Recommended Free Tools
In one Lenovo Yoga 9 15IMH5, the GeForce GTX 1650 Ti Max-Q decoded Gemma 4 E2B at a median 4.14 times the rate of the laptop’s six-core Intel Core i7-10750H. The same test reported a 3.42× GPU advantage in prefill and a 3.62× advantage end to end. These are results for one laptop, model quantization, software build and benchmark protocol—not a general rule about GPUs and CPUs.
What did the benchmark measure?
In a 2026 DEV Community article, author xbill compared local Gemma 4 E2B inference on the CPU and GPU in the same Lenovo Yoga 9 15IMH5. The measured result was a median GPU-to-CPU decode ratio of 4.14×. Decode is the stage that generates output tokens after the prompt has been processed; prefill processes the prompt, and end-to-end timing includes both stages. Read the benchmark article.
As an Amazon Associate I earn from qualifying purchases.
The result grid reported CPU decode rates from 16.10 to 18.31 tokens per second and GPU rates from 68.22 to 71.89 tokens per second. Across the eight combinations of prompt and output length, GPU-to-CPU decode ratios were 3.93×, 3.98×, 4.14×, 4.09×, 4.30×, 4.14×, 4.25× and 4.17×. The median of those ratios was 4.14×.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What hardware and software were used?
The comparison used a six-core, 12-thread Intel Core i7-10750H CPU with AVX2 and an NVIDIA GeForce GTX 1650 Ti with Max-Q Design and 4096 MiB of GPU memory. Both components were tested in the same laptop, so the result reflects that machine’s shared chassis and cooling as well as the processors themselves.
#1 Best Overall
- Desktop-Level Performance, Anywhere: Get legendary gaming performance with the Intel Core Ultra 9 275HX processor, delivering ultra-smooth gameplay and future-ready AI (Up to 13 NPU TOPS). Offload tasks like background removal and audio optimization to the NPU for seamless streaming and gaming, while Intel Application Optimization enhances performance on classic titles.
- Game-Changing Realism: Powered by NVIDIA Blackwell architecture, GeForce RTX 5070 Ti Laptop GPU unlocks the game changing realism of full ray tracing. Equipped with a massive level of 992 AI TOPS horsepower, the RTX 50 Series enables new experiences and next-level graphics fidelity. Experience cinematic quality visuals at unprecedented speed with fourth-gen RT Cores and breakthrough neural rendering technologies accelerated with fifth-gen Tensor Cores.
- Supreme Speed. Superior Visuals. Powered by AI: DLSS is a revolutionary suite of neural rendering technologies that uses AI to boost FPS, reduce latency, and improve image quality. DLSS 4 brings a new Multi Frame Generation and enhanced Ray Reconstruction and Super Resolution, powered by GeForce RTX 50 Series GPUs and fifth-generation Tensor Cores.
- The Ultimate in Ray Tracing and AI: NVIDIA RTX is the most advanced platform for full ray tracing and neural rendering technologies that are revolutionizing the ways we play and create. Over 700 games and applications use RTX to deliver realistic graphics and incredibly fast performance with cutting-edge AI features like DLSS Multi Frame Generation.
- Immersive Depth and Detail: At 18 inches with a 16:10 aspect ratio, the pristine WQXGA screen offering vibrant colors with up to 100% DCI-P3 operates at a fast 240Hz refresh and 3ms overdrive response time. Alongside the suite of features from NVIDIA G-SYNC and NVIDIA Advanced Optimus, you're guaranteed that whatever's on-screen is a distinct viewing delight.
- Computer: Lenovo Yoga 9 15IMH5.
- Model:
google/gemma-4-E2B-it-qat-q4_0-gguf, a 3.35 GB quantization-aware GGUF, as described by the author. - Inference software:
llama-serverfrom llama.cpp commitf95b0d9, build 318. - System and toolchain: Debian forky/sid with kernel 7.2.6, gcc 16.2.0, CUDA 13.4 (V13.4.92) and NVIDIA driver 615.71.09.
Server settings were matched except for -ngl, which controls GPU layer offload. Shared settings included an 8192-token context, f16 key/value cache, flash attention, six CPU threads, 12 batch threads, one parallel request and metrics enabled.
How was run order controlled?
The author used an ABBA sequence: CPU, GPU, GPU, CPU, with two passes per device. Before each pass, the script waited at least 120 seconds and also required CPU package temperature to be at or below 50 °C and GPU temperature at or below 45 °C. The report describes the prompt cache as cold and the page cache as warm.
Each pass covered four prompt lengths—94, 516, 998 and 1959 tokens—and two output lengths—32 and 128 tokens. Each prompt/output combination was repeated three times, with one request running at a time. This gives eight paired workload cells for comparing the devices.
Free tools Windows power users keep installed
One-click scans. No signup required.
