Short answer: an SSD stores model files and software, system RAM supports the operating system and CPU-side work, and GPU VRAM is the GPU’s local workspace for model weights, the key-value (KV) cache and other runtime allocations. An SSD cannot substitute for the RAM or VRAM required while a model is running.
Some runtimes can place part of a model in system RAM, and unified-memory platforms expose CPU and GPU memory through a shared address space. Those are specific software or hardware designs—not proof that ordinary RAM and discrete VRAM are interchangeable. The model’s precision or quantization, context length, batch size, backend and other applications determine whether a particular workload fits.
As an Amazon Associate I earn from qualifying purchases.
SSD, system RAM and VRAM have different jobs
Local AI systems usually involve three memory layers. Treating them as one pool leads to wrong capacity estimates and disappointing upgrades.
| Resource | What it does | What it does not do | Typical planning question |
|---|---|---|---|
| SSD | Persists model downloads, applications, container images, caches and multiple model versions when the computer is powered off. | It is not active inference memory. It does not become VRAM merely because a model file is stored on it. | Do I have enough capacity for the models, runtimes and containers I want to keep locally? |
| System RAM | Holds the operating system and applications, supports CPU inference, and can participate in supported CPU/GPU offload or fallback paths. | It is not automatically as fast as a GPU’s local memory, and every backend does not transparently spill into it. | Can the chosen runtime use enough host memory for this model and context without starving the OS? |
| Discrete GPU VRAM | Provides fast, local working memory for GPU-resident weights, KV cache and execution buffers. | Its capacity is not determined by SSD size or total installed system RAM. | Can the GPU keep the intended layers and runtime state in VRAM at the required precision and context? |
| Unified memory | On selected architectures, CPU and GPU can address a shared memory pool or address space. | It does not describe the memory behavior of every desktop with a discrete graphics card. | What capacity, bandwidth and software support does this specific platform provide? |
For hardware and backend choices, NVIDIA’s local-AI overview is a useful starting point, but its examples are platform-specific rather than universal requirements: NVIDIA’s local-AI overview.
#1 Best Overall
- MEET THE NEXT GEN: Consider this a cheat code; Our Samsung 990 PRO Gen4 SSD helps you reach near max performance with lightning-fast speeds; Whether you’re a hardcore gamer or a tech guru, you’ll get power efficiency built for the final boss
- REACH THE NEXT LEVEL: Gen4 steps up with faster transfer speeds and high-performance bandwidth; With a more than 55% improvement in random performance compared to 980 PRO, it’s here for heavy computing and faster loading
- THE FASTEST SSD FROM THE WORLD'S FLASH MEMORY BRAND: The speed you need for any occasion; With read and write speeds up to 7450/6900 MB/s you’ll reach near max performance of PCIe 4.0 powering through for any use
- PLAY WITHOUT LIMITS: Give yourself some space with storage capacities from 1TB to 4TB; Sync all your saves and reign supreme in gaming, video editing, data analysis and more
- IT’S A POWER MOVE: Save the power for your performance; Get power efficiency all while experiencing up to 50% improved performance per watt over the 980 PRO; It makes every move more effective with less consumption
What the SSD contributes to a local AI setup
Persistent model and software storage
When you download a model, its weights are files on the SSD. The drive also stores the inference application, tokenizer data, configuration, logs and, in many workflows, container images. A larger SSD lets you keep several quantized variants or model families without repeatedly deleting and downloading them.
Storage capacity and runtime memory are separate budgets. A 100 GB model file still has to be represented in working memory when loaded, and the runtime may need additional memory for the KV cache and temporary buffers. Conversely, a model can fit in VRAM while the SSD is too full to hold another version or the container required to run it.
Container footprints can be significant
NVIDIA’s current AI Workbench support matrix lists a Windows Docker Desktop PyTorch container example at around 41 GB total. That figure describes the listed container and configuration’s disk footprint, not the model’s live VRAM requirement. The same matrix gives product-specific RAM and storage examples for its supported configurations: AI Workbench Support Matrix.
Does a faster SSD make generation faster?
The sources do not establish that a faster SSD generally increases tokens per second after active model data is resident in RAM or VRAM. An SSD may determine how much local content you can keep and can affect file-management or loading behavior, but it does not add GPU memory. If generation is constrained by VRAM, KV-cache capacity or compute throughput, replacing the drive alone does not remove that constraint.
