
HighPoint RocketStor 8631DD
GPU-to-GPU P2P Routing: Scalable Dual-GPU Compute Expansion Solution Bundle
The HighPoint RocketStor 8631DD is a PCIe Accelerator Expansion (PAE) solution bundle engineered to deploy multi-slot workstation GPUs—such as the NVIDIA® RTX™ PRO 6000 Blackwell—from standard 1U and 2U enterprise servers with scalable dual-accelerator physical capacity and unshared switch-level interconnects without sacrificing adjacent riser slots.
The single-width Rocket 1638E host card installs cleanly into a single 20.32 mm server slot (Full-Height Half-Length, compatible with standard 1U horizontal butterfly risers and 2U+ server risers), provisioning two independent external 64 GB/s CopprLink ports. By establishing native PCIe Gen5 Peer-to-Peer (P2P) DMA crossbar switching across its dual downstream ports, the onboard Broadcom 48-lane switch fabric delivers dual dedicated Gen5 x16 hardware links.
This physical foundation provides the dedicated, non-blocking data paths required by multi-GPU communication frameworks, such as NVIDIA CUDA P2P and NCCL, to stream inter-device traffic directly across the switch fabric without requiring motherboard slot bifurcation or routing through host CPU memory bounce buffers. ►Learn More
Crucially, the RocketStor 8631DD isolates power and heat outside the server chassis: each 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 acoustically tuned power supplies and dedicated cooling fans, each enclosure dissipates full GPU thermal loads externally at a measured 65.9 dBA under full load, preventing aggressive server fan ramp-up and enabling dual-GPU workstation accelerator capacity in space-constrained edge deployments.
Deployment Challenges & Solutions
The Mechanical, Electrical, Acoustic & Bandwidth Barrier
Compute-intensive enterprise workloads in Computational Fluid Dynamics (CFD), Finite Element Analysis (FEA/CAE), and distributed multi-agent LLM scaling demand multi-GPU compute power and high-speed inter-accelerator communication. However, deploying dual multi-slot workstation GPUs (such as the 600W NVIDIA RTX PRO 6000 Blackwell) inside standard 1U and 2U enterprise servers creates four critical bottlenecks:
1 | Physical Clearance in 1U/2U Racks
Challenge: Due to compact chassis design requirements, standard 1U and 2U servers typically utilize standard single-slot spacing (20.32 mm / 0.8 inches). Two 88 mm multi-slot workstation GPUs cannot physically fit inside 1U servers, and when installed into in 2U chassis, they may block as many as 6 adjacent slots—eliminating room for additional PCIe cards or storage devices.

The RocketStor 8631DD solution: The Rocket 1638D only requires a single PCIe slot - it's dual PCIe Gen5 x16 CopprLink ports can each support a RocketStor 8631D enclosure, enabling the GPU's to be moved outside of the host chassis.
2 | Host Server Power Delivery Constraints
Challenge: Deploying two 600W NVIDIA BWE GPUs adds up to 1,200W of GPU board-power demand, placing substantial load on the host's PSU, internal power distribution, and auxiliary GPU power connections. Under demanding workloads, this additional load and transient power excursions can exceed the available power budget or protection thresholds of conventional 1U/2U servers, potentially causing power throttling, instability, or Over-Current Protection (OCP) shutdowns.

The RocketStor 8631DD solution: Each enclosure includes a dedicated 1300W CRPS power supply to power the GPU independently from the host server’s GPU power budget. Add GPU capability without relying on the host PSU to supply the external GPU’s operating power.
3 | Host Server Thermal & Acoustic Burden
Challenge: Dissipating up to 1,200W of additional waste heat within dense 1U or 2U server chassis creates severe thermal management challenges. Confined internal spaces and restricted airflow paths frequently force host system cooling fans to ramp up aggressively to high-speed profiles, generating intrusive acoustic noise that disrupts office-adjacent edge environments.

The RocketStor 8631DD solution: The RocketStor 8631D enclosures enable the GPUs to be located outside the host chassis. A dedicated acoustically-tuned cooling system relocate the GPU heat load away from the host chassis at a controlled 65.9 dBA under full load.
4 | Single-Slot PCIe Bifurcation & GPU P2P Bottleneck
Challenge: Deploying two external GPUs from a single PCIe Gen5 x16 server slot through passive bifurcation requires an x8/x8 lane configuration, which limits each GPU to approximately 32 GB/s of theoretical bandwidth per direction—half the bandwidth of a dedicated Gen5 x16 connection. Unlike an active PCIe switch, passive bifurcation provides no local inter-GPU switching fabric, leaving direct GPU-to-GPU P2P communication dependent on the host root-complex topology and platform support, potentially increasing communication latency and host-side data movement.


