For the purpose of this review, Phison supplied the 7.68TB U.2 variant of the Pascari X200P for comprehensive testing. To thoroughly assess its performance under real-world enterprise workloads and pressure, we subjected the SSD to our full suite of rigorous enterprise benchmarks. These tests evaluated key performance metrics, including throughput, latency, and stability, across a diverse range of workload profiles to provide a complete picture of its capabilities in enterprise environments.
For this review, we focus on the Phison-supplied 7.68TB U.2 variant of the Pascari X200P for comprehensive testing. To thoroughly evaluate its performance under real-world enterprise workloads and operational pressures, we subjected the SSD to our full suite of rigorous enterprise-grade benchmarks. These comprehensive assessments measured key performance metrics—encompassing throughput, latency, and stability—across a wide spectrum of diverse workload profiles to fully characterize its capabilities in enterprise environments.
For this review, Phison provided the 7.68TB U.2 variant of the Pascari X200P for comprehensive testing, and we subjected the SSD to our full suite of rigorous enterprise benchmarks to thoroughly assess its performance under real-world enterprise workloads and pressure, evaluating key metrics like throughput, latency, and stability across diverse workload profiles to fully illustrate its capabilities in enterprise environments.
Build and Design
The Solidigm D5-P5336 122.88TB shares the same core architecture as the previously reviewed 61.44TB model, utilizing 192-layer QLC NAND. This consistency guarantees predictable performance, thermal behavior, and interface compatibility across capacities—critical for scale-out deployments. As a 32KB I/O Unit drive (an upgrade from the 16KB in the 61TB variant), the 122TB D5-P5336 is optimized for mid-sized I/O patterns typical in object storage and AI data pipelines, delivering greater workload flexibility while preserving efficiency.
What distinguishes this model is its 122.88TB capacity, which doubles storage volume without requiring additional physical space. Encased in a standard 2.5-inch U.2 15mm form factor, it is also available in E3.S 7.5mm and E1.L 9.5mm configurations to cater to diverse hyperscale requirements. The drive employs a PCIe Gen4 x4 NVMe interface, delivering up to 7GB/s sequential read throughput and 3GB/s write throughput. While it does not adopt PCIe Gen5, Gen4 provides ample bandwidth for the read-intensive workloads the D5-P5336 is targeting, such as AI pipelines, content distribution, and object storage.
From a performance perspective, the drive achieves up to 900,000 IOPS for random reads (4K, QD256) and 19,000 IOPS for random writes (16K, QD256). Read latency is specified at 110 microseconds (4K) and write latency at 40 microseconds (32K). Sequential access latency is even lower, with reads at 8 microseconds (4K) and writes at 21 microseconds (32K), enabling highly responsive operation in large-scale deployments.
When comparing the 122TB P5336 to the previous 61TB model, the higher-capacity SSD features slightly lower listed write performance. Sequential 128K transfers decrease to 3GB/s from 3.3GB/s, and 16K random write performance drops more noticeably, from 43K IOPS to just 19K IOPS. As we proceed with the evaluation, it is important to note that performance will differ as specific workloads stress the drives in terms of sequential or random transfer capabilities.
The drive incorporates SK hynix DRAM cache and power-loss protection capacitors. These components ensure reliable data buffering and safeguard data during unexpected power outages— a requirement in enterprise-scale environments. The drive’s reliability specifications include a mean time between failures (MTBF) rating of two million hours and an unrecoverable bit error rate of less than one bit error per 100 quadrillion bits read.
Organizations are concerned about the overall lifespan of SSDs, particularly the number of writes they can handle over years of use. The endurance rating for the Solidigm D5-P5336 is 0.6 drive writes per day (DWPD), based on a 32K random write workload, which translates to 134.3 petabytes written (PBW) over the warranty period. The 122TB D5-P5336 SSD from Solidigm sets a new benchmark for endurance, engineered for 24/7 continuous operation over a five-year period. It can handle either 32KB random writes, retaining 5% of its endurance after five years, or 4K random writes, with 12% endurance remaining. While it maintains a 0.60 DWPD rating, the increased NAND capacity allows it to support continuous workloads more effectively.
The drive features passive cooling and is housed in a sturdy aluminum case. It operates with a modest power profile—24 watts active and 5 watts idle—enabling easy integration into existing infrastructures. Weighing approximately 166.4 grams, it supports an operating temperature range of 0 to 70 degrees Celsius, vibration resistance up to 2.17 GRMS, and shock resistance up to 1,000 G, all backed by a five-year warranty. It is specifically designed for environments prioritizing density, efficiency, and rack consolidation, delivering massive storage capacity within a familiar enterprise form factor.
