Russian
Key Takeaways
- Flash-first NVMe RAID: PERC13 H975i moves off SAS/SATA entirely, built on Broadcom SAS51xx for an NVMe-native, AI-ready architecture.
- Big generational jump: PCIe Gen5 x16 with up to 16 NVMe drives per controller (32 with two) delivered 52.5 GB/s and 12.5M IOPS per controller in testing, with gains vs PERC12 including +88% read bandwidth, +318% write bandwidth, +31% 4K read IOPS, and +466% 4K write IOPS.
- AI server fit: Front-integrated design frees rear PCIe slots for GPUs, shortens MCIO runs, and enables a dedicated storage pipe per accelerator for steadier, more deterministic throughput with no CPU overhead.
- Resiliency under stress: Supercapacitor-protected cache and faster rebuilds cut time to as low as 10 min/TiB while maintaining high performance during rebuilds (up to 53.7 GB/s reads, 68 GB/s writes, 17.3M/5.33M 4K IOPS).
- End-to-end security: Hardware Root of Trust, SPDM device identity, and full-spectrum encryption that covers drives, in-flight data, and controller cache.
Dell PERC12 H965i and PERC13 H975i Specifications
| Feature | PERC12 H965i Front | PERC13 H975i Front |
|---|---|---|
| RAID Levels | 0, 1, 5, 6, 10, 50, 60 | 0, 1, 5, 6, 10, 50, 60 |
| Non-RAID (JBOD) | Yes | Yes |
| Host Bus Type | PCIe Gen4 x16 | PCIe Gen5 x16 |
| Side-band Management | I2C, PCIe VDM | I2C, PCIe VDM |
| Enclosures per Port | Not applicable | Not applicable |
| Processor / Chipset | Broadcom RAID-on-Chip, SAS4116W | Broadcom RAID-on-Chip, SAS5132W |
| Energy Pack / Power-backup | Battery | Supercapacitor |
| Local Key Management Security | Yes | Yes |
| Secure Enterprise Key Manager | Yes | Yes |
| Controller Queue Depth | 8,192 | 8,192 |
| Non-volatile Cache | Yes | Yes |
| Cache Memory | 8 GB DDR4 3200 MT/s | Integrated RAID cache |
| Cache Functions | Write-back, read-ahead, write-through, always write-back, no read-ahead | Write-back, write-through, always write-back, no read-ahead |
| Max Complex Virtual Disks | 64 | 16 |
| Max Simple Virtual Disks | 240 | 64 |
| Max Disk Groups | 64 | 32 |
| Max VDs per Disk Group | 16 | 8 |
| Max Hot-spare Devices | 64 | 8 |
| Hot-swap Devices Supported | Yes | Yes |
| Auto-Configure (Primary & Execute once) | Yes | Yes |
| Hardware XOR Engine | Yes | Yes |
| Online Capacity Expansion | Yes | Yes |
| Dedicated & Global Hot Spare | Yes | Yes |
| Supported Drive Types | NVMe Gen3 and Gen4 | NVMe Gen3, Gen4 and Gen5 |
| VD Strip Element Size | 64KB | 64KB |
| NVMe PCIe Support | Gen4 | Gen5 |
| Configuration Max NVMe Drives | 8 drives per controller | 16 drives per controller |
| Supported Sector Sizes | 512B, 512e, 4Kn | 512B, 512e, 4Kn |
| Storage Boot Support | UEFI-only | UEFI-only |
The PERC13 H975i Front controller in Dell PowerEdge servers is designed for seamless integration into the system architecture. Unlike traditional add-in cards that occupy rear PCIe slots, the H975i connects directly to the front drive backplane and interfaces with the front MCIO connectors on the motherboard through dedicated PCIe 5.0 interfaces. This integrated design preserves rear PCIe slots for high-performance GPUs and additional PCIe expansion, while significantly reducing the length of cables. This aids in maintaining signal integrity, making the system more reliable and easier to service. The result is a cleaner internal layout and improved airflow for dense, compute-intensive deployments.
The H975i implements a comprehensive security architecture that spans from silicon-level hardware attestation through full-spectrum data encryption of data in place with SED drives. At its foundation, Hardware Root of Trust establishes an immutable chain of cryptographic verification from the Internal Boot ROM through each firmware component, ensuring only authenticated Dell-certified firmware can execute on the controller. This hardware-based security extends through Security Protocol and Data Model (SPDM) implementation, where each controller contains a unique Device Identity certificate enabling iDRAC to perform real-time authentication verification. The controller extends cryptographic protection beyond traditional data-at-rest scenarios to include cache memory. It maintains encryption keys in secure memory regions that are inaccessible to unauthorized firmware. As a result, sensitive data remains protected whether residing on drives or actively being processed in cache.
