I thought 16GB of VRAM was overkill until I started automating my work with local LLMs

Published Aug 9, 2026, 8:30 AM EDT Nick Lewis is an editor at How-To Geek. He has been using computers for 20 years --- tinkering with everything from the UI to the Windows registry to device firmware. Before How-To Geek, he used Python and C++ as a freelance programmer. In college, Nick made extensive use of Fortran while pursuing a physics degree. Nick's love of tinkering with computers extends beyond work. He has been running video game servers from home for more than 10 years using Windows, Ubuntu, or Raspberry Pi OS. He also uses Proxmox to self-host a variety of services, including a Jellyfin Media Server, an Airsonic music server, a handful of game servers, NextCloud, and two Windows virtual machines. He enjoys DIY projects, especially if they involve technology. He regularly repairs and repurposes old computers and hardware for whatever new project is at hand. He has designed crossovers for homemade speakers all the way from the basic design to the PCB. Nick enjoys the outdoors. When he isn't working on a computer or DIY project, he is most likely to be found camping, backpacking, or canoeing. When I first picked up a 5070 Ti, I did so with gaming in mind. I didn't anticipate constantly running into the VRAM ceiling. Trying to build out a ton of AI automations is a completely different world. 16GB—which is generous for gaming—disappeared almost instantly when I started using local AI for more things. Automation used up all of my VRAM AI models add up fast There are two reasons your memory gets used up. One of them is the models themselves. Depending on what you're doing, a small AI model might use up 2–3GB of VRAM. Smarter LLMs will occupy 10GB of VRAM or more. Once you build a few different workflows that involve different AI models, you can easily use up 16GB. The way to mitigate that is to swap models in and out of VRAM from an SSD as-needed, but that means that any AI task you run has a delay while the pertinent model gets loaded from storage into the GPU. If you're running jobs while you're asleep or away from your PC, that probably won't matter too much, but it is very noticeable if you're sitting there waiting for a model to load. KV Cache in an automated workflow can be a problem The other reason your memory disappears with automation comes down to the KV cache. This cache stores the attention state for every token in your context, and it is stored in your VRAM right alongside the model weights. In a standard chat, like something you might do using LM Studio, you often reset or start fresh. An agent loop, however, accumulates. Every tool call, retrieved result, and step of reasoning adds to the context that must be reprocessed. You have to budget 25% more memory on top of your model weights for an 8K context, and up to double if you're trying for 32K. I'd recommend carefully tuning your default context to try to minimize these sorts of problems. Ollama's default context is 4,096, which is fine for many things. The moment you raise it to handle some complex job, you'll find that your memory fills up abruptly. Quiz 8 Questions · Test Your KnowledgeDIY PC buildingTrivia ChallengeFrom socket types to cable chaos — test your knowledge of building computers from scratch.HistoryHardwareTroubleshootingQuirksTipsBegin What year did Intel release the first consumer processor that popularized the DIY desktop PC market — the Intel 8086?A1972B1975C1978D1981Correct! The Intel 8086 launched in 1978 and gave birth to the x86 architecture still used in PCs today. It was a 16-bit processor running at 5–10 MHz — a far cry from today's multi-GHz giants. This chip laid the foundation for decades of DIY computing.Not quite — the Intel 8086 debuted in 1978. It introduced the x86 instruction set that still underpins virtually every desktop and laptop processor sold today. IBM later used the cheaper 8088 variant for its first PC in 1981, which is sometimes confused as the origin point.Continue When building a PC, what does 'POST' stand for in the context of the boot process?AProcessor Output Self TestBPower-On Self-TestCPrimary Operating System TransferDPower Output Signal ThresholdCorrect! POST stands for Power-On Self-Test, a diagnostic routine your motherboard runs every time you boot up. It checks that critical components like RAM, CPU, and GPU are present and functional. If POST fails, you'll often get beep codes or LED indicators to help diagnose the problem.The correct answer is Power-On Self-Test. Every time you press the power button, your motherboard runs POST to verify that essential hardware is connected and