Solidigm D5-P5336 Series (122.88TB) Specifications
| Specifications Overview |
Solidigm D5-P5336 Series (122.88TB) |
| Capacity |
122.88TB |
| Form Factor |
U.2 15mm or E1.L 9.5mm |
| Interface |
PCIe 4.0 x4, NVMe |
| Use Case |
Server / Enterprise |
| Sequential Read |
7000MB/s |
| Sequential Write |
3000MB/s |
| Random Read (IOPS) |
900,000 (4K, QD256) |
| Random Write (IOPS) |
19,000 (16K, QD256) |
| Latency (Read/Write) |
Read: 110μs (4K) / Write: 40μs (32K) |
| Sequential Latency (typ.) |
Read: 8μs (4K) / Write: 21μs (32K) |
| Power (Active/Idle) |
Active: 24W / Idle: 5W |
| Endurance |
0.6 DWPD (32K RW) / 134.3 PBW |
| MTBF |
2 million hours |
| UBER |
<1 sector per 10 bits read |
| Operating Temp |
0°C to 70°C |
| Vibration/Shock |
2.17 GRMS (operating), 1,000 G (shock) |
| Warranty |
5 years |
| Weight |
166.4g ± 10g |
Performance Testing
Drive Testing Platform
We leverage a Dell PowerEdge R760 running Ubuntu 22.04.02 LTS as our test platform across all the workloads in this review. Equipped with a Serial Cables Gen5 JBOF, it offers wide compatibility with U.2, E1.S, E3.S and M.2 SSDs Our system configuration is outlined below:
- 2 x Intel Xeon Gold 6430 (32-Core, 2.1GHz)
- 16 x 64GB DDR5-4400
- 480GB Dell BOSS SSD
- Serial Cables Gen5 JBOF
Drives Compared
As noted in the introduction, the high-capacity enterprise drive market is complex, with various form factors, NAND types, and cost-performance profiles to consider. For this review, we have a small set of SSDs to compare against the 122.88TB Solidigm P5336, including the smaller 61.44TB Solidigm P5336 and the 61.44TB Micron 6550.
The Micron 6550 stands out as a Gen5, TLC-based drive—and one of the few currently in production at this capacity point. This gives the Micron drive an advantage in terms of higher I/O speeds.
As we analyze the performance results, it is crucial to understand this context. In real-world deployments, these drives may not compete directly, but they do overlap in the storage capacities they offer. To provide a reference for scale, we have included the Micron drive in this review.
CDN Performance
To simulate a realistic, mixed-content CDN workload, the SSDs were subjected to a multi-phase benchmarking sequence designed to replicate the I/O patterns of content-heavy edge servers. The testing procedure includes a range of block sizes—both large and small—distributed across random and sequential operations, with varying concurrency levels.
Prior to the main performance tests, each SSD underwent a full device fill using a 100% sequential write pass with 1MB blocks. This process employed synchronous I/O and a queue depth of four, enabling four simultaneous jobs. This phase ensures the drive enters a steady-state condition that mirrors real-world usage. Following the sequential fill, a secondary three-hour randomized write saturation stage was conducted using a weighted bssplit (blocksize/percentage) distribution, with a strong emphasis on 128K transfers (98.51%) and minor contributions from sub-128K blocks down to 8K. This step mimics the fragmented, uneven write patterns commonly observed in distributed cache environments.
The main testing suite focused on scaled random read and write operations to measure the drive’s performance under variable queue depths and job concurrency. Each test ran for five minutes (300 seconds), followed by a three-minute idle period to allow internal recovery mechanisms to stabilize performance metrics.
-
This was executed using a fixed block size distribution favoring 128K (98.51%), with the remaining 1.49% of operations consisting of smaller transfer sizes ranging from 64K to 8K. Each configuration varied across 1, 2, and 4 concurrent jobs, with queue depths of 1, 2, 4, 8, 16, and 32, to profile throughput scalability and latency under typical edge-write conditions.
A heavily mixed block size profile, simulating CDN content retrieval, was utilized, starting with a dominant 128K (83.21%) component followed by a long tail of over 30 smaller block sizes—ranging from 4K to 124K—each with fractional frequency representation. This distribution reflects the diverse request patterns encountered during video segment fetching, thumbnail access, and metadata lookups, and these tests were also run across the complete matrix of job counts and queue depths.
