For RTX 4090 NVLink support, the short answer is no: NVIDIA lists NVLink (SLI-Ready) as unsupported for the GeForce RTX 4090. The Nvidia GeForce RTX 4090 is the flagship GPU of Nvidia's Ada Lovelace architecture. The RTX 5090 has since replaced it at the top of the GeForce line, and the 4090's 24 GB of VRAM fits many common inference and fine-tuning jobs.
This guide covers everything you need to know: full specs, VRAM capacity and bandwidth, current pricing, how it compares to the RTX 5090 and data-center GPUs like the H100, and which GPUs on Runpod offer 24 GB of VRAM or more.
Does the RTX 4090 support NVLink?
NVIDIA removed NVLink from the GeForce line with the Ada Lovelace generation. The RTX 3090 and 3090 Ti supported NVLink bridges; the RTX 4090 does not, and neither does the RTX 5090.
What this means in practice:
- No memory pooling. Two 4090s are two separate 24GB GPUs, not one 48GB GPU. A model needing 40GB will not fit across them without sharding.
- Multi-GPU still works over PCIe for data-parallel training, where each GPU holds a full copy of the model and only gradients are exchanged. That is the common case and it is fine.
- Model-parallel workloads suffer. Splitting a single model across cards means heavy inter-GPU traffic over PCIe rather than NVLink, and throughput drops noticeably.
If you need NVLink-class interconnect, the SXM data center cards are where it lives – H100 SXM at $4.39/hr, H200 at $5.29/hr, A100 SXM at $2.19/hr. For data-parallel jobs that fit in 24 GB per GPU, Runpod's 24 GB options include the RTX A5000 at $0.27/hr and the L4 at $0.69/hr on Secure Cloud.
How Much VRAM Does the RTX 4090 Have?
The RTX 4090 has 24 GB of GDDR6X VRAM on a 384-bit memory bus, with an effective speed of 21 Gbps and roughly 1 TB/s of memory bandwidth. This is the most VRAM available on any consumer GPU and is the primary reason the 4090 remains relevant for AI workloads even after the RTX 5090 launch.
24 GB is enough to run inference on most open-source large language models at full precision, including 7B, 13B, and quantized versions of 70B+ parameter models. For fine-tuning, 24 GB handles models up to roughly 7-13B parameters depending on precision and batch size. For creative workloads, it handles high-resolution Stable Diffusion and video generation without memory errors that plague smaller cards.
For context:
- RTX 4090: 24 GB GDDR6X, ~1 TB/s bandwidth
- RTX 5090: 32 GB GDDR7, ~1.8 TB/s bandwidth
- Nvidia H100 SXM: 80 GB HBM3, ~3.35 TB/s bandwidth
- Nvidia A100 80GB: 80 GB HBM2e, ~2 TB/s bandwidth
The 4090's VRAM advantage over cheaper cards is significant. The RTX 3090 had 24 GB, but the 4090 delivers that capacity with dramatically faster bandwidth and compute throughput.
RTX 4090 Specs
The RTX 4090 is built on Nvidia's Ada Lovelace architecture using TSMC's 5nm process. It was released in October 2022 as the company's most powerful consumer GPU to date. Here are the full specifications:
- CUDA Cores: 16,384
- Tensor Cores: 512 (4th generation), the primary driver of AI/ML performance, delivering 2-4x faster training vs. the previous generation
- RT Cores: 128 (3rd generation)
- VRAM: 24 GB GDDR6X on a 384-bit bus
- Memory Bandwidth: ~1 TB/s
- Base Clock: 2.23 GHz
- Boost Clock: 2.52 GHz
- FP32 Throughput: ~82.6 TFLOPS
- TDP: 450W (16-pin PCIe 5.0 connector)
- Process Node: TSMC 4N (5nm-class)
- NVLink: Not supported (40-series dropped multi-GPU support)
- DirectX: 12 Ultimate
- Key Features: DLSS 3 with frame generation, AV1 encode/decode (NVENC/NVDEC), Nvidia Broadcast, Reflex
RTX 4090 Price: What Does It Cost in 2026?
The RTX 4090 launched at an MSRP of $1,599 in October 2022. As of mid-2026 it sells for roughly $1,800 to $3,500 depending on model and retailer: Founders Edition cards sit near the bottom of that range where you can find them, while factory-overclocked partner cards from the likes of ASUS and MSI reach the top of it. Typical street prices through major US retailers have hovered in the $2,400 to $2,800 band through the first half of 2026.