The RocketStor 8631DD solution: Dedicated PCIe Gen5 Switch technology combined with an external CopprLink interface provides a dedicated x16 lanes of Gen5 bandwidth (up to 64GB/s) for each GPU, requiring zero host-side bifurcation.
Keep the Server. Upgrade the AI.
RocketStor PAE Solution Bundles help enterprises extend the useful life of existing servers and workstations by deploying NVIDIA RTX PRO Blackwell GPUs externally. Preserve host networking and storage resources while adding a dedicated PCIe Gen5 connection, independent GPU power, and external thermal management—without replacing an otherwise capable server simply to fit a larger GPU.
Validated GPUDirect P2P Performance: Overcoming the Server Root-Complex Barrier
Standard 1U and 2U enterprise server motherboards often restrict direct peer-to-peer PCIe transactions across root complexes, forcing multi-GPU traffic through host CPU memory and causing severe communication latency stalls. Equipped with an onboard Broadcom 48-lane PCIe Gen5 switch core, the RocketStor 8631DD establishes a native hardware-level crossbar switching fabric directly between both external accelerators. Testing conducted using the standard CUDA p2pBandwidthLatencyTest on production Intel and AMD enterprise platforms confirms wire-speed peer-to-peer performance:

*Test Environment: RS8631DD (Rocket 1638E + dual RS8631D enclosures) with dual NVIDIA RTX PRO 4500 Workstation Edition GPUs, Ubuntu 24.04 (kernel 6.8.0-31-generic), CUDA Toolkit 13.4.2, Driver 615.71.09, CUDA Sample v13.2.
• Cutting GPU-to-GPU latency down to 0.41 µs (sub-500 ns, a 97.3% reduction): eliminates synchronization stalls during iterative mesh boundary exchanges, keeping GPU compute cores running at peak utilization.
• Delivers 106+ GB/s bidirectional P2P throughput (up to a 391.9% bandwidth increase): removes bus bottlenecks during collective communication operations (such as NCCL All-Reduce), accelerating multi-agent tensor parallelism without requiring specialized multi-GPU server chassis.
• Preserves Host CPU Processing Headroom: Slashing host CPU access latency by up to 71.9% removes the burden of managing intermediate GPU bounce buffers, leaving host CPU threads 100% available for simulation meshing, preprocessing, and system I/O.
Request the Complete RocketStor 8631DD GPUDirect P2P 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, SR-IOV, Re-Size BAR, and ACS/IOMMU configurations across Dell PowerEdge R660 and Supermicro 2025HS-TNR platforms), exact benchmark tool parameters (CUDA Sample p2pBandwidthLatencyTest v13.2 execution scripts), and raw bidirectional bandwidth and latency matrices.
How to Request the Full Benchmark Report:
Why Choose the RocketStor 8631DD?

Dual Dedicated Gen5 x16 GPU Connections: Provides two independent, unshared PCIe Gen5 x16 external ports from a single server slot, eliminating slot bifurcation compromises and preserving dedicated 64 GB/s line-rate bandwidth for each connected accelerator.

Direct GPU-to-GPU P2P Routing via PCIe Gen5 Switch Fabric: The onboard Broadcom 48-lane PCIe Gen5 switch core establishes direct hardware-level Peer-to-Peer (P2P) DMA data paths between downstream ports. This provides the physical switching fabric required by multi-GPU software stacks (such as NVIDIA CUDA P2P and NCCL) to stream gradient updates and collective operations directly between accelerators at line-rate speeds, completely bypassing host CPU cores and eliminating system memory bounce buffers.*

Distributed Power & Acoustic Optimization: Relocates dual 600W workstation accelerators into two independent 1300W CRPS power supplies (0W host server draw) compatible with standard AC power sources, while dedicated cooling fans and acoustically tuned power supplies maintain a measured 65.9 dBA sound level under full load.

Modular On-Demand Dual-GPU Scalability: Ships with one RocketStor 8631D 1300W enclosure out of the box for immediate single-GPU acceleration, allowing enterprise teams to scale to dual-GPU deployments on demand by adding an optional second enclosure without purchasing a new host adapter.

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

Physical, Electrical & Thermal Decoupling: Completely isolates each 600W+ accelerator's power and thermal footprint outside the server chassis, drawing 0W from the server power plane and maintaining full GPU boost clocks without thermal throttling.
*Architectural Note: Applications leveraging NVIDIA CUDA P2P and NCCL can utilize this direct interconnect architecture when supported by the GPU driver, platform configuration (including PCIe ACS and IOMMU settings), and application software.
Solution Architecture

