieved ~57 million IOPS (4k Random Read) and ~346GBps (128k Sequential Read), meaning that performance scaled linearly from 1-24 drives. Solidigm D7-PS1010 NVMe SSDs are an ideal choice when configuring servers powered by 5th Gen AMD EPYC processors.
Notes
- Refer to the Solidigm D7-PS1010 Product Brief for DWPD and warranty information.
- Testing used pre-production 5th Gen 96-core AMD EPYC processors (OPN 1536-03 96c). The production version of these processors will be the 5th Gen AMD EPYC 9655.
Tests document performance of components on a particular test in specific systems. Differences in hardware, software, or configuration will affect actual performance. Consult other sources of information to evaluate performance as you consider your purchase.
Performance results are based on testing as of dates shown in the configurations and may not reflect all publicly available updates. No product or component can be absolutely secure.
Solidigm optimizations, for Solidigm compilers or other products, may not provide optimized performance to the same degree for non-Solidigm products. Solidigm technologies may require enabled hardware, software, or service activation.
Your costs and results may vary.
Solidigm does not control or audit third-party data. You should consult other sources to evaluate accuracy.
© Solidigm 2024
“Solidigm” is a trademark of SK hynix NAND Product Solutions Corp (d/b/a Solidigm). AMD, the AMD arrow logo, EPYC, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other names and brands may be claimed as the property of others.