R740xd – Dual A40 Build, Refurbished
Warranty: Lifetime
Dell PowerEdge R740xd 24SFF Rack Server – Dual A40 GPU Build (Refurbished)
Powerful 2U Rack Server with 2x NVIDIA A40 48GB GPUs (96GB VRAM) – Ready for AI, Virtualization & Compute Workloads
This fully configured Dell PowerEdge R740xd is a high-performance 2U rack server built for demanding applications. Equipped with two NVIDIA A40 48GB GPUs for 96GB of total GPU memory, dual Xeon Silver processors, and enterprise-grade networking, it delivers excellent performance for larger-model AI inference, multi-user LLM serving, deep learning, VDI, rendering, and virtualized environments.
Key Specifications:
- Server Model: Dell PowerEdge R740xd 24SFF (24x 2.5" Small Form Factor bays)
- Processors: 2x Intel Xeon Silver 4210R (2.4GHz base, 10 cores / 20 threads each) – 20 Cores / 40 Threads total
- Memory: 128GB DDR4 ECC Registered (expandable)
- GPU: 2x NVIDIA A40 48GB GDDR6 (96GB total, professional/data center GPUs) with GPU Enablement Kit
- Storage: 1x Dell 960GB 6G SATA 2.5" SSD
- RAID/Controller: PERC HBA330 (Non-RAID / Pass-Through)
- Networking: Intel X710 2x 10Gb SFP+ + I350 2x 1Gb Base-T rNDC
- Remote Management: iDRAC9 Enterprise (with dedicated management port)
- Power: 2x 1600W AC Hot-Plug Power Supplies (redundant)
- Cooling: High Performance Fan Kit
- Rails: Included Rail Kit
- Condition: Professionally refurbished and tested
- OS: No OS included (supports Windows Server, Linux, VMware ESXi, etc.)
Highlights:
- 2x NVIDIA A40 48GB GPUs (96GB total) – Room for larger models, higher-precision quantization, or two independent AI workloads at once
- 24-bay 2.5" SFF chassis with room for massive storage expansion
- Strong dual-socket performance with 20 cores / 40 threads
- Enterprise remote management via iDRAC Enterprise
- High-bandwidth 10Gb SFP+ networking included
- Redundant 1600W power supplies + high-performance cooling for 24/7 reliability
Perfect For:
- AI / Machine Learning inference on 70B-class models
- Multi-user LLM serving for teams
- Virtual Desktop Infrastructure (VDI)
- Rendering and 3D graphics
- High-density virtualization
- Scientific computing & HPC
This server has been fully tested and is ready to deploy. Ships with all necessary power cables, rails, and bezel (where applicable).
Note: This is a refurbished unit in excellent cosmetic and functional condition. Minor signs of previous rack use may be present.
Compare Our R740xd Builds:
All of these builds share the same GPU-ready platform (24SFF chassis, 128GB RAM, Intel X710/I350 networking, iDRAC Enterprise, GPU enablement kit, high-performance fans, 2x 1600W supplies) — they differ in CPU, storage controller and GPU:
- R740xd – GPU Ready (No GPU Included) – $2,995 – 2x Silver 4210R, HBA330 pass-through. The lowest-cost GPU-ready host; bring your own card.
- R740xd – GPU Ready, Gold 6248 / H740p – $3,493.33 – 2x Gold 6248 (40 cores), PERC H740p RAID. More CPU and hardware RAID; bring your own card.
- R740xd – Dual T4 Build – $4,693 – Silver 4210R platform + 2x Tesla T4 16GB (32GB total), slot-powered.
- R740xd – Dual V100 Build – $5,849 – Silver 4210R platform + 2x Tesla V100 32GB (64GB total). Best value for multi-GPU training and parallel workloads.
- R740xd – A40 Build – $12,159 – Silver 4210R platform + 1x NVIDIA A40 48GB. Best single-GPU performance for AI inference and visualization.
