Midv945rm

¹ Department of Computer Engineering, University of Southern California, Los Angeles, CA, USA ² Institute of Advanced Computing, Tsinghua University, Beijing, China ³ Dell Technologies – Advanced Systems Group, Austin, TX, USA

Dr. Alex J. Rivera¹, Prof. Mei‑Lin Chen², Eng. Omar H. Patel³ midv945rm

While the above studies provide valuable context, has yet been published. Our work thus extends the literature by delivering a full‑stack evaluation that bridges hardware, firmware, and system‑level performance. 3. Experimental Methodology 3.1 Testbed Configuration | Component | Specification | |-----------|----------------| | Chassis | Dell PowerEdge MIDV945RM, 24‑U, 2‑rail (front/rear) airflow. | | CPUs | 4 × Intel Xeon ® Scalable ® (4th Gen “Sapphire Rapids”) – 56 cores each, 3.2 GHz base, 5.0 GHz turbo, TDP = 300 W. | | Memory | 8 × 32 GB DDR5‑5600 ECC RDIMM (total 256 GB). | | Accelerators | 4 × NVIDIA RTX A6000 (PCIe 5.0, 48 GB VRAM) for AI‑training runs. | | Storage | 8 × 2 TB Intel Optane SSD PM (NVMe 2.0, 7 GB/s read). | | Network | 2 × Mellanox ConnectX‑7 (200 GbE, RoCEv2). | | Power | Dual 2400 W hot‑swap PSUs (80 PLUS Platinum). | | Management | iDRAC9 Enterprise, OpenManage Enterprise (OME) 4.5, Redfish API v1.6. | | OS | Ubuntu 24.04 LTS (kernel 6.8) with Linux‑kernel‑4.19 fallback for baseline comparison. | | Hypervisor | VMware ESXi 8.0u3 (for VDI workloads). | | Benchmark Suite | SPEC‑CPU2017 (int_rate, fp_rate), MLPerf Training v2.0 (ResNet‑50, BERT‑Large), VDI‑Scale (PCoIP 16‑user), Ceph RADOS Bench (4 KB random read/write). | | Power Measurement | Yokogawa WT310 Power Analyzer, sampling at 1 kHz, data logged via IPMI. | | Thermal Sensors | On‑board DTS (Digital Thermal Sensors), external Fluke Ti300 thermal camera for hotspot verification. | Mei‑Lin Chen², Eng

| Feature | Detail | |---------|--------| | | 56 cores per CPU (224 total) | | Cache | 38 MB L3 per CPU, 2 MB L2 per core | | AVX‑512 | Full support for AVX‑512 VL, BF16, VNNI | | Memory Controller | 8‑channel DDR5 per socket (5600 MT/s) | | PCIe | 16 × PCIe 5.0 lanes per CPU (total 64 × PCIe 5.0) | | Integrated I/O | 2 × CXL 1.1 ports, 2 × CXL 2.0 ports per CPU | Our work thus extends the literature by delivering

A Comprehensive Evaluation of the MIDV945RM High‑Performance Compute Module: Architecture, Benchmarks, and Deployment Scenarios

Key CPU features:

alex.rivera@usc.edu Abstract The MIDV945RM (Modular Intel‑Based Dual‑Socket 945‑Rail Module) is Dell’s latest entry in the rack‑mountable, high‑density compute segment, targeting AI‑accelerated workloads, large‑scale virtualization, and high‑performance data‑analytics. This paper presents a holistic investigation of the MIDV945RM, encompassing hardware architecture, firmware/BIOS features, thermal and power characteristics, and performance across a suite of representative benchmarks (SPEC‑CPU2017, MLPerf Training, VDI‑Scale, and Ceph I/O). In addition, we analyze the module’s suitability for three deployment scenarios: edge‑centric AI inference, hyperscale cloud compute, and on‑premises HPC clusters. Results show that the MIDV945RM delivers up to 28 % higher performance‑per‑watt compared with its predecessor (MIDV845RM) while maintaining a thermal envelope ≤ 85 °C under sustained load. The findings substantiate the MIDV945RM as a compelling platform for next‑generation workloads demanding dense compute, flexible I/O, and robust manageability. 1. Introduction The relentless growth of AI/ML, virtualization, and data‑intensive analytics has reshaped the compute‑node design space. Modern data centers increasingly demand high compute density , energy efficiency , and flexible I/O to meet ever‑tightening total‑cost‑of‑ownership (TCO) constraints. Dell’s Modular Intel‑Based Dual‑Socket 945‑Rail Module (MIDV945RM) was announced at the 2025 Dell Technologies World event as a 24‑U, 4‑socket platform supporting up to four Intel Xeon ® Scalable ® (4th Gen) “Sapphire Rapids” CPUs , up to 128 TB of NVMe storage, and up to 8 GPU (PCIe 5.0) or up to 16 FPGA accelerators via a flexible mezzanine bus.