This combination of preconditioning, saturation, and mixed-size randomized access tests is designed to reveal how SSDs handle sustained CDN-like environments, emphasizing responsiveness and efficiency in bandwidth-heavy and highly parallelized scenarios.
CDN Workload Read 1
In this single-threaded read test simulating light content delivery traffic, the Solidigm P5336 122.88TB and Solidigm P5336 61.44TB exhibit consistent scaling characteristics. The 122.88TB model reaches 7,109MB/s at QD32, slightly ahead of the 61.44 TB’s 7,002MB/s. This near-identical scaling suggests that Solidigm’s higher-capacity model retains the same efficiency under light read pressure without performance degradation. In contrast, the Micron 6550 61.44TB scales much more aggressively, topping out at 12,288MB/s.
CDN Workload Read 2
With two threads applied, the Solidigm P5336 122.88TB and P5336 61.44TB deliver nearly identical performance, scaling from 840MB/s at QD1 to approximately 7,467MB/s and 7,469MB/s respectively at QD32. Both drives demonstrate consistent gains up to QD16, after which throughput levels off, indicating a saturation point in their current architecture. For applications with moderate parallelism, this provides a reliable baseline for predictable scaling. The Micron 6550, in contrast, shows a higher overall scaling range, starting at 1,384MB/s and continuing upward to 13,312MB/s at QD32, reflecting the benefits of its TLC NAND and Gen5 interface.
CDN Workload Read 4
This high-demand read scenario places greater stress on the drives with increased concurrency. The Solidigm P5336 122.88TB and P5336 61.44TB drives exhibit consistent scaling, reaching approximately 7,466–7,469MB/s at QD16 and maintaining stability through QD32. The results between the two capacities remain effectively identical, reinforcing Solidigm’s consistent controller behavior across its high-capacity lineup. In comparison, the Micron 6550 achieved 13,107MB/s by QD16 and sustained that bandwidth through the remainder of the test.
CDN Workload Write 1
Transitioning into write performance under single-threaded conditions, the Solidigm P5336 122.88TB begins at 1,742MB/s and reaches approximately 2,572MB/s at QD32. The Solidigm P5336 61.44TB starts lower at 461MB/s but scales more aggressively, peaking at 3,029MB/s. The Micron 6550 starts at 984MB/s and continues scaling consistently through the queue depth range, reaching 6,288MB/s by QD32. The Solidigm models exhibit differing scaling characteristics, whereas Micron maintains a more linear progression across the test.
CDN Workload Write 2
Moving into dual-threaded write performance, bandwidth increases for all three drives. The Solidigm P5336 61.44TB starts at 2,771MB/s and maintains relatively steady output through QD32, with only minor fluctuations. The Solidigm P5336 122.88TB operates within a narrower range, staying between 2,468MB/s and 2,620MB/s across all queue depths. The Micron 6550 exhibits continued scaling, beginning at 2,035MB/s and reaching 6,743MB/s by QD32. The Solidigm drives maintain consistent throughput, while the Micron shows a broader scaling profile over the same range.
CDN Workload Write 4
Under maximum concurrency, both Solidigm P5336 models show stable but limited scaling. The P5336 61.44TB starts at approximately 2,935MB/s and peaks at 3,062MB/s, while the P5336 122.88TB begins at 2,529MB/s and ends slightly lower at 2,562MB/s. This results in a roughly 16% lower peak throughput for the 122.88TB model compared to the 61.44TB version. The Micron 6550, on the other hand, scales steadily from 2,323MB/s to 6,731MB/s by QD32.
ObjectStorage Performance
This test leverages an FIO script approximating an ObjectStorage workload, with 65% of requests issued at a 64 KiB transfer size to represent common small-block operations, 15% at 8 MiB for mid-range streaming workloads, and another 15% at 64 MiB to stress the drive’s large-block handling. The final 5% at 1 GiB payload pushes the maximum sequential throughput. By interleaving these four block sizes in the specified proportions, it simulates a mixed workload that reveals both the controller’s agility under small I/O and its raw bandwidth capabilities under massive transfers.
