Skip to content

RTX 4090 vs RTX 5090

Side-by-side comparison of the NVIDIA RTX 4090 and the NVIDIA RTX 5090 for AI inference workloads.

Specifications

SpecRTX 4090RTX 5090
Generationadablackwell
Memory TypeGDDR6XGDDR6X
VRAM24 GB32 GB
Memory Bandwidth1,008 GB/s1,792 GB/s
BF16 TFLOPS165210
FP16 TFLOPS165210
FP8 TFLOPS330420
INT8 TOPS330420
TDP450 W575 W
Interconnectpciepcie
Max GPUs per Node44
PCIe GenGen 4Gen 5
CUDA Compute Capability8.98.9

Pricing

RTX 4090

ProviderOn-DemandReservedSpot
runpod$1.10/hr-$0.79/hr
lambda$0.89/hr--
vast ai$0.74/hr-$0.44/hr
tensordock$0.69/hr-$0.44/hr
fluidstack$0.59/hr-$0.39/hr

RTX 5090

ProviderOn-DemandReservedSpot
runpod$1.69/hr-$1.29/hr
vast ai$1.29/hr-$0.89/hr
tensordock$1.19/hr-$0.89/hr

Cheapest available rate: RTX 4090 at $0.39/hr vs RTX 5090 at $0.89/hrRTX 4090 is +128% cheaper

Efficiency Metrics

TFLOPS / Watt

0.4

RTX 4090

vs

0.4

RTX 5090

BF16

VRAM / Dollar

61.5

RTX 4090

vs

36.0

RTX 5090

GB/$/hr

Bandwidth / Watt

2.2

RTX 4090

vs

3.1

RTX 5090

GB/s/W

Models (FP16, 1 GPU)

124.0

RTX 4090

vs

148.0

RTX 5090

Model Compatibility (FP16, Single GPU)

Only on RTX 4090 (0)

None

Both (124)

  • Yi 1.5 9B
  • Yi Coder 9B
  • GTE Qwen2 7B
  • Marco O1
  • Qwen 2 Audio 7B
  • Qwen 2.5 3B
  • OLMo 2 7B
  • OpenELM 3B
  • BGE Large EN v1.5
  • BGE M3
  • StarCoder2 3B
  • StarCoder2 7B
  • Aya 23 8B
  • Command R 7B
  • DeepSeek Coder 6.7B
  • DeepSeek Math 7B
  • DeepSeek R1 Distill 8B
  • Falcon 7B
  • Gemma 1.1 2B
  • Gemma 2 2B
  • +104 more

Only on RTX 5090 (24)

  • Qwen 1.5 MoE A2.7B
  • Qwen 2.5 14B
  • OLMo 2 13B
  • Amazon Nova Lite
  • Baichuan 2 13B
  • DeepSeek R1 Distill 14B
  • Gemma 3 12B
  • Vicuna 13B
  • Code Llama 13B
  • Llama 2 13B
  • Orca 2 13B
  • Phi 3 Medium 14B
  • Phi-4
  • Mistral Nemo 12B
  • Pixtral 12B
  • VILA 1.5 13B
  • Qwen 2.5 Coder 14B
  • RWKV-6 14B
  • StableLM 2 12B
  • ELYZA 13B
  • +4 more

Summary

The RTX 4090 (ada generation) offers 24GB of GDDR6X with 165 BF16 TFLOPS and 1,008 GB/s memory bandwidth at 450W TDP.

The RTX 5090 (blackwell generation) offers 32GB of GDDR6X with 210 BF16 TFLOPS and 1,792 GB/s memory bandwidth at 575W TDP.

The RTX 5090 has +33% more VRAM, allowing it to run larger models without multi-GPU setups.

From a cost perspective, the RTX 4090 is more affordable at $0.39/hr vs $0.89/hr for the RTX 5090.

More GPU Comparisons