Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Local large language models (LLMs) can be slow even when your CPU is busy and your operating system reports that a GPU is available. Generation is an autoregressive workload: the model repeatedly computes the next token, moving large amounts of model data through memory for every step. Actual speed depends on the exact model, quantization, context length, runtime, backend, processor instructions, memory system, drivers, and device configuration. An integrated GPU adds parallel compute, but usually shares system memory with the CPU rather than using dedicated video memory. That can create both capacity limits and transfer overhead.
There is no universal rule that a GPU beats a CPU, that more cores guarantee higher tokens per second, or that adding RAM fixes slow generation. The reliable approach is to identify the bottleneck on your machine, verify that the intended backend is really being used, and compare equivalent CPU-only and GPU-enabled runs.
As an Amazon Associate I earn from qualifying purchases.
What the computer is doing during generation
An LLM does not calculate an entire answer in one operation. It processes the prompt, then predicts one token, adds that token to the sequence, and repeats the computation for the next token. This repeated loop is why a model can require substantial compute and memory bandwidth even when the prompt is short.
Prompt processing and token generation are different workloads
Most runtimes expose two phases. Prompt processing (often called prefill) handles many input tokens at once and can benefit from parallel hardware and larger batches. Token generation (decode) produces output one token at a time and is often limited by memory movement, synchronization, or latency. A configuration that improves prompt processing may not improve the speed at which text appears. Record both figures when your runtime reports them.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
Why a high CPU percentage is not a diagnosis
A CPU at 100% means its available scheduling capacity is being used; it does not identify whether arithmetic, cache misses, memory bandwidth, thread synchronization, or an inefficient kernel is limiting the run. Core count, clock speed, and RAM capacity each matter in some configurations, but none predicts tokens per second by itself. CPU instruction-set support and optimized kernels can change the result substantially. The llama.cpp project supports several CPU instruction sets, quantization levels, and hardware-specific paths, so two processors with similar marketing specifications can behave differently.
Why CPUs can be slow with local models
The model repeatedly moves a large working set
Every generated token requires access to model weights and a growing key-value (KV) cache. If the working set does not fit comfortably in cache, the processor repeatedly fetches data from system memory. The resulting limit may be memory bandwidth or latency rather than the number of arithmetic units. Longer context windows increase KV-cache use and can make later tokens slower or reduce the amount of memory available for the model itself.
Quantization changes the workload
Quantization stores weights with fewer bits. llama.cpp supports integer formats ranging from roughly 1.5-bit through 8-bit representations. Lower-bit formats generally reduce the memory required to load a model and can make it possible to run a model that otherwise would not fit. They also change dequantization work and numerical behavior. A lower-bit model is not guaranteed to generate faster on every CPU, and output quality can vary with the model, quantization method, and task. Treat quantization as a memory-and-quality trade-off, then measure speed on your own workload.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Threading and build choices matter
Too few threads leave compute resources idle; too many can increase contention, cache pressure, and scheduling overhead. A runtime built without the processor’s available vector instructions can be much slower than an optimized build. Thermal throttling, background processes, power-management modes, and a laptop’s battery profile can also change sustained performance. These variables explain why a brief benchmark can look better than a long generation session.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Why an integrated GPU may not be faster
Shared memory is both capacity and a bandwidth constraint
An integrated GPU normally has no separate pool of high-bandwidth video memory. In Intel’s documented SYCL path, the GPU uses host shared memory, which is also needed by the operating system, applications, model buffers, and CPU-side runtime. The nominal amount of installed RAM is therefore not the same as memory available to the model. Firmware-reserved graphics memory, operating-system use, the context, and other applications reduce the usable pool.
Shared memory can also create competition between CPU and GPU. When both devices access the same memory system, transfers and synchronization can consume bandwidth. If the model is partly on the CPU and partly on the GPU, data may move between devices repeatedly. Offloading more layers is not automatically better if the transfer and coordination costs exceed the saved computation.
Execution resources and support are device-specific
Intel’s example for a llama2–7b-Q4 workload recommends an integrated GPU with at least 80 execution units and more than 4.5 GB of shared memory on a system with at least 16 GB of host memory. The llama.cpp SYCL guide gives another example: it lists at least 8.0 GB for its llama-2-7b.Q4_0 integrated-GPU case and warns that an integrated GPU with fewer than 80 execution units will likely be too slow for practical use. These figures belong to those documented examples, model variants, and software paths; they are not universal minimums for every LLM, context length, driver, or backend.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesDetection is not the same as acceleration
An operating system can list a GPU while the inference program uses a fallback path, an unoptimized backend, or only a small fraction of the model on the device. llama.cpp supports CPU, Vulkan, SYCL, and other hardware paths, as well as hybrid CPU/GPU inference when a model exceeds available device memory. The runtime log—not the desktop’s hardware panel—is the useful evidence. Look for the selected backend, detected device, model-buffer placement, and the number of layers or tensors actually offloaded.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
Why a CPU can beat GPU acceleration
A 2025 paper, Challenging GPU Dominance: When CPUs Outperform for On-Device LLM Inference, measured a one-billion-parameter model on an iPhone 15 Pro. Under the paper’s stated settings, a CPU-only F16 run using two threads reached 17 tokens per second, while GPU acceleration reached 12.8 tokens per second. The authors discuss memory-transfer overhead and CPU thread optimization, and note that restricted low-level profiling limited their explanation of the result. It demonstrates that CPU can win in a particular device, model, precision, and software configuration; it is not evidence that CPUs are generally faster.
