
HighPoint RocketStor 8631DH
Direct GPU-to-Storage P2P Routing: Compute & NVMe Storage Expansion Solution Bundle
Single multi-slot workstation GPU—such as the NVIDIA® RTX™ PRO 6000 Blackwell—alongside direct internal front-panel NVMe storage expansion from existing general-purpose 1U and 2U enterprise servers without sacrificing adjacent riser slots.
The single-width Rocket 7638D hybrid host card installs cleanly into a single 20.32 mm server slot (Low-Profile or Full-Height), integrating an external 64 GB/s CopprLink port with dual internal MCIO x8 ports (SFF-TA-1016) for direct connectivity to front-panel NVMe storage backplanes.
By establishing native PCIe Gen5 Peer-to-Peer (P2P) DMA switching across its downstream ports, the onboard Broadcom 48-lane switch fabric allows NVMe storage to stream directly into GPU VRAM across a local hardware crossbar, eliminating host CPU memory bounce buffers and substantially reducing CPU I/O overhead.
Crucially, the solution isolates power and heat outside the server chassis: the RocketStor 8631D enclosure features an independent 1300W CRPS industrial power supply (drawing 0W from the host server) compatible with separate standard AC power sources. Paired with an acoustically tuned PSU and dedicated cooling fans, the enclosure delivers a measured 65.9 dBA under full load, preventing aggressive server fan ramp-up and enabling high-throughput GPU + NVMe compute acceleration in office-adjacent edge environments.
Deployment Challenges & Solutions
Deploying high-power workstation GPUs and high-throughput NVMe storage inside existing general-purpose 1U and 2U enterprise servers frequently exposes four compounding infrastructure bottlenecks:
1 | Physical Clearance & PCIe Slot Starvation
Challenge: Standard enterprise 1U and 2U server risers standardize on single-slot pitch (20.32 mm / 0.8 inches). Modern multi-slot workstation GPUs (measuring up to 88 mm in width) physically cannot fit inside standard 1U chassis and mechanically obstruct two to three adjacent slots in many conventional 2U server risers. This can lead to severe slot starvation, leaving limited physical space for dedicated internal NVMe storage HBAs or high-speed clustering network cards.

The RocketStor 8631DH Solution: The single-width Rocket 7638D host adapter occupies a single riser slot, converging an external 64 GB/s CopprLink GPU port and dual internal MCIO x8 NVMe storage ports onto a single PCIe Gen5 x16 host connection. It externalizes the multi-slot GPU while simultaneously driving internal NVMe storage (direct-attached for up to four drives, or via switched UBM backplanes for up to sixteen drives), preserving remaining riser slots for 100GbE/200GbE networking.
2 | Host Server Power Delivery Constraints
Challenge: Deploying a 600W-class workstation GPU introduces substantial board-power demand, placing heavy additional load on the host server's PSU, internal power planes, and auxiliary GPU power connections. Under demanding matrix workloads, transient power excursions can challenge server power budgets or protection thresholds in certain compact server designs, potentially risking power throttling or Over-Current Protection (OCP) shutdowns.

The RocketStor 8631DH Solution: The accelerator's 600W+ footprint is powered entirely by the RocketStor 8631D enclosure's independent 1300W CRPS industrial power supply (drawing 0W from the host server). This isolates high-power demands and transient current spikes from server power planes and connects directly to standard AC power outlets.
3 | Host Chassis Thermal & Acoustic Burden
Challenge: Dissipating 600W+ of trapped GPU heat inside dense 1U/2U server chassis significantly increases internal cooling demands. To prevent thermal throttling, server chassis fans frequently ramp up to aggressive high-speed profiles, generating intrusive acoustic whine that disrupts noise-sensitive engineering labs, edge IT closets, or office-adjacent equipment rooms.

The RocketStor 8631DH Solution: Acoustically Optimized External GPU Cooling (65.9 dBA at Full Load) — By relocating the 600W-class GPU's thermal load outside the host server, the RocketStor S8631DH reduces internal cooling demands and helps minimize aggressive chassis fan operation. The RocketStor 8631D enclosure incorporates an acoustically tuned industrial PSU and dedicated cooling fans, delivering a measured 65.9 dBA under HighPoint's specified full-load test conditions. This makes the solution particularly relevant for on-premises AI installations in dedicated IT laboratories, edge-computing rooms, and office-adjacent equipment areas where cooling noise is an operational concern.
4 | GPU-to-NVMe Data Movement Bottlenecks
Challenge: In traditional server storage architectures, streaming high-resolution sensor, video, or checkpoint datasets from front-panel NVMe storage into GPU memory requires intermediate CPU management and system RAM bounce buffers. This host memory staging consumes substantial host CPU cycles, introduces latency, and caps overall ingestion throughput.

