Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
The documented system is not a completed dual-RTX 5090 machine. It is a working dual-RTX 4090 prototype built around a dual-socket AMD EPYC server platform, with a planned upgrade to two RTX 5090 cards. That distinction matters: the hardware appears unusually well suited to multi-GPU experimentation, but there is no published evidence here of completed dual-5090 testing, benchmark results, thermals, or application compatibility.
What the prototype is
The build was documented on May 30, 2024, in an [H]ard|Forum build thread titled “Ready for Dual 5090s, functional prototype on dual 4090s.” At that time, the RTX 5090 was a future upgrade target. The published working configuration used two NVIDIA RTX 4090 Founders Edition cards.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
VIPERA NVIDIA GeForce RTX 4090 Founders Edition Graphic Card | $4,440.00 | Buy on Amazon |
| 2 |
|
ASUS TUF Gaming NVIDIA GeForce RTX 4090 OC Edition Gaming Graphics Card (24GB GDDR6X, PCIe 4.0, HDMI... | $4,039.95 | Buy on Amazon |
Since then, the RTX 5090 has become a shipping product. NVIDIA lists 21,760 CUDA cores, 32GB of GDDR7 memory, a 512-bit memory interface, PCIe 5.0, a 600W power specification, and no NVLink or SLI support. NVIDIA announced availability on January 30, 2025, with a U.S. starting price of $1,999. That product information does not change what the original build log demonstrated: dual 4090 operation, not dual 5090 operation.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →The published hardware configuration
| Component | Configuration | Evidence |
|---|---|---|
| GPUs | Two NVIDIA RTX 4090 Founders Edition cards | Builder-reported |
| Processors | Two AMD EPYC 7773X CPUs | Builder-reported; specifications from AMD |
| Motherboard | Gigabyte MZ72-HB0 dual-socket SP3 board | Builder-reported; specifications from Gigabyte |
| Memory | 1TB ECC DDR4 LRDIMM | Builder-reported |
| Power supply | 1600W or higher digital PSU | Builder-reported; exact model not documented |
| Primary storage | 8TB Sabrent Rocket 4 Plus | Builder-reported |
| Data storage | Two Micron 9300 Max 15.4TB drives in a RAID 0 array | Builder-reported |
| Backup storage | Two 8TB Micron 5300 drives | Builder-reported |
| Operating systems | Windows Server 2022 Datacenter and Ubuntu | Builder-reported |
| Display | ASUS PA32UCG-K | Builder-reported |
The builder described the storage arrangement as a 77TB RAID 0 array, overprovisioned to 64TB. The available post does not document its filesystem, stripe configuration, RAID implementation, benchmark software, or independent verification. RAID 0 can improve throughput, but it provides no redundancy: a failed drive can make the entire array unavailable.
#1 Best Overall
- 16.384 NVIDIA CUDA Core
- Supports 4K 120Hz HDR, 8K 60Hz HDR and Variable Refresh Rate as specified in HDMI 2.1a
- New Flow Multiprocessors: Up to 2x performance and power efficiency
- Fourth Generation Tensor Cores: up to 2x AI performance
- Third Generation RT Cores: Up to 2x ray tracing performance
Why use two EPYC 7773X processors?
AMD specifies each EPYC 7773X as a 64-core, 128-thread processor with 768MB of L3 cache, eight memory channels, PCIe 4.0 connectivity with 128 lanes, a 280W default TDP, and support for one- or two-socket systems. Two processors therefore provide 128 cores and 256 threads, plus a nominal 1.536GB of aggregate L3 cache.
That is a strong foundation for highly parallel work, virtualization, data processing, rendering, AI experimentation, and multiple simultaneous workloads. It is not automatically a good gaming choice. Games typically benefit more from strong per-core performance, low latency, and predictable boost behavior than from server-class core counts.
A dual-socket system also introduces NUMA behavior. Each processor has its own attached memory and PCIe-connected devices. An application may perform better when a CPU thread, system memory allocation, and GPU are kept close to the same socket. Poor placement can add latency and reduce effective throughput. Software and workload design matter more than the raw core count suggests.
Can the MZ72-HB0 run two RTX 5090s?
It appears to have the slot resources, but that is not the same as official RTX 5090 certification. Gigabyte’s MZ72-HB0 provides five PCIe Gen4 expansion slots: three physical x16 slots wired at Gen4 x16 and two physical x16 slots wired at Gen4 x8, with slots distributed across the two CPU sockets. The board therefore offers a plausible electrical foundation for two GPUs.
There are several separate questions to answer:
- Which exact slots did the two 4090s use, and which CPU owns each slot?