working. Failed POST is one of the first hurdles new PC builders encounter, often caused by unseated RAM or a forgotten power connector.Continue Why do experienced PC builders recommend touching a metal part of the case before handling components?ATo warm up your hands for better gripBTo check if the case is properly grounded before installingCTo discharge static electricity from your bodyDTo magnetize your fingers for handling small screwsCorrect! Static electricity built up on your body can silently destroy sensitive PC components in an instant — a phenomenon called electrostatic discharge (ESD). Touching bare metal grounds you and neutralizes that charge before it can zap your CPU or RAM. Anti-static wrist straps work even better for extended build sessions.The answer is to discharge static electricity. Your body can carry thousands of volts of static charge without you feeling a thing, but that invisible zap can permanently damage a CPU or RAM stick. It's one of the oldest and most important safety habits in PC building — cheap insurance for expensive parts.Continue A newly built PC powers on, fans spin, but there's no display output. What is the MOST common first thing to check?AReplace the CPU thermal pasteBEnsure the monitor is plugged into the GPU, not the motherboardCReflash the BIOS using a USB driveDCheck the SATA cables are firmly connectedCorrect! This is arguably the most common rookie mistake in PC building — plugging the monitor into the motherboard's video output when a dedicated GPU is installed. The motherboard's HDMI or DisplayPort is disabled by default when a GPU is present. Always connect your display directly to the graphics card.The most common culprit is having the monitor plugged into the motherboard's video port instead of the dedicated GPU. When a graphics card is installed, most systems disable the motherboard's integrated video outputs automatically. It's such a frequent mistake that it has become a running joke in PC building communities.Continue What is the purpose of thermal paste when installing a CPU cooler?AIt acts as an adhesive to hold the cooler in place permanentlyBIt fills microscopic imperfections between the CPU and cooler to improve heat transferCIt insulates the CPU from electrical interference caused by the coolerDIt lubricates the cooler's fan bearings for quieter operationCorrect! Even finely machined metal surfaces have tiny imperfections and air gaps at the microscopic level. Thermal paste — also called thermal interface material (TIM) — fills those gaps to ensure maximum heat conduction from the CPU to the cooler. Without it, air pockets act as insulation and temperatures can skyrocket dangerously.Thermal paste fills microscopic gaps between the CPU lid and the cooler's base plate. Metal surfaces may look flat and smooth, but at a microscopic scale they're riddled with tiny ridges and valleys that trap air — and air is a terrible heat conductor. A thin, even layer of thermal paste eliminates those gaps and keeps temperatures in check.Continue The ATX motherboard form factor, which became the standard for DIY desktop PCs, was introduced by which company and in what year?AAMD in 1990BASUS in 1993CIntel in 1995DIBM in 1988Correct! Intel introduced the ATX (Advanced Technology Extended) standard in 1995, replacing the older AT form factor. ATX standardized component placement, power supply connectors, and airflow direction — making DIY builds far more practical and interchangeable. Nearly 30 years later, ATX and its derivatives like Micro-ATX and Mini-ITX still dominate the market.ATX was introduced by Intel in 1995. It was a major leap forward from the previous AT standard, defining a common layout for motherboards, cases, and power supplies that made mixing and matching components from different vendors straightforward. That standardization is a huge reason DIY PC building became so accessible.Continue When installing RAM into a motherboard with four slots, where should you install two sticks to enable dual-channel mode on most boards?ASlots 1 and 2 (the first two slots nearest the CPU)BSlots 1 and 3, or slots 2 and 4 (alternating slots)CSlots 3 and 4 (the two slots farthest from the CPU)DAny two adjacent slots work equally wellCorrect! Dual-channel mode requires RAM to be installed in matched pairs on alternating slots — typically A2 and B2, or slots 2 and 4. This allows the memory controller to access both sticks simultaneously, effectively doubling memory bandwidth. Your motherboard manual will show the exact recommended slots, usually color-coded for convenience.To enable dual-channel mode, RAM should go in alternating slots — such as slots 2 and 4, often