The RocketStor 8631DH Solution: PCIe Architecture-Driven Direct P2P Storage Pipeline — The onboard Broadcom 48-lane PCIe Gen5 switch core establishes a native, hardware-level crossbar switching fabric between the internal MCIO NVMe ports and the external GPU port. By enabling direct peer-to-peer DMA transfers across the local switch fabric, storage data bypasses host CPU root complexes and system RAM bounce buffers entirely. This architecture substantially frees host CPU compute cycles for preprocessing and application logic while minimizing end-to-end data transfer latency.*
Why Choose the RocketStor 8631DH?

Converged GPU Compute & Storage Ingestion in One Slot: Provides dedicated Gen5 x16 bandwidth for GPU compute and dedicated Gen5 x16 bandwidth across two MCIO x8 ports (connecting directly to 4 or up to 16 x4 NVMe devices via UBM backplane infrastructure) directly from a single Low-Profile slot, eliminating server riser slot starvation.

Hardware-Level Direct P2P Streaming from Fast NVMe Storage to GPU: The onboard Broadcom 48-lane PCIe Gen5 switch establishes a native hardware-level P2P DMA data pipeline between front-panel NVMe storage and the external accelerator. This architectural foundation enables direct storage-to-accelerator frameworks like NVIDIA GPUDirect Storage (GDS) to stream data straight from NVMe arrays to GPU VRAM, bypassing host system memory, slashing CPU utilization to <10%, and eliminating latency-inducing bounce buffers.

Zero Adjacent Riser Slot Blockage: Occupies strictly one single-width riser slot (20.32 mm pitch), preserving valuable remaining riser slots for 100GbE/200GbE clustering network interface cards.

Physical, Electrical & Thermal Decoupling: Completely isolates the 600W+ accelerator's power and thermal footprint outside the server chassis, drawing 0W from the server power plane and maintaining full GPU boost clocks at a controlled 65.9 dBA.
Solution Architecture


Validated GPU-to-Storage Performance: NVIDIA GPUDirect Storage (GDS) Benchmarks
The Host CPU Root Complex Bottleneck: In conventional enterprise servers, PCIe peripheral slots and NVMe drive controllers are segregated across different CPU Root Complexes.
Moving data from storage to GPU memory forces a high-overhead staging loop:
NVMe SSD → PCIe Switch → CPU Root Complex → System Memory (RAM) → CPU Root Complex → GPU VRAM.
This host memory staging introduces substantial latency, limits throughput to host bus speeds, and consumes 85% to 99% of host CPU processing capacity.
Hardware-Level Direct P2P Architecture & Platform Independence: The Rocket 7638D consolidates the external accelerator and internal NVMe drive backplanes into a single, unified PCIe Root Complex under its onboard Broadcom 48-lane PCIe Gen5 switch core. By establishing dedicated downstream Gen5 x16 data paths (64 GB/s line rate) to both the GPU enclosure and the internal MCIO storage ports, the card provisions a direct hardware-level crossbar pipeline:
NVMe SSD ◄──► PCIe Gen5 Switch ◄──► GPU VRAM
Crucially, because this high-bandwidth transfer loop routes locally across the switch core, it operates independently of the upstream host server slot speed. When deployed in an existing PCIe Gen4 server or Gen4 riser slot, the RS8631DH delivers full wire-speed Gen5 throughput (57+ GB/s) and sub-microsecond P2P transfers—delivering near-100% Gen5 performance parity without forcing organizations to replace existing server infrastructure.
Understanding the Two Transfer Modes Evaluated:
• CPU/RAM-Staged Transfer (Baseline): The benchmark's conventional transfer configuration, in which data passes through host CPU root complexes and system RAM before reaching GPU memory.
• GDS Direct Transfer (RocketStor 8631DH): The tested NVIDIA GPUDirect Storage configuration, using eligible direct DMA transfers between NVMe storage and GPU memory across the local PCIe Gen5 switch fabric, completely bypassing host memory staging.
PCIe Gen5 Server Benchmark (Intel® Xeon® Silver 4410Y)
Test Setup: Rocket 7638D · NVIDIA RTX PRO 4500 · 16x KIOXIA NVMe SSDs (RAID0) · NVIDIA gdsio