The ordering check matters because laptop performance can shift as the chassis heats. The article reports these combined and order-specific median ratios:
Rank #2
- Intel Core i9 14th Gen 14900HX 1.6GHz Processor, NVIDIA GeForce RTX 5070 8GB GDDR7, 32GB DDR5-5600 RAM
- 1TB PCIe Gen4 x4 NVMe M.2 SSD
- 15.1" WQXGA OLED Glossy Display
- Gigabit LAN, 2x2 WiFi 7 (802.11be), Bluetooth 5.4
- 4.19 lbs. (1.90 kg)，Windows 11 Home
| Run grouping | Decode lead | Prefill lead |
|---|---|---|
| CPU-first | 4.09× | 3.38× |
| GPU-first | 4.17× | 3.47× |
| Combined ABBA result | 4.14× | 3.42× |
On this laptop, the author characterized the order effect as about 2% of the ratio. The CPU-first sequence gives the initially cooler CPU pass an advantage, while the later GPU pass runs after more heat has accumulated in the chassis. ABBA helps expose that order sensitivity; it does not eliminate every possible thermal effect.
What do the prefill and timing results add?
The GPU’s reported median advantage was 3.42× for prefill and 3.62× for end-to-end performance, both smaller than its 4.14× decode lead. Longer prompts increased time to first token on both devices: across the tested prompt range, the GPU’s time-to-first-token advantage ranged from 2.84× to 3.56×. Decode rates remained within the device-specific ranges reported above.
For interactive use, decode speed affects how quickly an answer continues to appear after generation begins; prefill affects the wait before the first generated token, particularly when a prompt is longer. The benchmark therefore suggests the GPU was faster across the measured stages, but its relative advantage varied by stage.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11How consistent were the repeated runs?
The article reports a median GPU pass-to-pass decode drift of +0.6%, ranging from 0.0% to +1.3%. CPU pass-to-pass drift had a median of -1.1%, ranging from -9.3% to +0.1%. Within a workload cell, CPU repeat spread reached 14.05%, compared with a maximum GPU repeat spread of 1.14%; the CPU’s second pass also had twice as many throttle events as its first.
Rank #3
- [Top Performance Processors] KAIGERR Light gaming laptop R7-5700U by ΑΜD ZEN 3 architecture, matched with 16MB of L3 cache, built by TSMC 7nm process with 8 cores & 16 threads (turbo up to 4.3GHz). KAIGERR office light gaming laptops makes it easy to qualify for your PC work and PC games which have amazing loading and processing power for a smoother PC used experience
- [Huge Capacity Storage] KAIGERR laptop comes with 16GB SODIMM DDR4 RAM, advantages of large operating memory capacity both can reduce read latency of memory data and improve CPU utilization. KAIGERR laptop computer configured with an M.2 2280 NVMe 512GB SSD which offers fast startup and loading of applications, as well as a large amount of storage space for your various files
- [Brilliant Display & Integrated Graphics] KAIGERR Light gaming laptop features an innovative thin-bezel display that provides more usable onscreen space for immersive FHD viewing. KAIGERR laptop integrates with ΑΜD Radeon Graphics and delivers strong graphics processing like a rich level of image detail making it possible to play computer games or edit pictures with a great experience on this laptop
- [Rich Interfaces & Wireless Connectivity] KAIGERR traditional laptop offers a variety of connectivity options, including HDMI, Type-C, 3.5mm TRRS Jack, Memory Card Slot and USB3.2 ports. You can easily connect to various devices and peripherals to expand your capabilities. Mini laptop computers equipped with WiFi6 & Bluetooth 5.2 which offer strong wireless signal, fast wireless connections, and reliable transmission speed
- [Portable Design & Durable] KAIGERR laptop compact design makes it easy to carry with you wherever you go. Also, you can enjoy the benefits of a powerful computer without the bulk of a traditional desktop. KAIGERR laptop computers are built with high-quality components and designed to handle heavy workloads and deliver consistent performance and longevity. If you encounter any problems, please contact us and we will help you solve the problem within 12 hours
Those observations support saying that CPU measurements were more thermally variable in this particular run. They do not establish that GPUs are generally more stable: that would depend on the hardware, cooling, workload and test conditions.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What can—and can’t—you conclude?
For this setup
If choosing between these two inference paths on this tested laptop, the GPU was substantially faster for serving the specified Gemma 4 E2B quantization. The CPU remains a workable fallback when the GPU is unavailable or occupied, but its measured throughput was lower in this test.
For other machines and workloads
The result covers one laptop, one GGUF model and quantization, one llama.cpp commit, one concurrency setting, eight paired workload cells and one request at a time. It should not be extrapolated directly to other GPUs, CPUs, model sizes, quantizations, software versions or cooling designs. A useful comparison with another benchmark needs to line up the model and quantization, exact CPU and GPU, software build and thread settings, prompt/output lengths, concurrency, run order, temperature controls, and separate decode, prefill and end-to-end measurements.
The author also mentions an earlier run reporting 4.27× decode and 3.63× prefill. Because the commit, thread flags and run order all changed between that run and this one, the difference cannot be attributed to any one change; it is not a controlled before-and-after comparison.
Thermal caveat
A commenter questioned whether a temperature gate before each pass can capture a thermal change that happens during the pass, and suggested recording clocks and temperatures alongside throughput. That is a methodological concern, not evidence that a thermal step invalidated the reported result. The article’s run report was not independently verified here, and no independent replication of this benchmark was established.
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.