color-coded on the motherboard. Placing both sticks in adjacent slots (like 1 and 2) forces single-channel operation, which can noticeably reduce performance in memory-intensive tasks. Always check your motherboard manual for the exact recommended configuration.Continue What is 'coil whine' in the context of a newly built gaming PC?AA rattling sound from an improperly secured case fanBA high-pitched noise produced by electromagnetic vibration in inductors under high electrical loadCThe sound of a CPU cooler fan spinning at maximum speed during stress testsDNoise caused by the hard drive's read/write head seeking data rapidlyCorrect! Coil whine is a high-pitched, sometimes whirring or buzzing noise caused by tiny electromagnetic coils (inductors) on a GPU or PSU vibrating at audible frequencies under heavy electrical load. It's technically a defect in manufacturing tolerances but is extremely common and not usually harmful to the component. Ironically, it's often loudest in high-end GPUs under uncapped framerates.Coil whine is that annoying high-pitched squeal coming from inductors on your GPU or power supply vibrating under electrical load. It tends to be loudest when framerates are uncapped or during heavy computational tasks. While alarming to new builders, it's usually harmless — though some manufacturers will replace components with severe coil whine under warranty.See My Score Challenge CompleteYour Score/ 8Thanks for playing!Try Again Multiple models eat up VRAM faster In a real-world pipeline, automation rarely means running one model in isolation. You are usually managing a stack of models resident in memory at once: the main LLM, an embedding model for retrieval, and sometimes a reranker to clean up any messy results. My own setups often involve four or even five models processing data sequentially. When your work involves screenshots or scanned documents, you're also using vision-capable models. Even if you have two agents sharing a single loaded model, you still need enough VRAM to accommodate the largest single context they might generate. The moment more than one process needs to be active, that 16GB of headroom disappears completely. What can 16GB actually do? Surprisingly smart local AI Generally, 8GB runs a 7B model, 12GB handles 14B, and 16GB gets you into the 24B range. For context, a 27B model at Q4_K_M requires about 16 GB for the model alone—before you've ever added anything to context. Realistically, a 14B model is actually pretty smart. I've been using one as a brain for an AI-powered Discord bot, AI home assistant, and an AI researcher, and I haven't had too many problems with hallucinations since I switched to Qwen-3-Embedder-0.6b. If you have 16GB of VRAM but an older card, you should look at Mixture-of-Experts (MoE) architectures. These models activate only a fraction of their parameters per token, giving you larger model quality with the output rate of a much smaller model. A 20B-range MoE (quantized) with a 60K context on my 5070 Ti can deliver more than 100 tokens per second, which is comfortably usable for automation. Additionally, the GDDR7 bandwidth on the 5070 Ti (896 GB/s) ensures that these models feel significantly more responsive than they would on older cards with the same capacity. How to get more out of the card you have If you're sticking with 16GB, you can stretch your resources by being careful about your configuration. You should use KV cache quantization (q8 or q4), which can halve the cache footprint. That is a one-line configuration change in llama.cpp or Ollama. You should also cap your context to exactly what the task requires; just because a model "supports" 128K context doesn't mean you should allocate that much. You should also consider switching to a mixture-of-experts (MoE) model instead of a dense model. MoE models take up as much VRAM as a similarly-sized dense model, but they only activate a subset of their total parameters for any given input. That means they'll be significantly faster. Finally, don't be afraid to use your system RAM. Having 32GB or 64GB of system memory allows you to offload, which keeps larger, more complex tasks running, albeit much more slowly. VRAM is everything VRAM doesn't just determine the size of the model you can load, it also determines how much information you can keep about a given job. 16GB of VRAM with 32GB of system RAM is a very capable gaming setup, but it no longer feels excessive. In fact, I frequently find myself wishing I'd purchased an RTX 5090 before they reached the "Decent used car" price range. However, in the interim, I've taken to just offloading to my CPU when necessary. It still works, it is just slower.