Date: Jan 13, 2026 Read: 17
Share:

Facing the diverse storage challenges posed by AI large models, Intel and Union Memory have worked closely to develop a high-performance storage solution capable of flexibly addressing high-throughput, low-latency, and massive-scale expansion scenarios. This solution is built on the Xeon® 6 processor and Union Memory UH812a SSD, leveraging dual-path optimization technology at the kernel and SPDK levels.
High Throughput and Low Latency: AI training relies on high throughput for rapid data loading and processing; AI inference depends on low latency to ensure ultra-fast model response and result delivery.
High Reliability and Availability: While maintaining data security and storage reliability, the solution enables rapid responses to various computing side failures and continuously guarantees the availability of business data.
Massive Data Scalability: Supporting flexible scalability from single-node to large-scale clusters, the solution provides full data lifecycle management and a unified storage platform to significantly enhance data utilization efficiency.
Hardware-Software Co-Design: Through deep integration of the latest software technologies with the storage platform, the solution unlocks the full potential of CPUs, GPUs, and enterprise SSDs, comprehensively meeting critical storage requirements in AI scenarios.
|
Server Platform |
2U2S Storage Server
|
|
|
Processor |
2 Intel Xeon 6767P processors (2.4GHz, 64 cores/128 threads each) |
|
|
Memory Configuration |
512GB DDR5 (16 ×32GB SK Hynix 5600MT/s) Memory Slot Configuration: 1 DIMM per channel |
|
|
CPUO |
CPU1 |
|
|
CPUO_DIMM_A1 |
CPU1_DIMM_A1 |
|
|
CPUO_DIMM_B1 |
CPU1_DIMM_B1 |
|
|
CPUO_DIMM_C1 |
CPU1_DIMM_C1 |
|
|
CPUO_DIMM_D1 |
CPU1_DIMM_D1 |
|
|
CPUO_DIMM_E1 |
CPU1_DIMM_E1 |
|
|
CPUO_DIMM_F1 |
CPU1_DIMM_F1 |
|
|
CPUO_DIMM_G1 |
CPU1_DIMM_G1 |
|
|
CPUO_DIMM_H1 |
CPU1_DIMM_H1 |
|
|
Storage Configuration |
16 Union Memory UP2097T6HKO15LX PCIe Gen5 TLC NVMe SSDs (7.68TB per drive) SSD Installation Layout: 8 drives connected to each socket |
|
|
|
BIOS Options |
Default |
Recommended |
|
Socket config & power configuration |
Socket Config -Advanced PM Config > Package C State Control > Package C state |
Auto |
C0/C1 state |
|
Socket Config -lIO Config > Intel VT-d for Directed l/O > PRS Capability for PCle |
Enable |
Disable |
|
|
Socket Config > Advanced PM Config > CPU -Advanced PM Tuning > Energy Perf BIAS > Workload Config |
Balanced |
I/O sensitive |
|
|
Socket Config > Advanced PM Config > CPU Advanced PM Tuning > Latency Optimized Mode |
Disable |
Enable |
|
|
Socket Configuration --> Processer Configuration Enable LP [Global] (Intel® Hyper-Threading Technology) |
ALL LP |
Enable |
|
|
Socket Configuration >llO Configuration > Global Configuration -> Relax Ordering |
Enable |
Enable |
|
|
Socket Configuration>lIO Configuration > Intel VT for Directed I/O (VT-d) -> lntel® VT-d |
Enable |
Enable |
|
|
Socket Configuration -> Processor Configuration --> Global Configuration --> PCle ASPM |
Per-Port |
Disable |
|
|
OS |
openEuler 22.03 LTS and openEuler 24.03 LTS |
||
|
Kernel |
5.10 and 6.6, combined with Intel-developed kernel NVMe driver optimizations |
||
|
SPDK Version |
25.05 (https://github.com/spdk/spdk) |
||
|
FIO Version |
3.34 |
||
|
NUMA Affinity |
Recommend to bound CPU cores, NVMe SSDs, and memory to the same NUMA node. Use bound CPU cores for each NVMe SSD. |
||
Test results confirm that the dual-path optimization of both kernel and SPDK not only maximizes hardware-software synergy for optimal efficiency, but also consistently delivers industry-leading high throughput and low latency performance, providing applications with robust "deterministic performance" assurance.
SPDK-Based Storage Software
Storage Driver Based on Optimized Kernel
The Intel® Xeon® 6 processor has undergone deep optimization to deliver significantly improved single-thread performance. With more cores, doubled memory bandwidth, and built-in AI acceleration per core, this processor can boost overall performance for diverse workloads—including AI—by up to 2 times. It excels in compute-intensive tasks such as AI inference and machine learning, outperforming comparable general-purpose CPUs. Moreover, the Xeon® 6 processor is better optimized for public cloud environments, providing higher single-vCPU performance for scenarios such as storage services and transactional databases.
The Union Memory UH812a PCIe 5.0 enterprise SSD is built for mission-critical workloads. It delivers sequential read/write speeds up to 14,900/10,500 MB/s, and its proprietary LDPC + DSP engine extends flash endurance by 5 times. Through dynamic adjustment technology, it ensures efficient operation across all scenarios, with peak power consumption kept at ≤ 24W. Certified by multiple Intel tests and passing neutron irradiation testing, its MTBF exceeding 2.5 million hours and an annual failure rate below 0.35% provide high-performance, highly reliable data storage assurance for diverse enterprise scenarios.
Moving forward, Intel and Union Memory will continue to explore more storage best practices, applying them in AI and other related fields, to help enterprise users achieve stable, reliable, and high-performance storage deployments and unlock the core value of data.