Free tools Windows power users keep installed
One-click scans. No signup required.
In a single benchmark reported by BuildZn on September 24, 2026, an RTX 4090 running Llama 3 8B Instruct Q4_K_M produced 115.2 tokens per second in a virtio-nvgpu KVM guest, compared with 120.5 tokens per second bare metal. That is 95.6% of the bare-metal throughput in that test—not a general performance guarantee for other GPUs, models, or configurations.
What the benchmark measured
BuildZn author Umair Bilal compared bare-metal inference, PCI passthrough, and a virtio-nvgpu KVM guest. The article reports averages for throughput and first-token latency:
As an Amazon Associate I earn from qualifying purchases.
| Configuration | Average throughput | Average first-token latency |
|---|---|---|
| Bare metal | 120.5 tokens/s | 150 ms |
| PCI passthrough | 119.8 tokens/s | 155 ms |
| virtio-nvgpu KVM guest | 115.2 tokens/s | 168 ms |
These are figures reported in the BuildZn article, not independently reproduced measurements. Its 95.6% result is the ratio of the guest average (115.2 tokens/s) to the bare-metal average (120.5 tokens/s). In the same comparison, the guest trailed bare metal by 5.3 tokens/s and PCI passthrough by 4.6 tokens/s. Throughput measures generated tokens over time; first-token latency measures the wait before generation begins, so the two figures describe different aspects of responsiveness.
How to interpret the 95% figure
The result supports a narrow statement: in the workload and setup described by BuildZn, the virtio-nvgpu guest delivered 95.6% of the bare-metal average throughput. It does not establish that virtio-nvgpu generally runs LLMs at 95% of bare-metal speed. GPU model, driver and kernel compatibility, QEMU revision, model and quantization, and workload can all differ between deployments.
The project architecture documentation describes GPU-rendering performance in the 95–99% bare-metal range. That is a separate workload and does not independently confirm the LLM inference result. The project describes its goal as “Near-native NVIDIA GPU access inside a KVM guest,” but that description should not be read as a measured guarantee for every workload.
What hardware and software were used
According to the September 24, 2026 BuildZn article, the host used an AMD Ryzen 9 7950X, 64 GB of DDR5 memory, an NVIDIA GeForce RTX 4090, and Ubuntu 22.04. The guest was configured with 16 vCPUs, 32 GB of RAM, Ubuntu 22.04, and a custom Linux 6.8 kernel with a virtio-nvgpu driver. The article also reports using a custom QEMU build and NVIDIA user-mode libraries inside the guest.
The inference workload was Llama 3 8B Instruct quantized to Q4_K_M, run via Ollama/llama.cpp. The article says it sent 100 requests, each with an approximately 100-token input and a target generation of approximately 200 output tokens. It describes measuring tokens per second and first-token latency.
What the comparison says about passthrough
In this reported test, PCI passthrough was close to bare metal on average throughput: 119.8 versus 120.5 tokens/s. The virtio-nvgpu guest reached 115.2 tokens/s. First-token latency was 155 ms with passthrough and 168 ms with virtio-nvgpu, compared with 150 ms bare metal. These results describe the tested configuration only; they do not establish which approach will be faster or simpler for a different host, GPU, or workload.
Rank #3
- System Compatibility Note: 2-slot card, 271x112x39mm, single 8-pin power, 200W TDP. Verify chassis clearance and PSU capacity before purchase.
- Dedicated Support: Please contact us directly through Amazon for any product questions or assistance you may require.
- 24GB GDDR6 on 192-Bit Bus: Massive 24GB memory with 456 GB/s bandwidth – ideal for LLMs, AI inference, 3D rendering, and generative design.
- Intel Xe2-HPG Architecture: Built on Intel's next-gen architecture with 20 Xe cores and 160 XMX engines for AI acceleration (197 INT8 TOPS).
- PCIe 5.0 Support: PCI Express 5.0 x16 interface for maximum bandwidth with the latest workstation platforms.
How virtio-nvgpu works—and what remains uncertain
The project architecture document describes virtio-nvgpu as a virtio device and guest kernel driver that forward NVIDIA kernel-driver ioctls between a KVM guest and its host. The stated rationale is to avoid API-level translation while allowing the guest to use NVIDIA user-mode drivers. This architecture is distinct from the benchmark outcome: the document’s rendering discussion is not a second LLM benchmark.
The BuildZn article characterizes its setup as experimental and reports custom QEMU and patched guest-kernel requirements. The available information does not establish an official production-support status or a complete, current compatibility matrix covering GPUs, NVIDIA drivers, QEMU revisions, and kernel revisions. Check current project documentation for the specific versions and hardware you intend to use.
Rank #4
- ATTENTION BUYERS: EXPORTING this item outside the U.S. requires strict compliance with the Bureau of Industry and Security (BIS) U.S. Export Administration Regulations (EAR). Non-compliance carries fines up to $1 MILLION PER VIOLATION AND UP TO 20 YEARS IMPRISONMENT for both the buyer and seller. To uphold our regulatory obligations, we strictly follow BIS "Know Your Customer" guidance by screening each order and immediately aborting transactions showing unresolved regulatory red flags. ***
- [ Local Generative AI & Data Science ] Supercharge your local generative AI pipelines with the RTX PRO 6000 96GB Max-Q. Engineered for agentic workflows and advanced data science, this dual-slot workstation monster packs a massive 96GB of next-gen GDDR7 ECC memory. Eliminate VRAM bottlenecks and host heavy open-source LLMs locally with a 512-bit bus width. It fits seamlessly into standard desktops, enabling teams to scale intelligence rapidly without cloud latency or subscription overhead.
- [ High-Performance Computing (HPC) ] Accelerate your high-performance computing tasks using the cutting-edge NVD Blackwell architecture. This efficient Max-Q Edition delivers up to 3,511 TOPS of AI performance while capping maximum power consumption at a modest 300 Watts. Run complex simulations, massive deep learning datasets, and parallel enterprise tasks effortlessly. This card offers unmatched thermal management, making multi-GPU desktop scaling perfectly viable for demanding research labs.
- [ AI-Driven Rendering & Visuals ] Transform raw data into cinematic reality with ai-driven rendering graphics. Powered by 24,064 CUDA cores and 188 4th-generation Ray Tracing cores, the RTX PRO 6000 Max-Q handles real-time 3D environments and complex visual effects with ease. Experience flawless path tracing and hardware-accelerated rendering. It bridges the gap between intense calculation and breathtaking visual output, maximizing creativity for studios and independent developers alike.
- [Reliability Guaranteed] Whether you are investing in high-speed DDR5 RAM or a powerhouse GPU, we protect your build against defects and performance failures. We stand firmly behind the quality of our hardware, ensuring that your setup remains fast, stable, and secure for years to come. 3YR Manufacturer Warranty. NOTE: OEM/Bulk Packaging. Retail packaging is NOT included.
How strong is the evidence?
The BuildZn report provides averages but not raw run-by-run observations, distributions, confidence intervals, or a public benchmark harness. It also does not provide enough detail to independently reproduce its exact token accounting. The figures are therefore useful as a reported result for one setup, but they do not show how much performance may vary between runs or whether another system will achieve similar results.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteThe article separately claims a 30% reduction in “operational latency” for a multi-tenant workflow, attributing it to faster VM instantiation and teardown. It does not clearly define that metric or provide the measurement method or raw data in the reported text, so the percentage cannot be treated as independently verified or directly compared with the inference latencies in the table.
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.