Power protection in the H975i is another significant evolution from traditional battery-backed systems through its integration of a supercapacitor. The supercapacitor provides instantaneous power delivery during unexpected power loss events, ensuring an encrypted and complete cache flush to non-volatile storage, where data remains protected indefinitely. In addition, unlike battery-based systems that require 4-8 hours for learn cycles, the H975i’s supercapacitor completes its Transparent Learn Cycle within 5-10 minutes without any performance degradation during calibration. This design eliminates the maintenance overhead and degradation concerns inherent in battery solutions, while providing superior reliability for mission-critical data protection.
Integrated Monitoring and Management
Dell’s PERC13 RAID controller, like many of Dell’s RAID solutions, can be managed and monitored in many ways, including during platform boot via System Setup in the BIOS, through the iDRAC web GUI, the PERC12 utility, and even Dell OpenManage UI and CLI.
iDRAC Controller Management
When viewing the iDRAC management interface, the controllers tab offers an overview of the server’s storage hardware. Alongside the BOSS card, you’ll see the dual PERC H975i controllers, complete with information on firmware versions, cache memory, and battery health. This summary enables you to quickly verify the controllers’ readiness and configuration without needing to access the BIOS or use CLI tools.
The Virtual Disks tab in iDRAC shows the storage arrays that have been created, including their RAID level, size, and caching policy. In this system, two RAID-10 groups are listed, all built on SSDs. From this view, administrators can confirm volumes are online, create new virtual disks, or use the Actions menu to adjust or delete existing configurations.
RAID Controller Configuration Utility
The image above shows an example of entering the PERC H975i Front Configuration Utility System Setup on the PowerEdge R7715 platform. From this interface, you can manage all key RAID controller settings, including Configuration Management, Controller Management, Device Management, and more. This utility provides a streamlined way to set up virtual disks and monitor hardware components directly during the platform boot process.
After selecting the RAID level, we move on to choosing physical disks for the array. In this example, all available NVMe SSDs are listed and marked as RAID-capable. We select multiple 3.2 TiB Dell DC NVMe drives from the unconfigured capacity pool. Filters like media type, interface, and logical sector size help narrow the selection. Once the desired drives are checked, we can proceed by clicking “OK” to finalize the disk selection and continue creating the Virtual Disk.
Before finalizing the virtual disk creation, the system displays a warning that confirms all data on the selected physical disks will be permanently deleted. To proceed, we check the “Confirm” box and select “Yes” to authorize the operation. This safeguard helps prevent accidental data loss during the RAID creation process.
Once the virtual disk is created, it appears under the “Virtual Disk Management” menu. In this example, our new RAID 5 virtual disk is listed with a capacity of 43.656 TiB and a status of “Ready.” With just a few simple steps, the storage is configured and ready for use.
While the PERC BIOS Configuration Utility and iDRAC interface offer intuitive options for local and remote management, Dell also provides a powerful command-line tool called PERC CLI (perccli2). This utility supports Windows, Linux, and VMware, making it ideal for scripting, automation, or managing PERC controllers in headless environments. Dell also provides detailed documentation on installation and command usage for PERC CLI on their support site.
Dell PERC13 Performance Testing
Before diving into performance testing, we prepared our environment using the Dell PowerEdge R7715 platform configured with dual PERC H975i front controllers. These were paired with thirty-two 3.2TB Dell NVMe drives, each rated for up to 12,000 MB/s sequential reads and 5,500 MB/s sequential writes using 128 KiB block sizes. This high-performance foundation enables us to push the limits of the PERC13 controller’s throughput and evaluate RAID behavior at scale.
- Platform: Dell PowerEdge R7715
- CPU: AMD EPYC 9655P 96-Core Processor
- Ram: 768GB (12 x 64GB) DDR5-5200 ECC
- Raid Controller: 2 x PERC13 H975i
- Storage: 32 x 3.2TB Dell CD8P NVMe Drives
- PCIe Accelerators: 2 x NVIDIA H100 GPU
NVIDIA Magnum IO GPU Direct Storage: AI Meets Storage
Modern AI pipelines are often I/O-bound, not compute-bound. Data batches, embeddings, and checkpoints must be transferred from storage to GPU memory quickly enough to keep accelerators busy. NVIDIA’s Magnum IO GDS (via cuFile) short-circuits the traditional “SSD → CPU DRAM → GPU” path and lets data DMA directly from NVMe to GPU memory. That removes CPU bounce-buffer overhead, lowers latency, and makes throughput more predictable under load, all of which translates to higher GPU utilization, shorter epoch times, and quicker checkpoint save/load cycles.