The card is well above its launch MSRP nearly four years on, which is unusual. Three things keep it there: end-of-life production status, sustained demand from AI developers, and thin retail inventory. Secondary market prices have held up unusually well compared to previous GPU generations, reflecting that 24 GB of VRAM at this price point has no direct consumer-GPU competition (the RTX 5090 launched at $1,999 for 32 GB).
Renting cloud GPUs avoids the upfront cost, electricity, cooling, and depreciation that come with owned hardware. For 24 GB of VRAM on Runpod, the options include the RTX A5000 and L4. The RTX A6000 and L40S have 48 GB if you need more headroom.
RTX 4090 Performance and Use Cases
The RTX 4090's 82.6 TFLOPS of FP32 compute and 4th-gen Tensor Cores make it one of the most capable single GPUs available for a wide range of workloads:
- AI Inference: The 4090 handles inference for most open-source models, including LLaMA 3, Mistral, Stable Diffusion XL, Flux, and Whisper, at production-quality throughput. Its 24 GB VRAM is the key enabler, allowing full-precision loading of models that would require quantization on 16 GB or 12 GB cards.
- AI Training and Fine-Tuning: QLoRA and LoRA fine-tuning of 7B-13B parameter models runs efficiently on a single 4090. Full fine-tuning of smaller models (up to ~3B parameters) is also practical. The 4th-gen Tensor Cores deliver 2-4x the training throughput of the RTX 3090.
- Stable Diffusion and Image Generation: The 4090's 24 GB handles high-resolution image generation. Full-resolution SDXL, ControlNet stacks, video generation (Wan, Mochi, CogVideoX), and real-time generation workflows all run without memory constraints.
- 4K and 8K Gaming: The 4090 is capable of native 4K at high refresh rates in the most demanding titles, with headroom for ray tracing and DLSS 3 frame generation.
- 3D Rendering and Content Creation: Blender, DaVinci Resolve, and other GPU-accelerated creative tools see dramatic speedups. The large VRAM allows complex scenes and high-resolution footage to be processed entirely in memory.
RTX 4090 vs. RTX 5090: How Do They Compare?
The RTX 5090 launched in January 2025 at $1,999 and introduces Nvidia's Blackwell architecture with 32 GB of GDDR7 VRAM and significantly improved memory bandwidth (~1.8 TB/s vs ~1 TB/s on the 4090). For most AI workloads, the 5090 offers roughly 30-40% better throughput depending on precision and model type.
Other factors to weigh:
- Price gap: The 5090 launched $400 above the 4090's MSRP and commands a premium on the secondary market.
- VRAM is often the bottleneck, not compute: For inference workloads where the model fits in 24 GB, the 4090 and 5090 will perform similarly. The bandwidth improvement matters most at the memory ceiling.
- Software maturity: The 4090 has years of driver optimization, community tooling, and benchmarked configurations behind it. Quantization strategies and inference libraries are well-tuned for Ada Lovelace.
Which card fits depends on VRAM first. Workloads that fit in 24 GB run on either card, and the 5090's extra 8 GB and bandwidth matter most near the memory ceiling. On Runpod, the RTX A5000 and L4 cover 24 GB workloads, and the RTX A6000 and L40S have 48 GB for more headroom. These are matched on VRAM, not speed. The A5000 and L4 are slower than a 4090, so benchmark your own job before sizing a long run.
RTX 4090 vs. A100 and H100: How Do They Compare?
The A100 and H100 are Nvidia's data-center GPUs, built for scale, reliability, and large-model training. Here is how they compare to the 4090 for AI workloads:
- VRAM: Both A100 and H100 offer up to 80 GB of HBM memory, more than 3x the 4090's 24 GB. This is the decisive advantage for models that require more than 24 GB at inference or for large-batch training runs.
- Memory Bandwidth: A100: ~2 TB/s. H100 SXM: ~3.35 TB/s. The 4090's ~1 TB/s is meaningfully lower, which matters for memory-bound workloads like large transformer inference.
- Multi-GPU Support: A100 and H100 support NVLink for high-bandwidth multi-GPU communication, critical for training models across multiple GPUs. The 4090 does not support NVLink.
- Enterprise Features: A100/H100 include ECC memory, MIG virtualization, and are designed for 24/7 datacenter operation. The 4090 is a consumer card and lacks these.
- Cost on Runpod: An A100 is $2.09/hr and an H100 is $2.89/hr on Secure Cloud. For workloads that fit in 24 GB, the RTX A5000 is $0.27/hr.
For workloads that fit in 24 GB or 48 GB, the RTX A5000, L4, RTX A6000, and L40S on Runpod cover that range. For large-scale training or models that require 40-80 GB of VRAM, an A100 or H100 is the right tool, and both are available on Runpod at on-demand rates.