- R740xd – Dual A40 Build (this listing) – $18,920 – Silver 4210R platform + 2x NVIDIA A40 48GB (96GB total). The most GPU memory in the lineup, for 70B-class models with headroom or two workloads side by side.
Need a custom configuration, more storage, or a third GPU? Message us — we can build to your requirements.
What This Server Runs
Two A40s give you 96GB of GDDR6. Inference frameworks such as vLLM and llama.cpp split a model across both cards (tensor or pipeline parallelism), or you can run a separate model on each card. Measured on a single A40 at 4-bit quantization, single-stream — each card in this server delivers the same:
- 14B models (Qwen2.5 14B) — ~9GB VRAM, ~44 tokens/sec
- 27–34B models (Gemma2 27B, Qwen2.5 32B, QwQ 32B, Llava 34B) — 16–20GB VRAM, 23–29 tokens/sec
- 70B models (Llama 3.1 / 3.3 70B) — 39–43GB VRAM, 12–13.5 tokens/sec on one card
What the second card changes:
- 70B at 4-bit with real context headroom — on a single A40 a 70B model leaves little room for KV cache. Split across two cards, roughly 50GB is left for long context and concurrent requests.
- 70B at 8-bit — about 70–75GB of weights. Does not fit on one A40; fits across two.
- Two models at once — for example a 70B model on one card and a 27–34B coding or embedding model on the other.
- Double the serving capacity — one A40 handles 100 concurrent requests on an 8B model at ~1,705 tokens/sec of aggregate output under vLLM. Two cards running independent instances roughly double that.
The two cards communicate over PCIe, which is how most multi-GPU inference deployments run. Full single-card benchmark table, KV cache math and the FP8/MIG limitations: What can you actually run on a 48GB NVIDIA A40?
Configuration Requirements — Already Handled
The A40 is a 300W passively cooled card with no fan of its own, and this server runs two of them. Dropping them into a stock R740xd does not work — Dell's installation guidance requires a specific set of parts, all of which are installed and tested in this build:
- GPU enablement kit (risers, GPU power cables, GPU air shroud, mylar insulation) — Dell 330-BBMO / 85CYN in the US, 490-BEIX / JKFGX in EMEA
- Low-profile heat sinks in both CPU sockets — standard heat sinks physically foul the GPU air shroud, and Dell requires both sockets populated
- High-performance fan kit — mandatory for any GPU configuration in this chassis, not optional
- A dedicated GPU power cable for each card — required for any card above 75W; each A40 draws 300W, and both are cabled and powered
- Identical GPUs — Dell requires all installed GPUs to match in type and model; both cards are NVIDIA A40 48GB
- 1600W power supplies — two 300W cards plus dual processors put this build well above Dell's 1100W floor for GPU configurations; 1600W supplies are the right size for a two-card load
- 30°C maximum inlet temperature — the thermal ceiling once GPUs are installed alongside 150W/8C, 165W/12C, 200W or 205W processors
Assembled from separate secondary-market purchases, that is a stack of line items, condition claims and return windows. Here it is one part number, one warranty and one place to call. The full breakdown, including the 330-BBMO vs 490-BEIX part-number confusion and the two mistakes that stop most self-assembled builds: Putting an NVIDIA A40 in a PowerEdge R740xd
Frequently Asked Questions
What is the largest model this server can run?
With 96GB across two A40s, a 70B parameter model at 8-bit quantization (about 70–75GB of weights) fits, as does a 70B model at 4-bit with roughly 50GB left over for long context and concurrent users. A single A40 tops out at 70B at 4-bit with little headroom.
How do the two GPUs work together?
Over PCIe. Frameworks like vLLM and llama.cpp split models across both cards, or you can run a separate model on each card.
Does the NVIDIA A40 support FP8 or MIG?
Neither. The A40 is an Ampere GA102 card — FP8 tensor support began with Ada Lovelace and Hopper, and Multi-Instance GPU partitioning is limited to A100 and H100. The A40 does support SR-IOV for virtualization.