A separate 2025 study of tightly coupled CPU-GPU servers found that bottlenecks moved with workload and batch size. In its studied configuration, a Grace Hopper system achieved 1.9×–2.7× lower prefill latency for Llama 3.2-1B at large batch sizes than loosely coupled systems, while also identifying CPU-bound regions and benefits from kernel fusion at low batch sizes. Those are server findings, not laptop benchmarks. They reinforce the practical rule: the bottleneck depends on architecture and workload.
How to find the bottleneck on your machine
1. Record a reproducible baseline
- Write down the exact model family and parameter count, quantization format, context length, runtime version, backend, operating-system version, processor, integrated-GPU generation, and installed RAM.
- Use the same prompt, output length, temperature, and sampling settings for every comparison. Stop other heavy applications and run on wall power if the laptop has a performance power mode.
- Record prompt-processing speed and token-generation speed separately when available. Also note peak memory use, whether the run swaps to disk, and whether the speed declines as context grows.
2. Read the runtime log
Confirm that the intended backend is selected and that a real device is recognized. In the llama.cpp SYCL documentation, logs show model-buffer size; use that information to determine whether the model fits the available shared memory. Check for messages indicating CPU fallback, insufficient device memory, rejected layers, or an unsupported kernel. A GPU name in the log is not enough if the model remains resident on the CPU.
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 matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstall3. Compare CPU-only and offloaded runs
Run the same task once with GPU offload disabled and once with the documented backend enabled. Keep quantization, context, prompt, and generation length unchanged. If GPU mode lowers prompt time but leaves decode speed unchanged, decode is likely limited elsewhere. If it is slower in both phases, investigate transfers, unsupported kernels, driver versions, or a device that is too small for the chosen model.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
4. Test a smaller memory footprint
Try a smaller model or a lower-memory quantization that your runtime supports. If the smaller configuration is dramatically faster, capacity pressure, paging, or memory bandwidth is likely involved. Do not infer a fixed percentage improvement: gains vary by CPU, GPU, backend, context, and quantization.
Intel integrated-graphics checks
For Intel’s documented llama2–7b-Q4 SYCL example, check all of the following before judging performance:
- The integrated-GPU generation and execution-unit count fall within the guide’s support guidance.
- The graphics driver and required oneAPI runtime are installed and current for that backend.
- The inference program recognizes the GPU device and reports model buffers or layers on it.
- Enough host memory remains after the operating system, applications, model, KV cache, and runtime allocations are counted.
- The selected model and context fit without swapping or repeated allocation failures.
If any check fails, changing thread counts will not turn an unsupported or memory-starved path into a fast one. Fix the backend or reduce the workload first.
CPU-only, integrated-GPU, or another backend?
| Option | When it can make sense | Risks and trade-offs | What to measure |
|---|---|---|---|
| CPU-only | The model fits in system memory and the runtime has optimized CPU kernels for the processor. | Decode may be limited by memory traffic; long contexts and larger models increase work. | Tokens per second, thread scaling, sustained temperature, and memory use. |
| Integrated-GPU offload | The exact GPU, driver, and backend are supported and shared memory can hold the workload. | CPU and GPU compete for host memory; transfers and partial offload can erase gains. | Device recognition, offloaded buffers/layers, prompt speed, decode speed, and peak shared memory. |
| Supported alternative backend | Your runtime offers a suitable Vulkan, SYCL, or other path for the exact device. | Installation and driver complexity vary; a supported path may still be slower for a particular model. | Backend initialization, kernel selection, memory placement, and identical-task throughput. |
Do not rank these options in the abstract. Choose the path that supports your hardware, fits the model and context, and wins a controlled comparison on the task you actually run.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Does adding RAM make a local LLM faster?
More RAM can solve a capacity problem: it may let the model, KV cache, and runtime fit without swapping, or give an integrated GPU a larger shared-memory budget. It does not prove that token generation will accelerate once the workload already fits. Shared-memory performance also depends on memory channels, speed, firmware settings, CPU/GPU architecture, and contention.
Before buying a kit, verify the computer’s memory type, maximum supported capacity, number of upgradeable slots, channel configuration, and whether memory is soldered. Upgrade only when measurements show that capacity or swapping is the limiting factor. If memory use is comfortably below the limit, investigate backend support, quantization, thermals, and thread configuration instead.
Common symptoms and targeted fixes
“GPU enabled” but speed is unchanged
- Check the runtime log for actual device use and offloaded buffers.
- Compare prompt and decode rates separately; only one phase may benefit.
- Test a smaller model to reveal whether shared-memory pressure is masking acceleration.
- Verify that the selected backend is optimized for the device rather than a compatibility fallback.
Generation starts quickly, then slows
- Measure whether the context and KV cache grow until memory pressure appears.
- Check for thermal throttling during a sustained run.
- Reduce context length or use a model/quantization that leaves more headroom.
The model fails to load or the system becomes unusable
- Close other memory-heavy applications and disable disk swapping as a supposed performance strategy; swapping is a sign the workload does not fit.
- Reduce model size, context, or precision until the runtime can allocate buffers without exhausting shared memory.
- Inspect logs for the exact allocation or backend error before changing hardware.
Practical expectations
Published figures are useful only with their conditions attached. The 17-versus-12.8-token-per-second result was a one-billion-parameter iPhone 15 Pro experiment with stated CPU and F16 settings. The 1.9×–2.7× prefill result came from a particular server architecture and large-batch workload. Intel’s memory and execution-unit numbers describe documented examples, not guarantees for your computer. A benchmark that omits model, quantization, context, backend, and hardware details cannot predict your result.
Free tools Windows power users keep installed
One-click scans. No signup required.
The fastest path is therefore empirical but controlled: make the model fit, verify the backend, hold the task constant, separate prompt from decode performance, and change one variable at a time. That process tells you whether the limiting factor is model size, shared memory, CPU kernels, transfers, thermals, or software configuration.
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.