- Do both cards receive the intended link width?
- Will two particular RTX 5090 models physically fit without blocking each other?
- Can the chassis provide adequate intake and exhaust airflow?
- Are nearby connectors, memory slots, storage devices, or expansion cards obstructed?
- Does the board firmware enumerate both cards correctly?
- Can the power supply handle two 600W-class GPUs, two 280W CPUs, memory, storage, and transient loads?
The RTX 5090 supports PCIe 5.0, while the MZ72-HB0 is a PCIe 4.0 platform. That is a generation mismatch, not automatically a fatal incompatibility. A Gen5 GPU may operate on a Gen4 slot, but the actual link behavior must be verified on the specific board, firmware, slot, driver, and application.
Card dimensions are equally important. NVIDIA lists its RTX 5090 Founders Edition as a two-slot, 304mm-long design, but partner cards can be considerably thicker or longer. A pair of large aftermarket cards may not fit the same way as the original Founders Edition 4090 arrangement.
What changes from two RTX 4090s to two RTX 5090s?
One RTX 5090 has 32GB of GDDR7 memory and a 600W power specification. Two cards provide two separate 32GB memory pools. They do not automatically become one addressable 64GB pool.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Whether both GPUs help depends on the software. A workload may split batches, models, frames, jobs, or rendering tiles across devices. Another application may use only one GPU. Some frameworks support model or data parallelism; others duplicate data in each card’s memory. Communication and synchronization over PCIe can also reduce scaling efficiency.
NVIDIA lists the RTX 5090 as not NVLink or SLI-ready. That is especially relevant to gaming and tightly coupled workloads that would otherwise benefit from a fast GPU-to-GPU interconnect.
Gaming: two installed GPUs are not two GPUs in every game
For ordinary gaming, this is usually an impractical platform. Modern games do not guarantee useful scaling across two consumer GPUs, and the absence of SLI support makes the old “add a second card for double the frame rate” expectation even less realistic.
A second GPU can still be useful for:
- Running a compute workload while another GPU handles interactive graphics.
- Driving separate GPU-intensive applications simultaneously.
- Professional renderers or other software with explicit multi-GPU support.
- Independent jobs that can run concurrently without sharing a single workload.
But the dual EPYC platform adds cost, power draw, physical size, and possible NUMA latency without necessarily improving game responsiveness. A modern single-socket gaming or workstation system with one RTX 5090 will usually be simpler and easier to cool.
AI, rendering, and parallel workloads are the stronger use cases
Two RTX 5090s could make sense for parallel image-generation jobs, batch inference, separate model instances, GPU rendering, CUDA workloads, and local-model experimentation. The most obvious benefit may be throughput: two independent jobs can run at the same time even when one large job cannot scale efficiently across both cards.
Large-model inference is more complicated. If a model exceeds one card’s 32GB capacity, the framework must explicitly shard or distribute it. That can work, but it introduces synchronization, memory-transfer, and configuration overhead. Installing two cards does not make an oversized model run automatically.
NVIDIA’s Blackwell materials promote FP4 support and large generative-AI gains compared with the RTX 4090. Those are NVIDIA claims and should not be treated as independent benchmark results. Real performance depends on the framework, precision, model, batch size, quantization, driver, PCIe placement, and whether the workload uses one GPU or both.
Rank #2
- NVIDIA Ada Lovelace Streaming Multiprocessors: Up to 2x performance and energy efficiency
- Tensor Cores of the 4th Generation: up to 2x AI performance
- RT-cores of the 3rd Generation: up to 2x raytracing performance
- OC mode: Boost clock 2595 MHz (OC mode) / 2565 MHz (gaming mode)
- Axial Tech fans deliver up to 23% higher airflow
Power and cooling are the hard parts
The builder reported GPU temperatures below 48°C under load and benchmarks, CPU temperatures below 46°C under load, and approximately 24°C CPU and 31°C GPU temperatures during normal operation. The post also described an open-design case arrangement and unusual card orientation.
Those figures must remain attributed to the builder. They are not a reliable prediction for two RTX 5090s in a conventional enclosed case. Temperature results depend on ambient temperature, load duration, power limits, fan curves, card spacing, airflow, and whether the cards were undervolted.
Two 600W-class GPUs combined with two 280W CPUs already create a substantial heat and power envelope before adding memory, storage, fans, and motherboard consumption. “1600W” alone is not proof that a PSU is suitable. The exact model, connector design, rail behavior, transient response, cable arrangement, and continuous capacity all matter.