Our GDSIO test is geared to measure the storage-to-GPU data path itself, sweeping block sizes and thread counts to show how quickly PERC13-backed NVMe set can stream into H100 memory. With each H975i on a PCIe 5.0 x16 link (theoretical ~64 GB/s per controller, unidirectional), two controllers set an aggregate ceiling near ~112 GB/s; where our curves plateau tells you whether you’re link or media-limited. For practitioners, read the charts as proxies for real workloads: large sequential reads map to dataset streaming and checkpoint restores; large sequential writes map to checkpoint saves; smaller transfers with concurrency reflect dataloader shuffles and prefetch. In short, strong GDSIO scaling means fewer GPU stalls and more consistent performance during both training and high-throughput inference.
GDSIO Read Sequential Throughput
Starting with sequential read, throughput began modestly at lower block sizes and thread counts, starting around 0.3 GiB/s at 8K blocks with a single thread. Performance scaled sharply between 16K and 512K blocks, particularly when increasing thread count from 4 to 16. The most substantial gains occurred at 1M, 5M, and 10M block sizes, where throughput jumped dramatically, peaking at 103 GiB/s at 10M block size with 256 threads. This progression shows that the PERC13 array benefits from larger block sizes and multithreaded parallelism, with optimal saturation around 64-128 threads, beyond which gains plateau.
GDSIO Read Sequential Throughput Differential
In sequential read testing across block sizes from 8K to 10M, the PERC13 (H975i) consistently outperformed the PERC12 (H965i), with percentage gains scaling dramatically at larger block sizes and higher thread counts.
At smaller block sizes (8K-16K), improvements were modest (typically ranging from 0-20%), and in some isolated cases the H975i trailed slightly due to test variability at low queue depths. By 32K-64K block sizes, the advantage became more consistent, with the H975i delivering 30-50% higher throughput across most thread counts.
The most significant differences were observed at larger block sizes (128K through 10M), where the PERC13 controller unlocked the full sequential read potential of the system. Here, the H975i demonstrated gains of 50-120% compared to the H965i. For example, at 1M block size with 8-16 threads, throughput was over 55 GiB/s higher, equating to roughly a 90% uplift. At block sizes of 5M and 10M, improvements regularly exceeded 100%, with some configurations showing nearly double the performance compared to the previous generation.
Overall, the PERC13 (H975i) established a commanding lead in sequential read workloads, especially as block size and thread count scaled. While smaller block sizes showed incremental improvement, at 256K and above, the newer controller consistently delivered 50-100%+ higher performance, clearly highlighting the architectural advancements in Dell’s latest RAID platform.
GDSIO Read Sequential Latency
As sequential read throughput increased, latency remained manageable at smaller block sizes and lower thread counts. For example, latency stayed below 100 µs up to 64K blocks and 16 threads, showing efficient handling of reads in that range. Once block sizes and thread counts scaled higher, especially at 5M and 10M with 64 or more threads, latency climbed rapidly, peaking at 211.8 ms at a 10M block size with 256 threads. This highlights how controller or queuing bottlenecks emerge under extreme workloads, even though throughput remains high.
The best balance of performance and efficiency was observed at the 1M block size with 8-16 threads, where the array sustained 87.5-93.7 GiB/s throughput while keeping latency between 179-334 µs. This zone represents the sweet spot for maximizing bandwidth while keeping delays well under a millisecond.
GDSIO Write Sequential Throughput
Write performance showed strong early scaling as block sizes increased, with throughput climbing from 1.2 GiB/s at 8K and 1 thread to 13.9 GiB/s by 256K. The most substantial growth appeared between 128K and 1M block sizes, where throughput reached over 80 GiB/s at 8 to 16 threads. Peak performance came at the 5M and 10M block sizes, sustaining 100 to 101 GiB/s from 8 threads onward.
