Almost any modern GPU can be passed through to a Proxmox VM once. The hard part is the second time: stopping the VM, starting it again, and finding the card still answers. That is why the best GPU for Proxmox passthrough is not the fastest card you can afford. It’s the one that resets cleanly, behaves with the guest driver, fits your case, and, if you want it, can be shared between several VMs.
We haven’t run every card here in our own lab, so the passthrough notes come from vendor documentation, the Proxmox forum and Level1Techs, linked as we go. Here are seven picks for different jobs.
Quick picks
| Job | Card | Price | Power |
|---|---|---|---|
| Best overall passthrough VM | ASUS Dual RTX 5060 Ti 16GB | About $794 | 180W, 8-pin |
| Best low-profile | PNY RTX 2000 Ada 16GB | About $770 | 70W, slot only |
| Best for sharing one GPU (SR-IOV) | GUNNIR Arc Pro B60 TF 24GB | About $1,050 | 200W, 2x 8-pin |
| Cheapest clean reset | Sparkle Arc A380 ELF 6GB | About $170 | Slot only |
| Budget low-profile | PNY RTX A2000 12GB | About $640 | 70W, slot only |
| Single-slot, slot-powered | PNY T1000 8GB | About $340 | 50W, slot only |
| Most VRAM (best in an LXC) | ASRock Radeon AI PRO R9700 32GB | About $1,950, resellers only | 300W, 16-pin |
Prices are well above launch for most of these, because the 2026 memory shortage pushed GPU prices up across the board. Treat any MSRP you remember as history.
What makes a GPU good for passthrough
It resets after the VM stops
This is the trait that rules out most of AMD’s current lineup. The community vendor-reset module only supports Polaris, Vega and the RX 5000 series, so it does nothing for RX 6000, 7000 or 9000 cards. The author of that module wrote that RX 7000 cards do not reset, and an RX 9070 XT owner on the Proxmox forum still needs a pre-stop hookscript to hand the card back without hanging the host. One forum regular says most RX 9070 cards work on Proxmox kernel 7.0 and later, though that is second-hand, so the honest summary is “unreliable”. If you stop and start VMs often, plan on NVIDIA or Intel.
The guest driver does not fight you
NVIDIA ended the old Code 43 problem in 2021, when its R465 driver started officially supporting GeForce passthrough, so no hidden-KVM tricks are needed now. RTX 5060 series cards are the exception to know about: early units showed no VM BIOS screen until NVIDIA’s UEFI firmware update tool was applied, as covered in this Proxmox forum thread.
Turn on Above 4G Decoding and Resizable BAR in the host firmware for Intel Arc, which Intel recommends, and the VM needs a large 64-bit MMIO window or the guest sees a reduced BAR. Windows guests on Arc A-series cards also need a CPUID line in the VM config, which this tested Proxmox guide documents along with the correct driver INF.
It can be shared, if you need that
- Intel Arc Pro B-series supports SR-IOV with no license. Intel lists the Arc Pro B50 at 2 virtual functions, and a Level1Techs forum user saw a B60 expose 7 after a February 2026 firmware update. The setup guide we follow targets Proxmox VE 9.2 with kernel 7.0. Linux guests need kernel 6.18 or later with the xe driver, and Windows guests need the current Arc Pro driver.
- Consumer Arc cards, including the B580, have no SR-IOV. Intel’s own virtualization table marks them unsupported.
- NVIDIA vGPU works on Proxmox but needs a paid NVIDIA vGPU entitlement and a data-center or qualifying workstation card, per the Proxmox wiki. None of the GeForce or entry pro cards here qualify.
- AMD MxGPU is limited to Instinct accelerators and the Radeon PRO V710, so it is not a homelab option.
If you only want the GPU for containers, an LXC sidesteps all of this. The host keeps the driver, the container gets the device nodes, and nothing ever needs to reset.
It fits the box and its power supply
Small workstations and prebuilt SFF desktops often have a low-profile slot and no PCIe power cable. That’s where 50W and 70W cards that draw everything from the slot earn their price. Check lane width too: several cards here are PCIe x8, which is fine in a bifurcated slot. Our PCIe lanes guide covers the math.
1. ASUS Dual RTX 5060 Ti 16GB
For a single VM that gets the whole card, this is the one we’d buy. You get 16GB of VRAM, a current NVENC block with AV1 encode, and NVIDIA’s generally predictable passthrough behavior. NVIDIA’s encode matrix lists 12 concurrent encode sessions for GeForce RTX 50 cards, more than most homelab media servers need. It is a PCIe x8 card, so it loses little in a split slot.
Before you install it: apply NVIDIA’s RTX 5060 series UEFI firmware update, or the VM may not show a boot screen. Some Blackwell owners on the same forum thread still report host freezes when a VM powers off, so test a few stop and start cycles before you depend on it. It needs an 8-pin cable and a full-height case. Our VRAM guide shows what 16GB runs for local LLMs. About $794, sold by Amazon.
2. PNY RTX 2000 Ada 16GB
This is the best card for a small box. It is low-profile, draws 70W entirely from the slot, and still has 16GB of ECC memory and AV1 encode. NVIDIA lists its encoder as unrestricted, with no session cap. Pro cards like this one tend to be the least dramatic passthrough path, and it ships with both bracket heights.