PCIe Gen4 Server Benchmark (AMD EPYC™ 7282)
Test Setup: Rocket 7638D · NVIDIA RTX PRO 4500 · 16x KIOXIA NVMe SSDs (RAID0) · NVIDIA gdsio
★ Preserves Existing Gen4 Server Investment

*Test Environment: R7638D host AIC with RS8631D enclosure, NVIDIA RTX PRO 4500 GPU (Driver 590.48.01, CUDA 13.3, GDS 1.18.0.66), 16x KIOXIA NVMe SSDs in RAID0 via dual MCIO x8 UBM cabling, Ubuntu 22.04 LTS (kernel 5.15.0-122). Tested using official NVIDIA gdsio benchmark tool (IO block 2M, 64 workers for sequential; 4K, 256/64 workers for random). End-to-end performance depends on drive models, RAID layer, filesystem, and GPU application support.
Key Architectural Takeaways & Enterprise ROI Benefits:
1. Wire-Speed PCIe Gen5 Throughput: Direct P2P DMA routing over the Broadcom switch core achieves 57,328 MB/s sequential read and 55,670 MB/s sequential write, eliminating host memory bounce buffers and lifting throughput by up to 78.3% over CPU-relayed transfers.
2. 99.9% Gen5 Performance on Legacy Gen4 Servers (Capex Protection): Because the onboard switch fabric localizes the high-speed data loop directly between internal storage and the external GPU, the RS8631DH delivers 57,264 MB/s sequential read on a PCIe Gen4 host server—matching 99.9% of native Gen5 speeds. This enables enterprise organizations to deploy modern NVIDIA Blackwell workstation accelerators on existing Gen4 server nodes, eliminating the need for a $15,000+ full-system chassis refresh.
3. 32.9% to 40.2% Latency Reduction: Slashing direct memory access latency from 82 µs down to 55 µs on Gen5 (and 116 µs down to 69 µs on Gen4) removes data-waiting stalls, keeping GPU Tensor and CUDA cores continuously saturated during high-resolution volumetric rendering and iterative model checkpointing.
4. Up to 92.8% CPU Overhead Offload: Bypassing CPU-managed copy loops reduces host CPU load from 98.0%–99.5% down to just 7.1%–12.0% (a 92.8% drop), releasing massive CPU processing headroom for application preprocessing, sensor ingestion, and operating system tasks.
Request the Complete RocketStor 8631DH Performance Report
Evaluate the in-depth architectural and benchmark findings for your engineering team. The full technical report includes comprehensive Test System Environmental Setup details (including PCIe Gen5 hardware cabling, slot topologies, BIOS subsystem parameters, and configurations across the Intel & AMD based platforms, and benchmark tool parameters).
How to Request the Full Benchmark Report:
Solution Bundle Contents
The RocketStor 8631DH is shipped in a single, comprehensive retail package (Accelerator and host server sold separately):
1x Rocket 7638D Hybrid PCIe Gen5 x16 Host Interface Card (Pre-installed Low-Profile bracket; Full-Height bracket included)
1x RocketStor 8631D PCIe Gen5 Accelerator Enclosure (Integrated 1300W CRPS PSU & Astera Labs Retimer backplane)
1x Certified 1M PCI-SIG® CopprLink™ CDFP (SFF-TA-1032) Cabled Interconnect (64 GB/s)
1x 12VHPWR / 12V-2x6 to Dual 8-Pin Power Harness
1x Regional AC Power Cord & Quick Installation Guide (QIG)
- Specification ParameterRocketStor 8631DH: PAE Compute & NVMe Storage BundlePackaging & Retail ModelRocketStor 8631DH (Single Retail Box)Hardware Core Topology1:1 Compute + NVMe Storage FabricIncluded Host Interface Card1x Rocket 7638D Host Switch CardHost Switch ArchitectureBroadcom 48-Lane PCIe Gen5 Switching TechnologyHost Interface & Data RatePCIe 5.0 x16 (64 GB/s line-rate; backward compatible with Gen4)Host Riser Form FactorLow-Profile (68.9 mm x 165 mm); Single-widthIncluded BracketsFull-Height & Low-Profile BracketsExternal Interconnect Ports1x External CDFP Port (Dedicated PCIe 5.0 x16)CopprLink CompliancePCI-SIG CopprLink compliant (CDFP Pinout / SFF-TA-1032)Internal NVMe Storage Ports2x Internal MCIO x8 Ports (Default 4x4 mode; UBM & VPP ready)Storage Device CompatibilityDirect support for PCIe NVMe SSDs (U.2, U.3, E1.S, E3.S, PCIe AIC)Included Enclosures1x RocketStor 8631D EnclosureSignal Retiming TechnologyAstera Labs PCIe 5.0 Active RetimerGPU / Accelerator Slots1x PCIe 5.0 x16 Slot (GPU not included)Supported Accelerator ClearanceUp to 370 mm (L) x 170 mm (H) x 88 mm (W) (Full-length, triple-slot cards)Accelerator CompatibilityNVIDIA Blackwell RTX PRO 6000, Ada Lovelace RTX 6000, or any Gen5/Gen4 Workstation GPUDedicated Enclosure Power1x Integrated 1300W CRPS Industrial PSUSustained GPU Power Capability600W+ sustainedIncluded GPU Power Cables1x 12VHPWR to Dual 8-Pin (2x4) Power CableHost Power Isolation100% Externalized GPU PowerAcoustic Noise Profile65.9 dBA under sustained 100% full loadCooling & Fan ControlsAutomated Smart Fan (Linear ramp 65°C–75°C, Full >75°C) & 5 Manual Preset SpeedsHardware Fault Buzzer AlertAudio alerts for Fan Failure (<100 RPM) & Retimer Overheat (>100°C) with rear Mute buttonDriver & Software Architecture100% Driverless: Native In-Kernel OS Support (No proprietary host drivers required)Operating System SupportRHEL, Rocky Linux, Ubuntu Server LTS, Proxmox VE, Windows Server 2022/2025Direct DMA & Ingestion ProtocolsNVIDIA GPUDirect Storage (GDS) Compatible (Direct NVMe-to-VRAM bypass via MCIO)System RAS & SecurityHardware/Software Secure Boot, Stores VPDIncluded Package Contents (single box)• 1x Rocket 7638D Host Switch Card• 1x Full-Height (Pre-installed) & 1x Low-Profile Bracket• 1x RocketStor 8631D Enclosure• 1x 1m CopprLink Gen5 Cable• 1x 12VHPWR to Dual 8-Pin Cable• 1x UL AC Power Cord & 1x QIG
Application Deployment Ideas
3D Medical Imaging
(Cryo-EM, CT/MRI)