Performance flattened across 8 to 64 threads for these larger blocks, indicating the controllers reached saturation early in the scaling curve. At higher thread counts, especially 128 and 256 threads, throughput stability varied, holding steady at large 5M and 10M blocks at 101 GiB/s but declining for mid-range block sizes, such as 256K, falling from 61.2 GiB/s at 32 threads to 45.3 GiB/s at 256 threads.
GDSIO Write Sequential Throughput Differential
In sequential write testing, the PERC13 (H975i) delivered substantial gains over the PERC12 (H965i), particularly as block sizes and thread counts scaled. At small block sizes (8K-32K), improvements were modest, generally within 0-10%, with occasional test noise showing negligible differences.
From 64K onward, the advantage of the H975i became more pronounced. At 64K block size, improvements reached 40-70%, with throughput rising by more than 12-17 GiB/s compared to the H965i. At 128K-256K, the uplift grew stronger, where the H975i consistently delivered 50-70% higher throughput at moderate to high thread counts.
The most dramatic performance gap appeared at larger block sizes (512K through 10M). At 512K, the H975i achieved gains of +31 to +56 GiB/s, equating to 60-80% improvement over the H965i. At 1M block size, the lead extended further, with throughput jumps of +40 to +68 GiB/s, representing 70-90% gains. Finally, at 5M and 10M block sizes, the PERC 13 nearly doubled throughput compared to the PERC 12, with deltas of +75 to +79 GiB/s, translating into 100% improvement in some thread-rich scenarios.
Overall, the PERC 13 controller showed a clear generational leap in sequential write performance. While differences are minor at the smallest block sizes, once workloads scale past 64K, the H975i consistently delivers 50–100% higher throughput, firmly establishing its superiority over the H965i in write-intensive sequential workloads.
GDSIO Write Sequential Latency
Latency during sequential writes remained impressively low at smaller block sizes and lower thread counts, often staying under 50 µs through 128K blocks with up to 8 threads. As thread counts increased, latency scaled more noticeably. For example, latency reached 392 µs at 512K with 32 threads and exceeded 1 ms at 1M block size with 64 threads.
Saturation effects became more evident at the largest block sizes and highest concurrency levels. Latency rose to 12.4 ms at 5M with 128 threads and peaked at 50.3 ms at 10M with 256 threads.
The most efficient operating point for sequential write workloads occurred at the 1M or 5M block sizes with 8 to 16 threads, where throughput reached 87.9 to 101.2 GiB/s while latency remained within 178 µs – 1.7 ms, providing strong sustained performance without triggering excessive write queue delays.
MLPerf Storage 2.0 Performance
To evaluate real-world performance in AI training environments, we utilized the MLPerf Storage 2.0 test suite. MLPerf Storage is specifically designed to test I/O patterns in real, simulated deep learning workloads. It provides insights into how storage systems handle challenges such as checkpointing and model training.
Checkpointing Benchmark
When training machine learning models, checkpoints are essential for periodically saving the model’s state. This helps prevent loss of progress due to interruptions, such as hardware failures, enables early stopping during training, and allows researchers to branch from various checkpoints for experiments and ablations.
The checkpoint save duration comparison revealed that Dell PERC13 consistently outperformed PERC12 across all model configurations. PERC 13 achieved save times ranging from 7.61 to 10.17 seconds, while PERC12 required 10.41 to 20.67 seconds for the same operations. The performance gap was most pronounced with the 1T parameter model, where PERC13 completed saves in just over 10 seconds compared to PERC12’s 20+ seconds. This represents approximately a 50% reduction in save time for the largest models.
Examining the Save throughput results, the data showcases PERC13’s superior bandwidth utilization, consistently delivering higher data transfer rates. PERC13 achieves throughput between 11.46 and 14.81 GB/s, with peak performance on the 1T model. In contrast, PERC12 tops out at 9.49 GB/s and drops to 6.98 GB/s for the largest configuration. The newer controller maintains more stable performance across different model sizes, suggesting better optimization for handling large sequential writes typical of checkpoint operations.
Micron 7600 MAX Specifications
The table below outlines the supported specifications for the Micron 7600 MAX, a mixed-use PCIe Gen5 NVMe SSD rated for up to 3 drive writes per day (DWPD).
| Micron 7600 MAX Specifications (U.2 / E3.S / E1.S) | |||||
|---|---|---|---|---|---|
| Use Case |
Контактная информация
Beijing Qianxing Jietong Technology Co., Ltd.
Контактное лицо: Ms. Sandy Yang Телефон: 13426366826 | ||||