Can I add a third GPU?
The R740xd supports up to three double-width accelerators, and Dell requires all installed GPUs to be identical in type and model. Contact us before adding a third A40, as the thermal, power and riser assignments change and we can build it to order.
Does it come with an operating system?
No OS is included. The server supports Windows Server, Linux distributions, VMware ESXi and Proxmox. iDRAC9 Enterprise is licensed for remote installation and management.
Can I expand the storage?
Yes. The chassis has 24 x 2.5" SFF bays and ships with one 960GB SATA SSD, leaving 23 bays open. The PERC HBA330 runs in pass-through mode, which suits ZFS, Ceph and software-defined storage. Tell us your capacity target and we will quote the drives with the server.
What condition is it in and what warranty applies?
Professionally refurbished, fully tested and burned in, in excellent cosmetic and functional condition with minor signs of previous rack use possible. Backed by the Resilient Tec lifetime warranty, reachable directly rather than through a manufacturer ticket queue.
What networking is included?
An Intel X710 dual-port 10Gb SFP+ rNDC plus an Intel I350 dual-port 1Gb Base-T, with a dedicated iDRAC management port. We can supply matching 10G SFP+ DAC cables or transceivers with the server.
Further Reading
- Putting an NVIDIA A40 in a PowerEdge R740xd — every part the card needs, the enablement kit part numbers, and why a stock chassis throttles it
- What can you actually run on a 48GB NVIDIA A40? — benchmarked model sizes, tokens per second, and vLLM concurrency limits
- Why your GPU won't work in your server — the eight installation problems and the fix for each
- Custom Dell R740xd GPU servers: build options, prices, and the forum questions nobody answers — which GPU power cable (4VPD3 vs TR5TP), fans at 100%, LFF chassis limits, and the builds compared
- The cloud sent you a bill — on-prem inference versus rented GPU time
- GPU servers & on-prem AI guide, GPU & AI infrastructure guide and enterprise servers guide
Dell PowerEdge R740xd 24SFF Rack Server – Dual A40 GPU Build (Refurbished)
Powerful 2U Rack Server with 2x NVIDIA A40 48GB GPUs (96GB VRAM) – Ready for AI, Virtualization & Compute Workloads
This fully configured Dell PowerEdge R740xd is a high-performance 2U rack server built for demanding applications. Equipped with two NVIDIA A40 48GB GPUs for 96GB of total GPU memory, dual Xeon Silver processors, and enterprise-grade networking, it delivers excellent performance for larger-model AI inference, multi-user LLM serving, deep learning, VDI, rendering, and virtualized environments.
Key Specifications:
- Server Model: Dell PowerEdge R740xd 24SFF (24x 2.5" Small Form Factor bays)
- Processors: 2x Intel Xeon Silver 4210R (2.4GHz base, 10 cores / 20 threads each) – 20 Cores / 40 Threads total
- Memory: 128GB DDR4 ECC Registered (expandable)
- GPU: 2x NVIDIA A40 48GB GDDR6 (96GB total, professional/data center GPUs) with GPU Enablement Kit
- Storage: 1x Dell 960GB 6G SATA 2.5" SSD
- RAID/Controller: PERC HBA330 (Non-RAID / Pass-Through)
- Networking: Intel X710 2x 10Gb SFP+ + I350 2x 1Gb Base-T rNDC
- Remote Management: iDRAC9 Enterprise (with dedicated management port)
- Power: 2x 1600W AC Hot-Plug Power Supplies (redundant)
- Cooling: High Performance Fan Kit
- Rails: Included Rail Kit
- Condition: Professionally refurbished and tested
- OS: No OS included (supports Windows Server, Linux, VMware ESXi, etc.)