Direct streaming of multi-gigabyte tomographic datasets from front NVMe scratch volumes to GPU memory, cutting 3D mesh rendering and slice reconstruction cycles by up to 60%.
8K / 12K Virtual Production

Real-time display multi-camera rendering and uncompressed 8K/12K video raster playback without frame drops, sustained by line-rate NVMe-to-VRAM transfers.
High-Throughput
Model Checkpointing

Accelerated checkpoint save and restore cycles in multi-billion parameter model execution, minimizing training pipeline pauses by offloading I/O management from the host CPU.
Edge Multi-Sensor Fusion
& Autonomous Telemetry

Direct streaming of multi-gigabyte tomographic datasets from front NVMe scratch volumes to GPU memory, cutting 3D mesh rendering and slice reconstruction cycles by up to 60%.
HighPoint PCIe Accelerator Expansion (PAE) Solution Bundles
Which Bundle is Right for Me?
Selection Criteria | RocketStor 8631DW | RocketStor 8631DH | RocketStor 8631DD |
(Single-GPU Offload Bundle) | (Compute & Storage Bundle) | (Dual-Port Scalable Bundle) |
Application Deployment Ideas | • Local 70B+ LLM Inference (vLLM) • Edge Multi-Camera Video Analytics • Interactive 3D CAD/CAM Workstations | • 3D Medical Imaging (Cryo-EM, CT/MRI) • 8K/12K Virtual Production (UE5) • High-Throughput Model Checkpointing | • CAE / CFD Simulation (OpenFOAM, Ansys) • Distributed Multi-GPU AI Scaling • Large-Mesh Finite Element Solvers |
Server Deployment Requirement | Standard Server with One PCIe Slot (Single Width or wider) (Low-Profile or Full-Height Bracket Supported) | Standard Server with One PCIe Slot (Single Width or wider) (Low-Profile or Full-Height Bracket Supported) | Standard Server with One PCIe Slot (Single Width or wider) (Full-Height Bracket; Horizontal Riser Required for 1U) |
GPU Form-Factor & Capacity | Accommodates 1x Multi-Slot GPU (Up to 370 mm L × 170 mm H × 88 mm W) | Accommodates 1x Multi-Slot GPU (Up to 370 mm L × 170 mm H × 88 mm W) | Accommodates 1x Multi-Slot GPU Out of Box Scalable to 2x Multi-Slot GPUs (via 2nd chassis) |
Internal NVMe Storage Expansion | None (Dedicated 1:1 GPU Compute) | 2x MCIO x8 Ports (SFF-TA-1016) Direct connection to 4x4 NVMe backplanes | None (Dedicated Dual-GPU Compute) |
Primary Architectural Value | 1:1 Dedicated PCIe Gen5 x16 Pipeline Lowest-cost entry point to externalize a single multi-slot workstation GPU from a single-width server slot with dedicated 64 GB/s throughput | Direct GPU-to-Storage P2P Routing via PCIe Gen5 Switch Fabric Dedicated Gen5 x16 connectivity for GPU and internal NVMe storage backplanes | Direct GPU-to-GPU P2P Routing via PCIe Gen5 Switch Fabric Dual dedicated Gen5 x16 GPU connections for unshared multi-GPU scaling |
Included Hardware in Box | • 1x Rocket 7634D Host AIC (LP & FH) • 1x RS8631D 1300W Enclosure (65.9 dBA) • 1x 1m CopprLink Cable + 12VHPWR Cable | • 1x Rocket 7638D Host AIC (LP & FH) • 1x RS8631D 1300W Enclosure (65.9 dBA) • 1x 1m CopprLink Cable + 12VHPWR Cable | • 1x Rocket 1638E Host AIC (FHHL) • 1x RS8631D 1300W Enclosure (65.9 dBA) • 1x 1m CopprLink Cable + 12VHPWR Cable |
MSRP (USD) | $2,298.00 | $2,798.00 | $3,298.00 (Optional 2nd Enclosure: $1,299.00) |
