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.
| 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 | midv945rm
Key CPU features:
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 Results show that the MIDV945RM delivers up to
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. | and Deployment Scenarios
A Comprehensive Evaluation of the MIDV945RM High‑Performance Compute Module: Architecture, Benchmarks, and Deployment Scenarios