The fit catch: it is a dual-slot card. In a Minisforum MS-01, you’d need the single-slot RTX 2000E variant, which is hard to find in the US. For other mini workstations, check our MS-03, MS-01 and MS-A2 comparison for slot widths. About $770, and in stock when we checked.
3. GUNNIR Arc Pro B60 TF 24GB
If you want one GPU split across several VMs without a license, this is the card. The B60 supports SR-IOV, and a Level1Techs forum user saw it expose 7 virtual functions after the February 2026 firmware, though fewer have been reported in actual use. It also has 24GB of memory, which makes it the cheapest new, current-generation 24GB card we found on Amazon US. The ASRock Creator we first linked sold out, so the link now goes to GUNNIR’s version of the same GPU at about $1,050 from a third-party seller.
Know the limits: the 24GB is shared across every virtual function, so each VM gets a slice. SR-IOV wants a recent Proxmox kernel, a recent guest driver, the latest card firmware and a motherboard with enough MMIO space. Treat it as a capable homelab card rather than a proven production platform. It’s a 2-slot blower with two 8-pin connectors.
4. Sparkle Arc A380 ELF 6GB
The cheapest card here that should reset cleanly. The tested guide linked above found a working function-level reset on the Arc A310, which uses the same DG2 chip as the A380, so a Proxmox VM can stop and start without hookscripts. It has AV1 encode, no external power connector, and at about $170 it is currently cheaper than any A310 on Amazon US, which flips the usual “buy the A310” advice.
Setup notes: turn on Resizable BAR and Above 4G Decoding, and give Windows guests the CPUID line. It is full-height and dual-slot, so it will not fit low-profile boxes. Sold by Amazon.
5. PNY RTX A2000 12GB
The A2000 has been the classic homelab SFF card for years, and it’s still a solid choice when the RTX 2000 Ada is out of reach. It’s low-profile, slot-powered at 70W, and has 12GB of ECC memory and an unrestricted encoder.
What you give up: it is an Ampere card without AV1 encode, and it is dual-slot like the Ada. Listings move fast, with stock at just a few units when we checked. About $640.
6. PNY T1000 8GB
When the slot is single-width and low-profile, the T1000 is the card that fits. It draws 50W from the slot, has 8GB of memory and passes through without quirks. It is the cheapest 8GB card in this form factor.
Trade-offs: it’s a Turing card with no AV1 encode, and NVIDIA’s matrix caps it at 8 encode sessions. The RTX A400 is a newer single-slot alternative at about $245, but it only has 4GB. The T1000 was about $340.
7. ASRock Radeon AI PRO R9700 Creator 32GB
This pick comes with an asterisk. The ASRock listing sold out in October 2026, and the R9700 stock left on Amazon US is a handful of reseller units at about $1,950, so this pick has no buy link until a normal listing returns. For 32GB of memory at that price, nothing else here comes close, which matters for larger local models. But the R9700 uses the same RDNA4 chip as the RX 9070 XT, vendor-reset doesn’t support it, and we found no reports confirming it resets cleanly in a VM.
How to run it: put it in an LXC, where the host keeps the driver and ROCm, or in a VM that boots once and stays up. Don’t build a workflow around restarting a VM that owns it. It is a 300W card with a 16-pin connector. For LLM-focused GPU choices, see our best GPU for local LLMs guide.
Cards we left out, and why
- Intel Arc Pro B50. It is the low-profile SR-IOV card on paper, but the only Amazon US listings are resellers asking about $629, roughly 1.8 times its launch price, with almost no reviews. If you find it near $350 elsewhere, it is a good small-box SR-IOV card with 2 virtual functions.
- Radeon RX 9060 XT. At about $660 it’s cheaper than the RTX 5060 Ti 16GB, and fine in an LXC. In a VM you restart, the RDNA4 reset problem applies.
- Arc B580. Cheap and capable, but consumer Battlemage has no SR-IOV, and the A380 covers budget media work.
- RTX 3090. The 24GB used-market favorite now sells renewed for about $1,800 with weak seller ratings. The B60 gives you 24GB for a little over half of that.
- RTX 5090 and RTX PRO 2000 Blackwell. The 5090 is listed from about $6,650 to over $7,000, and the low-profile RTX PRO 2000 is more than double the RTX 2000 Ada for the same 16GB and 70W.
Still planning the host itself? Our Proxmox server build guide covers IOMMU groups and platform choice, and the AMD vs Intel homelab comparison covers the iGPU side.
Which GPU for which passthrough job
| If you want | Buy | Why |
|---|---|---|
| One fast VM with the whole card | RTX 5060 Ti 16GB | Clean NVIDIA passthrough, 16GB, AV1, 12 encode sessions |
| A GPU in a low-profile box | RTX 2000 Ada 16GB | 70W slot power, 16GB, unrestricted encoder |
| One GPU shared by several VMs | Arc Pro B60 24GB | SR-IOV with no license, up to 7 VFs exposed |
| The cheapest card that resets | Arc A380 ELF | Same DG2 chip as the A310 with working FLR, AV1, slot power |
| A single-slot, slot-powered card | T1000 8GB | 50W, fits single-width low-profile slots |
| The most VRAM for LLMs | Radeon AI PRO R9700 32GB | 32GB, but run it in an LXC |
| Frequent VM restarts on AMD | Nothing current | RX 7000 and 9000 reset unreliably, and RX 6000 is hit or miss |