Solution Bundle Contents
The RocketStor 8631DD is shipped in a single, comprehensive retail package (Accelerator and host server sold separately):
-
1x Rocket 1638E Dual-Port PCIe Gen5 x16 Host Interface Card (Pre-installed Full-Height bracket; Low-Profile 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 AC Power Cord & Quick Installation Guide (QIG)
- Specification ParameterRocketStor 8631DD: PAE Dual-Port Scalable Compute BundlePackaging & Retail ModelRocketStor 8631DD (Single Retail Box)Hardware Core Topology1:2 Dual-GPU P2P Switch FabricIncluded Host Interface Card1x Rocket 1638E 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 FactorFHHL (148 mm x 110 mm); Single-widthIncluded BracketsFull-Height BracketExternal Interconnect Ports2x External CDFP Ports (2x Dedicated PCIe 5.0 x16)CopprLink CompliancePCI-SIG CopprLink compliant (CDFP Pinout / SFF-TA-1032)Internal NVMe Storage PortsN/AStorage Device CompatibilityN/AIncluded Enclosures1x RocketStor 8631D Enclosure (Supports optional 2nd RS8631D enclosure)Signal Retiming TechnologyAstera Labs PCIe 5.0 Active Retimers (1 per enclosure)GPU / Accelerator Slots1x PCIe 5.0 x16 Slot (1 per enclosure; GPUs not included)Supported Accelerator ClearanceUp to 370 mm (L) x 170 mm (H) x 88 mm (W) per enclosure (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 PSU (Scales to 2600W with 2nd enclosure)Sustained GPU Power Capability600W+ sustained draw (Scales to 1200W+ sustained with 2nd enclosure)Included GPU Power Cables1x 12VHPWR to Dual 8-Pin (2x4) Power CablesHost Power Isolation100% Externalized GPU PowerAcoustic Noise Profile65.9 dBA per enclosure 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 ProtocolsCUDA P2P DMA Compatible (Switch-level inter-GPU routing; bypasses host CPU/RAM)System RAS & SecurityHardware/Software Secure Boot, Stores VPDIncluded Package Contents (single box)• 1x Rocket 1638E Host Switch Card• 1x Full-Height Bracket (Pre-installed)• 1x RocketStor 8631D Enclosures• 1x 1m CopprLink Gen5 Cables• 1x 12VHPWR to Dual 8-Pin Cables• 1x UL AC Power Cords & 1x QIG
Application Deployment Ideas
CAE / CFD Simulation & Finite Element Solvers

Accelerating high-cell aeromechanical, thermal, and fluid dynamics simulations (OpenFOAM, Ansys Fluent) through direct low-latency Peer-to-Peer memory exchanges between GPUs.
Distributed Multi-GPU
AI Scaling

Scaling distributed multi-agent LLM inference and localized fine-tuning workloads with hardware-level P2P DMA, delivering dedicated Gen5 x16 bus bandwidth to maximize NCCL All-Reduce efficiency across dual external accelerators.
Large-Mesh Finite
Element Analysis

Processing complex structural mechanics, seismic wave propagations, and large-mesh physics simulations across multiple high-performance GPUs without CPU bus contention.
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 8631DD - Frequently Asked Questions
A: Choose the RocketStor 8631DD if your workload requires multi-accelerator compute scaling and direct inter-GPU communication. The RocketStor 8631DD features the Rocket 1638E switch card, which is equipped with dual external CopprLink ports. This design supplies a Gen5 x16 switching fabric that allows software frameworks, such as NVIDIA CUDA P2P and NCCL, to stream data directly between two external accelerators. By doing so, it bypasses host CPU cores and eliminates slot-level bifurcation compromises while avoiding the severe thermal and electrical challenges of housing two 600W cards inside a compact server environment.
A: Rather than placing up to 1,200W of additional board-power demand and transient power excursions directly onto the host server's PSU, the RocketStor 8631DD separates compute power across two independent 1300W CRPS industrial power supplies (drawing 0W from the host server). This isolated architecture allows power delivery to be distributed across separate standard AC power sources without exceeding host server power thresholds.
A: No. Unlike passive bifurcation adapters that require split x8/x8 configurations (limiting each GPU to approximately 32 GB/s per direction), the Rocket 1638E features an onboard Broadcom 48-lane PCIe Gen5 switch that provisions dual dedicated Gen5 x16 connections. Inter-GPU communication routes locally across the switch fabric at full 64 GB/s line-rate speeds without saturating the upstream host bus.
A: Yes. The Rocket 1638E host adapter is a single-width Full-Height Half-Length (FHHL) card. In standard enterprise 1U servers (such as Dell PowerEdge R660, HPE ProLiant DL360, and Supermicro Ultra platforms), the card installs into horizontal butterfly riser cages, which orient the card parallel to the motherboard. It can also be installed into any standard 2U+ full-height riser slot.
A: The RocketStor 8631DD bundle ships with one RocketStor 8631D 1300W enclosure and the dual-port Rocket 1638E host card, allowing immediate single-GPU deployment. You can scale to dual-GPU compute on demand at any time simply by adding an optional second RocketStor 8631D enclosure (which includes its own dedicated 1M CopprLink cable and 1300W power supply) without replacing the host card.
A: No. Both accelerators draw 0W from the host server power plane. Each RocketStor 8631D enclosure houses an independent 1300W CRPS industrial power supply, absorbing all sustained 600W+ loads and transient current spikes externally while dedicated 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: RS8631DD | MSRP: $3,298.00
HighPoint PAE Solution Bundles are available through authorized distributors and resellers worldwide:
North America:

Asia:

Europe / United Kingdom:

.png)