Power connectors also need careful handling. Use the manufacturer-recommended cables, ensure every connector is fully inserted, provide the required independent connections, and maintain the specified bend clearance near the plug. Do not sharply bend a high-power GPU cable immediately at the connector.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Validation checklist for a dual-5090 conversion
The following is a practical validation plan, not a record of steps confirmed by the original builder.
Free tools Windows power users keep installed
One-click scans. No signup required.
Hardware checks
- Record the exact RTX 5090 model, dimensions, thickness, connector location, and power requirement.
- Confirm physical spacing for both cards, including fan intakes and adjacent connectors.
- Confirm that each card is fully seated and attached securely against its weight.
- Use a PSU with documented support for the required high-power GPU cabling and sufficient transient headroom.
- Inspect the connector insertion and cable bend radius before applying load.
- Update the motherboard firmware and verify relevant PCIe settings.
- Confirm that the cooling arrangement provides active airflow to both cards.
Software checks
- Confirm both GPUs appear in firmware.
- Confirm both devices are enumerated by the operating system.
- Record the operating-system version, driver version, and GPU firmware where available.
- Test each card separately before enabling a multi-GPU workload.
- Run a sustained workload while logging temperature, power, clocks, and throttling.
- Test the intended application rather than relying only on a synthetic benchmark.
- Check whether the application uses independent jobs, data parallelism, model sharding, peer-to-peer transfers, or only one GPU.
- Measure PCIe link width and CPU/GPU affinity where the workload is NUMA-sensitive.
Likely failure modes
The second GPU is not detected
Possible causes include an incorrect slot, a card that is not fully seated, inadequate power, firmware configuration, physical interference, resource allocation, or a driver problem. Test each card individually, try the second card in a known-good slot, inspect connectors, check firmware settings, and clean-install the driver only after the hardware has been isolated.
The system shuts down under load
Suspect PSU overload or transient response, a loose GPU connector, thermal protection, motherboard power limits, or unstable tuning. Return the system to stock settings, test one GPU at a time, log power, inspect every cable, and do not assume a nominal wattage rating guarantees safe operation.
Both GPUs work but performance does not scale
The application may support only one GPU, duplicate VRAM rather than pool it, incur too much synchronization overhead, or encounter CPU, PCIe, or NUMA bottlenecks. Compare separate processes, data-parallel modes, and model-parallel modes. The best result may be two independent jobs rather than one faster job.
Temperatures are much higher than the prototype’s results
An enclosed chassis, thicker cards, warmer ambient air, higher RTX 5090 power draw, poor intake spacing, or different fan behavior can explain the difference. Measure core, hotspot, and memory temperatures; improve airflow; increase spacing where possible; and consider a lower power limit or undervolt.
What the original build gets right—and what it does not prove
The platform is unusually capable. It combines large memory capacity, ECC support, abundant PCIe connectivity, enormous CPU parallelism, and substantial local storage. Those characteristics make it more suitable for a workstation/server hybrid than a conventional desktop.
It does not prove that:
- Two RTX 5090s were installed or benchmarked in the system.
- Two RTX 5090s provide a unified 64GB VRAM pool.
- Gaming performance doubles.
- The MZ72-HB0 officially supports RTX 5090 cards.
- The reported temperatures apply to a normal enclosed case.
- The RAID array delivers its reported throughput in every workload.
- A 1600W PSU is sufficient without knowing its exact model and measured behavior.
Who should consider this kind of system?
This design is defensible for someone who specifically needs several of the following: large ECC memory capacity, massive CPU parallelism, many PCIe devices, independent GPU jobs, local AI experimentation, GPU rendering, virtualization, or a highly configurable server-workstation platform.
It is a poor fit for someone whose main goal is gaming, quiet operation, low power use, simple maintenance, or the best performance per dollar. A single RTX 5090 workstation is easier to cool and configure. A modern single-socket workstation may provide newer PCIe and memory technology with less NUMA complexity. Cloud GPUs avoid the upfront purchase and power infrastructure but introduce recurring cost, data-transfer concerns, and less control.
Verdict
The dual-5090 concept is technically plausible, but the published evidence stops at a functional dual-4090 prototype and a planned upgrade path. The Gigabyte MZ72-HB0 has enough PCIe resources to make two GPUs conceivable, yet physical clearance, PCIe generation, firmware, power delivery, cooling, drivers, and application support all require validation.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →For gaming, two RTX 5090s would usually be an expensive and inefficient approach. For AI, rendering, and independent parallel workloads, the idea is much more compelling—but only when the software can use both cards effectively. The correct description is therefore not “a dual-RTX 5090 build,” but “a dual-RTX 4090 system designed as the basis for a possible dual-RTX 5090 upgrade.”
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