Random Read (1 Thread, 40QD)
| Drive |
Read Bandwidth (MB/s) |
Read IOPS |
Read Latency (ms) |
| Micron 6550 61TB |
13,444.10 |
3,165.10 |
12.5011 |
| Solidigm P5336 61TB |
7,117.38 |
1,673.76 |
23.4513 |
| Solidigm P5336 122TB |
7,101.97 |
1,674.78 |
23.4385 |
In this single-threaded, high-depth random read test, the Solidigm P5336 122.88TB and P5336 61.44TB deliver nearly identical performance. The 122.88TB model achieves 7,101.97MB/s and 1,674.78 IOPS with a latency of 23.44ms, while the 61.44TB variant hits 7,117.38MB/s and 1,673.76 IOPS at 23.45ms. The bandwidth difference between the two Solidigm capacities is less than 0.25%, underscoring consistent performance across the P5336 lineup for random read workloads.
The Micron 6550 offers significantly higher performance, reaching 13,444.10 MB/s and 3,165.10 IOPS with a lower latency of 12.50ms. Its edge in this scenario stems from its use of TLC NAND and a PCIe Gen5 interface—both of which contribute to stronger random read throughput and responsiveness compared to the QLC-based, Gen4 Solidigm drives.
Sequential Read (1 Thread, 40QD)
| Drive |
Read Bandwidth (MB/s) |
Read IOPS |
Read Latency (ms) |
| Micron 6550 61TB |
13,955.46 |
223.32 |
174.723 |
| Solidigm P5336 61TB |
7,098.64 |
114.12 |
341.727 |
| Solidigm P5336 122TB |
7,103.98 |
114.60 |
340.322 |
Moving on to sequential read performance, the Solidigm P5336 122.88TB and P5336 61.44TB deliver nearly identical results. The 122.88TB model hits 7,103.98MB/s with 114.60 IOPS and a latency of 340.32ms, while the 61.44TB version registers 7,098.64MB/s, 114.12 IOPS, and 341.73ms. The performance difference between the two is less than 0.1%, reflecting consistent behavior across both capacities in sustained sequential read workloads. The Micron 6550 performs significantly better, measuring 13,955.46MB/s and 223.32 IOPS with 174.72ms latency—offering roughly 96% higher throughput than either Solidigm model in this test.
Random Read (4 Thread, 10QD)
| Drive |
Read Bandwidth (MB/s) |
Read IOPS |
Read Latency (ms) |
| Micron 6550 61TB |
13,301.67 |
3,142.01 |
12.5619 |
| Solidigm P5336 61TB |
7,131.65 |
1,686.98 |
22.9787 |
| Solidigm P5336 122TB |
7,131.95 |
1,690.84 |
22.9315 |
Moving to a four-thread read test with a queue depth of 10, the Solidigm P5336 122.88TB registers 7,131.95MB/s, 1,690.84 IOPS, and 22.93ms latency. The Solidigm P5336 61.44TB trails slightly at 7,131.65MB/s and 1,686.98 IOPS, with a latency of 22.98ms. The bandwidth difference between the two models is less than 0.005%. Meanwhile, the Micron 6550 hits 13,301.67MB/s and 3,142.01 IOPS with 12.56ms latency, providing approximately 86% more throughput than either Solidigm drive.
Sequential Read (4 Thread, 10QD)
| Drive |
Read Bandwidth (MB/s) |
Read IOPS |
Read Latency (ms) |
| Micron 6550 61TB |
13,524.00 |
218.06 |
171.040 |
| Solidigm P5336 61TB |
7,130.97 |
115.03 |
315.565 |
| Solidigm P5336 122TB |
7,130.99 |
114.72 |
316.304 |
In this four-thread sequential read test at queue depth 10, the Solidigm P5336 122.88TB achieves 7,130.99MB/s, 114.72 IOPS, and a latency of 316.30ms. The Solidigm P5336 61.44TB delivers nearly matching performance, registering 7,130.97MB/s, 115.03 IOPS, and 315.57ms latency. Across capacities, the two models exhibit almost identical sequential performance, with a difference of less than 0.01%. Under the same test conditions, the Micron 6550 outputs 13,524.00MB/s and 218.06 IOPS with 171.04ms latency, offering approximately 89% higher throughput than either Solidigm drive.
