mreltera/qwen3-coder-gb10-vs-rtx5090-benchmark

NVIDIA GB10 vs. GeForce RTX 5090 - Local LLM Inference Benchmark

Model: Qwen3-Coder-30B-A3B-Instruct
Format: GGUF, Q4_K_M, 18.63 GB
Runtime: LM Studio / llama.cpp
Author: Efehan A.
Benchmark date: 5 August 2026

This repository contains a decode-focused local inference benchmark comparing an NVIDIA GB10 system with a Windows workstation containing two GeForce RTX 5090 GPUs. Telemetry shows that the inference workload was carried primarily by a single RTX 5090 (GPU 0), while GPU 1 remained mostly idle.

Headline result

Platform	Output	Runs	Mean tokens/s	Std. dev.	Mean TTFT	Mean generation time
NVIDIA GB10	512	5	44.93	0.02	0.0862 s	11.374 s
GeForce RTX 5090	512	5	243.46	1.32	0.0280 s	2.099 s
NVIDIA GB10	1024	5	43.80	0.03	0.1000 s	23.355 s
GeForce RTX 5090	1024	5	240.97	0.89	0.0320 s	4.245 s

Observed RTX 5090 throughput advantage: 5.42x at 512 tokens and 5.50x at 1024 tokens.

Test configuration

• Model: Qwen3-Coder-30B-A3B-Instruct
• Architecture reported by LM Studio: qwen3moe
• Quantization: Q4_K_M
• GPU offload: 48 layers / maximum
• Evaluation batch size: 512
• Max concurrent predictions: 1
• Unified KV cache: enabled
• KV cache GPU offload: enabled
• Flash Attention: enabled
• K/V cache quantization: disabled
• API request temperature: 0
• One warm-up request followed by five measured requests per batch

Benchmark batches

1. GB10, 512 completion tokens, five measured runs.
2. GB10, 1024 completion tokens, five measured runs.
3. RTX 5090, 512 completion tokens, five measured runs.
4. RTX 5090, 1024 completion tokens, five measured runs.

Repeatability

Throughput coefficients of variation were approximately 0.04% and 0.07% on GB10, and 0.54% and 0.37% on RTX 5090. The cross-platform difference is much larger than run-to-run variation.

1024-token sustained generation

At 1024 tokens, RTX 5090 averaged 240.97 tokens/s versus 43.80 tokens/s on GB10. Doubling output length reduced throughput by only 2.50% on GB10 and 1.02% on RTX 5090.

Interpretation

The tested 18.63 GB model fits inside the RTX 5090's 32 GB GDDR7 memory. NVIDIA specifies 1,792 GB/s peak memory bandwidth for the RTX 5090, compared with 273 GB/s for the GB10 unified-memory platform. The measured decode gap is directionally consistent with the large bandwidth difference, although operating system, runtime version and kernel differences prevent single-cause attribution.

The systems target different deployment priorities:

• RTX 5090: maximum single-stream decode speed when the model fits in 32 GB.
• GB10: 128 GB unified-memory capacity for larger models, larger KV caches and longer-context workloads that do not fit on a single RTX 5090.

Important limitations

• The 512-token benchmark output reports context 8192 on GB10 and 10132 on RTX, despite load-status captures showing 8192 on both systems. The corrected 1024-token results report context 8192 on both systems.
• Runtime builds differed: ARM64/Linux CUDA 13 versus x86-64/Windows CUDA 12.
• CPU thread settings differed: 20 on GB10 and 8 on RTX.
• No direct prompt-prefill benchmark was completed.
• Power efficiency is not compared because GB10 SoC telemetry and discrete-GPU board power are not directly equivalent.
• Only one model, quantization and inference application were tested.

Files

• benchmark_report.pdf - full professional report
• benchmark_summary.csv - aggregate results
• run_level_results.csv - all measured runs
• evidence/ - public-release screenshots

References

1. NVIDIA DGX Spark Hardware Overview: https://docs.nvidia.com/dgx/dgx-spark/hardware.html
2. NVIDIA GeForce RTX 5090 Specifications: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
3. NVIDIA RTX Blackwell GPU Architecture: https://images.nvidia.com/aem-dam/Solutions/geforce/blackwell/nvidia-rtx-blackwell-gpu-architecture.pdf
4. LM Studio REST API v0: https://lmstudio.ai/docs/developer/rest/endpoints