24 GB and 48 GB GPUs on Runpod
Runpod offers on-demand GPU instances without hardware procurement, setup, or maintenance. You get full root access to a containerized GPU environment and can be running within minutes of signing up.
- Secure Cloud pricing: From $0.27/hr for an RTX A5000 on Runpod-owned hardware, with per-second billing
- Serverless: Deploy models as HTTP endpoints on Serverless workers, paying per request rather than per hour
- Network volumes: Attach persistent storage to keep model weights, datasets, and outputs available across sessions. Standard storage starts at $0.07/GB/mo (first TB, then $0.05/GB/mo), with high-performance storage at $0.14/GB/mo for faster data loading. A single volume can be shared across multiple instances simultaneously.
- Templates: Pre-built environments for PyTorch, Stable Diffusion, vLLM, Jupyter, and more with no manual setup required
- Scalability: Spin up multiple instances in parallel, or switch to an RTX A6000 or L40S (48 GB), A100, H100, H200, B200 or B300 when your workload demands more VRAM
See current availability and pricing on the Runpod pricing page.
RTX 4090 FAQs
How much VRAM does the RTX 4090 have?
The RTX 4090 has 24 GB of GDDR6X VRAM on a 384-bit memory bus, with approximately 1 TB/s of memory bandwidth. This is the highest VRAM capacity available on any consumer GPU and is sufficient for inference on most open-source LLMs and generative AI models.
What is the RTX 4090?
The Nvidia GeForce RTX 4090 is the flagship GPU of Nvidia's Ada Lovelace (40-series) architecture, launched in October 2022. It features 16,384 CUDA cores, 512 Tensor Cores, 24 GB of GDDR6X VRAM, and 82.6 TFLOPS of FP32 compute. It is the most powerful consumer GPU Nvidia produced before the RTX 5090.
How much does the RTX 4090 cost?
The RTX 4090 launched at an MSRP of $1,599. As of mid-2026 it sells for roughly $1,800 to $3,500 depending on model and retailer, with typical street prices through major US retailers in the $2,400 to $2,800 band. Founders Edition cards sit at the lower end where available; factory-overclocked partner cards reach the top. For 24 GB of VRAM in the cloud, Runpod's RTX A5000 is $0.27/hr on Secure Cloud.
Why is the RTX 4090 still above its launch price?
End-of-life production status, sustained demand from AI developers, and thin retail inventory. Twenty-four gigabytes of VRAM at this price point has no direct consumer-GPU competition, and demand from people running local inference has kept the secondary market unusually firm compared with previous GPU generations.
Is the RTX 4090 good for AI and deep learning?
Yes, for workloads that fit in 24 GB. Its 24 GB of VRAM allows inference on most open-source LLMs, and its 4th-gen Tensor Cores deliver strong FP16/BF16 training throughput. It is particularly well-suited for fine-tuning (QLoRA/LoRA), Stable Diffusion, and inference serving of models up to 13B parameters at full precision. On Runpod, the RTX A5000 and L4 offer the same 24 GB of VRAM.
RTX 4090 vs RTX 5090: which is better for AI?
The RTX 5090 has 32 GB GDDR7 VRAM and ~1.8 TB/s bandwidth vs the 4090's 24 GB and ~1 TB/s. For workloads that fit in 24 GB, the performance difference is modest (30-40%). The 5090's extra 8 GB and bandwidth matter most for workloads near the 24 GB limit. On Runpod, the RTX A5000 and L4 cover 24 GB workloads, and the RTX A6000 and L40S offer 48 GB if VRAM is the limit. The A5000 and L4 are slower than a 4090, so they're matched on memory, not speed.
How does the RTX 4090 compare to the H100?
The H100 has 80 GB of HBM3 memory, ~3.35 TB/s bandwidth, NVLink support, and ECC memory, designed for datacenter-scale AI training. The 4090 has 24 GB of GDDR6X and costs considerably less to rent. For large-model training, multi-GPU jobs, or 80 GB VRAM requirements, the H100 is the right choice, and it's available on Runpod. For workloads that fit in 24 GB, Runpod also offers the RTX A5000 and L4.
Related guides
- Nvidia H200 GPU: Specs, VRAM, Price, and AI Performance
- Nvidia H100 GPU: Specs, VRAM, Price, and AI Performance
- Nvidia B200 GPU: Specs, VRAM, Price, and AI Performance
- AI Model Serving Architecture for Scalable APIs
- Deploy vLLM with Docker on Runpod: Container Config, Model Loading, and Production Tuning
- The LLM inference optimization playbook: architecting for latency, throughput, and cost