Highlights:
- 2x NVIDIA A40 48GB GPUs (96GB total) – Room for larger models, higher-precision quantization, or two independent AI workloads at once
- 24-bay 2.5" SFF chassis with room for massive storage expansion
- Strong dual-socket performance with 20 cores / 40 threads
- Enterprise remote management via iDRAC Enterprise
- High-bandwidth 10Gb SFP+ networking included
- Redundant 1600W power supplies + high-performance cooling for 24/7 reliability
Perfect For:
- AI / Machine Learning inference on 70B-class models
- Multi-user LLM serving for teams
- Virtual Desktop Infrastructure (VDI)
- Rendering and 3D graphics
- High-density virtualization
- Scientific computing & HPC
This server has been fully tested and is ready to deploy. Ships with all necessary power cables, rails, and bezel (where applicable).
Note: This is a refurbished unit in excellent cosmetic and functional condition. Minor signs of previous rack use may be present.
Compare Our R740xd Builds:
All of these builds share the same GPU-ready platform (24SFF chassis, 128GB RAM, Intel X710/I350 networking, iDRAC Enterprise, GPU enablement kit, high-performance fans, 2x 1600W supplies) — they differ in CPU, storage controller and GPU:
- R740xd – GPU Ready (No GPU Included) – $2,995 – 2x Silver 4210R, HBA330 pass-through. The lowest-cost GPU-ready host; bring your own card.
- R740xd – GPU Ready, Gold 6248 / H740p – $3,493.33 – 2x Gold 6248 (40 cores), PERC H740p RAID. More CPU and hardware RAID; bring your own card.
- R740xd – Dual T4 Build – $4,693 – Silver 4210R platform + 2x Tesla T4 16GB (32GB total), slot-powered.
- R740xd – Dual V100 Build – $5,849 – Silver 4210R platform + 2x Tesla V100 32GB (64GB total). Best value for multi-GPU training and parallel workloads.
- R740xd – A40 Build – $12,159 – Silver 4210R platform + 1x NVIDIA A40 48GB. Best single-GPU performance for AI inference and visualization.
- R740xd – Dual A40 Build (this listing) – $18,920 – Silver 4210R platform + 2x NVIDIA A40 48GB (96GB total). The most GPU memory in the lineup, for 70B-class models with headroom or two workloads side by side.
Need a custom configuration, more storage, or a third GPU? Message us — we can build to your requirements.
What This Server Runs
Two A40s give you 96GB of GDDR6. Inference frameworks such as vLLM and llama.cpp split a model across both cards (tensor or pipeline parallelism), or you can run a separate model on each card. Measured on a single A40 at 4-bit quantization, single-stream — each card in this server delivers the same:
- 14B models (Qwen2.5 14B) — ~9GB VRAM, ~44 tokens/sec
- 27–34B models (Gemma2 27B, Qwen2.5 32B, QwQ 32B, Llava 34B) — 16–20GB VRAM, 23–29 tokens/sec
- 70B models (Llama 3.1 / 3.3 70B) — 39–43GB VRAM, 12–13.5 tokens/sec on one card
What the second card changes:
- 70B at 4-bit with real context headroom — on a single A40 a 70B model leaves little room for KV cache. Split across two cards, roughly 50GB is left for long context and concurrent requests.
- 70B at 8-bit — about 70–75GB of weights. Does not fit on one A40; fits across two.
- Two models at once — for example a 70B model on one card and a 27–34B coding or embedding model on the other.
- Double the serving capacity — one A40 handles 100 concurrent requests on an 8B model at ~1,705 tokens/sec of aggregate output under vLLM. Two cards running independent instances roughly double that.
The two cards communicate over PCIe, which is how most multi-GPU inference deployments run. Full single-card benchmark table, KV cache math and the FP8/MIG limitations: What can you actually run on a 48GB NVIDIA A40?