RocketStor 8631DH - Frequently Asked Questions
A: Choose the RocketStor 8631DH if your application requires both GPU acceleration and direct high-speed access to front-panel NVMe storage. While the RocketStor 8631DW is designed strictly for dedicated 1:1 GPU compute offload, the RS8631DH integrates 2x internal MCIO x8 ports directly onto the host adapter. This establishes a native PCIe Gen5 P2P DMA path between your NVMe drives and the external accelerator through the card's onboard switch, providing the physical data conduit required by frameworks like NVIDIA GPUDirect Storage (GDS) while freeing up other server slots.
A: The included Rocket 7638D host adapter features two standard SFF-TA-1016 (MCIO x8) delivering a total of 16 PCIe Gen5 lanes. In direct-attached configurations, each MCIO x8 port connects directly to up to two x4 NVMe SSDs (four drives total across both ports). For configurations requiring up to sixteen drives, the MCIO ports connect to server backplanes equipped with onboard PCIe switches using Universal Backplane Management (UBM), allowing all 16 drives to access and dynamically share the 16 lanes of Gen5 storage bandwidth provided by the host adapter.
A: No. The Rocket 7638D utilizes an onboard Broadcom 48-lane PCIe Gen5 switch. The host interface receives a full 16-lane connection to the host server, while the downstream fabric provides an independent, unshared Gen5 x16 connection to the external accelerator and dedicated x16 lanes routed across the two MCIO x8 storage ports. Because direct P2P DMA transfers occur locally across the switch fabric, storage ingestion does not saturate the upstream host bus.
A: Yes. Because the Rocket 7638D's onboard Broadcom PCIe Gen5 switch establishes a local hardware crossbar between the internal MCIO storage ports and the external GPU port, P2P DMA transfers (such as NVIDIA GDS) route directly across the switch fabric at full Gen5 x16 line rates (64 GB/s). Benchmark tests verify 57,264 MB/s throughput on PCIe Gen4 platforms—delivering 99.9% performance parity with Gen5 systems and eliminating the need for expensive server replacements.
A: No. The Rocket 7638D conforms strictly to the standard PCI-SIG CEM single-slot pitch (20.32 mm / 0.8 inches). It installs cleanly into a single slot, leaving adjacent riser positions available for high-speed clustering network adapters or additional expansion cards.
A: No. The accelerator draws 0W from the host server power plane. All operational power—including sustained 600W+ draws and transient microsecond current spikes—is supplied independently by the RocketStor 8631D’s 1300W CRPS power supply, while dedicated cooling fans maintain a measured 65.9 dBA sound level under full load.
Ready to Modernize Your AI Infrastructure?
Verify your server platform, riser configuration, and NVIDIA Blackwell Workstation GPU models with HighPoint’s engineering team before deployment.
Global Channel Distribution
Part Number: RocketStor 8631DH | MSRP: $2,798.00
HighPoint PAE Solution Bundles are available through authorized distributors and resellers worldwide:
North America:

Asia:

Europe / United Kingdom:

.png)