DLIO Checkpointing Benchmark
To evaluate the real-world performance of SSDs in AI training environments, we employed the Data and Learning Input/Output (DLIO) benchmark tool. Developed by Argonne National Laboratory, DLIO is specifically engineered to test I/O patterns in deep learning workloads, offering insights into how storage systems address key challenges such as checkpointing, data ingestion, and model training. The chart below demonstrates how both drives handle this process across 99 checkpoints (198 for the 122TB model). During machine learning model training, checkpoints are critical for periodically saving the model’s state, preventing progress loss in the event of interruptions or power failures. This storage demand calls for robust performance, particularly under sustained or intensive workloads. We used DLIO benchmark version 2.0, released on August 13, 2024.
To ensure our benchmarking aligned with real-world scenarios, we based our testing on the LLAMA 3.1 405B model architecture. We implemented checkpointing via torch.save() to capture model parameters, optimizer states, and layer states. Our setup simulated an eight-GPU system, adopting a hybrid parallelism strategy with 4-way tensor parallelism and 2-way pipeline parallel processing distributed across the eight GPUs. This configuration resulted in checkpoint sizes of 1,636GB, which is representative of the storage requirements for modern large language model training.
When comparing the checkpoint performance of the 61TB and 122TB Solidigm P5336, the 122TB SSD experiences longer checkpoint times once the drive is fully filled. In the first pass, the 122TB version is approximately 20% faster than the 61TB model; however, in the second and third passes, it is 16.4% and 18.4% slower, respectively. The 61TB Micron 6550 achieves an average checkpoint time of 585 seconds in the third pass, compared to 640 seconds for the 61TB P5336 and 757 seconds for the 122TB P5336.
The 122TB Solidigm P5336 boasts a unique initial advantage when it comes to checkpoint storage: it can accommodate significantly more checkpoints. While the 61TB SSDs max out at 33 checkpoints per pass, the 122TB model can fit 66 checkpoints before reaching its capacity limit. Although the average time-per-pass chart above somewhat masks these quantities, the per-checkpoint time perspective helps highlight this capacity advantage. Both Solidigm SSDs stabilize their performance after completing the first pass of checkpoints, while the Micron 6550 maintains relative consistency throughout the test, with a trend toward faster checkpoint times.
FIO Performance Benchmark
To measure the storage performance of each SSD across common industry metrics, we leverage FIO. Each drive is run through the same testing process, which includes a preconditioning step of two full drive fills with a sequential write workload, followed by measurement of steady-state performance. As each workload type being measured changes, we run another preconditioning fill of that new transfer size.
In this section, we focus on the following FIO benchmarks:
- 128K Sequential
- 64K Random
- 16K Random
- 4K Random
With the high-capacity QLC SSDs designed for large transfer sizes, our write speed tests stop at 16K random. For 4K, we leverage the pre-filled state from the 16K workload to measure only 4K random read performance.
128K Sequential Precondition (IODepth 256 / NumJobs 1)
In this heavy queue-depth preconditioning test, the Solidigm P5336 122.88TB achieves 3,134 MB/s, while the P5336 61.44TB reaches 2,500.9 MB/s—representing a 25.3% improvement in write bandwidth for the higher-capacity model. The Micron 6550 leads the pack at 10,455.3 MB/s. While both Solidigm models lag behind the Micron in raw throughput, the performance gap between the 122TB and 61TB variants underscores scale-driven optimization within the P5336 platform, with the larger drive delivering clear gains in handling sustained sequential writes. Though the Micron 6550 features a notably smaller preconditioning stage, its higher write speed allowed it to complete the initial fill far more quickly.
128K Sequential Precondition Latency (IODepth 256 / NumJobs 1)
In terms of latency during the preconditioned 128K sequential write, the Micron 6550 registers the lowest value at 3.06ms. The Solidigm P5336 122.88TB follows with 10.21ms, while the P5336 61.44TB comes in at 12.80ms. This translates to a 20.2% latency reduction for the 122.88TB model compared to its 61.44TB counterpart, reflecting more efficient and consistent latency performance and highlighting the enhancements made within the Solidigm P5336 series.
128K Sequential Write (IODepth 16 / NumJobs 1)
In this sequential write test with a queue depth of 16 and a single job, the Solidigm P5336 122.88TB delivers 3,152.5MB/s and 25,220 IOPS. The P5336 61.44TB follows closely behind, recording 2,503.5MB/s and 20,030 IOPS—yielding a 25.9% throughput advantage for the higher-capacity Solidigm model. Leading the rankings, the Micron 6550 achieves top-tier performance at 10,456.4MB/s and 83,650 IOPS, outperforming both Solidigm drives in this scenario.
Russian