Configuration Requirements — Already Handled
The A40 is a 300W passively cooled card with no fan of its own, and this server runs two of them. Dropping them into a stock R740xd does not work — Dell's installation guidance requires a specific set of parts, all of which are installed and tested in this build:
- GPU enablement kit (risers, GPU power cables, GPU air shroud, mylar insulation) — Dell 330-BBMO / 85CYN in the US, 490-BEIX / JKFGX in EMEA
- Low-profile heat sinks in both CPU sockets — standard heat sinks physically foul the GPU air shroud, and Dell requires both sockets populated
- High-performance fan kit — mandatory for any GPU configuration in this chassis, not optional
- A dedicated GPU power cable for each card — required for any card above 75W; each A40 draws 300W, and both are cabled and powered
- Identical GPUs — Dell requires all installed GPUs to match in type and model; both cards are NVIDIA A40 48GB
- 1600W power supplies — two 300W cards plus dual processors put this build well above Dell's 1100W floor for GPU configurations; 1600W supplies are the right size for a two-card load
- 30°C maximum inlet temperature — the thermal ceiling once GPUs are installed alongside 150W/8C, 165W/12C, 200W or 205W processors
Assembled from separate secondary-market purchases, that is a stack of line items, condition claims and return windows. Here it is one part number, one warranty and one place to call. The full breakdown, including the 330-BBMO vs 490-BEIX part-number confusion and the two mistakes that stop most self-assembled builds: Putting an NVIDIA A40 in a PowerEdge R740xd
Frequently Asked Questions
What is the largest model this server can run?
With 96GB across two A40s, a 70B parameter model at 8-bit quantization (about 70–75GB of weights) fits, as does a 70B model at 4-bit with roughly 50GB left over for long context and concurrent users. A single A40 tops out at 70B at 4-bit with little headroom.
How do the two GPUs work together?
Over PCIe. Frameworks like vLLM and llama.cpp split models across both cards, or you can run a separate model on each card.
Does the NVIDIA A40 support FP8 or MIG?
Neither. The A40 is an Ampere GA102 card — FP8 tensor support began with Ada Lovelace and Hopper, and Multi-Instance GPU partitioning is limited to A100 and H100. The A40 does support SR-IOV for virtualization.
Can I add a third GPU?
The R740xd supports up to three double-width accelerators, and Dell requires all installed GPUs to be identical in type and model. Contact us before adding a third A40, as the thermal, power and riser assignments change and we can build it to order.
Does it come with an operating system?
No OS is included. The server supports Windows Server, Linux distributions, VMware ESXi and Proxmox. iDRAC9 Enterprise is licensed for remote installation and management.
Can I expand the storage?
Yes. The chassis has 24 x 2.5" SFF bays and ships with one 960GB SATA SSD, leaving 23 bays open. The PERC HBA330 runs in pass-through mode, which suits ZFS, Ceph and software-defined storage. Tell us your capacity target and we will quote the drives with the server.
What condition is it in and what warranty applies?
Professionally refurbished, fully tested and burned in, in excellent cosmetic and functional condition with minor signs of previous rack use possible. Backed by the Resilient Tec lifetime warranty, reachable directly rather than through a manufacturer ticket queue.
What networking is included?
An Intel X710 dual-port 10Gb SFP+ rNDC plus an Intel I350 dual-port 1Gb Base-T, with a dedicated iDRAC management port. We can supply matching 10G SFP+ DAC cables or transceivers with the server.
Further Reading
- Putting an NVIDIA A40 in a PowerEdge R740xd — every part the card needs, the enablement kit part numbers, and why a stock chassis throttles it
- What can you actually run on a 48GB NVIDIA A40? — benchmarked model sizes, tokens per second, and vLLM concurrency limits
- Why your GPU won't work in your server — the eight installation problems and the fix for each
- Custom Dell R740xd GPU servers: build options, prices, and the forum questions nobody answers — which GPU power cable (4VPD3 vs TR5TP), fans at 100%, LFF chassis limits, and the builds compared
- The cloud sent you a bill — on-prem inference versus rented GPU time
- GPU servers & on-prem AI guide, GPU & AI infrastructure guide and enterprise servers guide