How system RAM is used
Operating-system and CPU workload
System RAM is the computer’s main memory. The operating system, browser, desktop applications and CPU-side portions of an inference stack all compete for it. A model that technically fits in total RAM can still fail if the OS and other processes leave too little free memory.
Rank #2
- Ideal for high speed, low power storage
- Gen 4x4 NVMe PCle performance
- Up to 6,000MB/s read, 4,000MB/s write
- Includes Acronis cloning software
- 5-year limited warranty
CPU inference and partial offload
CPU-only inference keeps model data in system RAM and uses the processor for computation. Hybrid configurations keep some layers or runtime state in VRAM and place the remainder in host memory. The exact split, supported formats and performance depend on the backend, driver and model implementation.
Documented fallback is not a universal guarantee
llama.cpp documents a CUDA unified-memory option on Linux. Setting GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 allows the runtime to swap to system RAM instead of failing when VRAM is exhausted. Its build and backend documentation also points Windows users to a corresponding System Memory Fallback setting in NVIDIA Control Panel: llama.cpp build and backend documentation.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsThis mechanism proves only that the documented runtime and platform can use host memory in that way. It does not mean every inference backend will spill automatically, nor that RAM access will deliver the same responsiveness or throughput as keeping the data in VRAM. Treat fallback as a compatibility and capacity option, not as free VRAM.
How much system RAM should you install?
There is no model-independent minimum. As one product-specific reference, NVIDIA lists 16 GB as a minimum and 32 GB as recommended system RAM for its full-local-install configurations on Windows and Ubuntu in the current AI Workbench support matrix. Those numbers describe that product’s installation configurations, not a universal local-LLM recommendation. Leave additional headroom for the operating system, applications, CPU offload and the model’s own runtime allocations.
What GPU VRAM holds during inference
Weights are only the first allocation
When a GPU backend loads a model, VRAM may contain the model weights, the KV cache, activations, workspace and other execution buffers. The required capacity therefore depends on the representation (for example, FP16 or a quantized format), context length, batch or concurrency, backend and GPU architecture—not parameter count alone.
Rank #3
- SPEED UP PROJECTS. Launch creator applications fast with uncompromising PCIe 4.0 read speeds up to 7,100MB/s,[2] (1TB and 2TB[1] models) and write speeds up to 6,700MB/s[2] (1TB[1]-4TB[1] models).
- CREATE AND STORE MORE. Make more room for your 4K videos and high-resolution images with capacities from 500GB[1] up to 4TB[1] on M.2 2280 built with our trusted 8th generation SANDISK BiCS QLC 3D CBA NAND.
- IT GOES WHERE YOU GO. With an all-new power efficient design, your drive delivers high performance with low power, giving you more time to be productive while on the go.
- UNCOMPROMISED RELIABILITY. With up to 1,200 TBW[3] (4TB[1] model) endurance rating, your drive is designed for creators.
- KEEP YOUR DRIVE UPDATED. Monitor your SSD’s performance and check for updates with the downloadable SANDISK Dashboard application.[5]
The KV cache grows with context and batch
The KV cache stores attention information for tokens already processed. NVIDIA notes that it grows with context length and batch size. A short single-user prompt and a 128k-token context can have radically different memory footprints even with identical weights.
NVIDIA’s large-scale inference example makes the distinction concrete: its Llama 3 70B FP16 illustration uses about 140 GB for weights and approximately 40 GB for the KV cache for one user at a 128k-token context. These are vendor-published figures for that model, precision, context and workload, not requirements for every quantized build or inference stack: NVIDIA’s KV-cache and CPU-GPU memory-sharing article.
More VRAM generally keeps more work local
Additional VRAM can allow more layers, a larger context or more concurrent sequences to remain on the GPU. It does not guarantee a particular speedup because compute capability, kernel support, quantization and software scheduling also matter. The reviewed sources do not provide a controlled, universal speed ratio for RAM fallback versus all-VRAM execution.
Model-sizing examples with their qualifications attached
The following figures are useful anchors only when their model, precision and software context are kept intact.
| Source and workload | Published estimate | How to interpret it |
|---|---|---|
| NVIDIA technical article, Llama 3 70B, FP16 weights | About 140 GB for weights | A model-specific FP16 weight estimate; quantization changes the weight footprint. |
| NVIDIA technical article, the same Llama 3 70B example at 128k context, one user | About 40 GB KV cache | Illustrates why context and runtime state can add tens of gigabytes beyond weights. |
| NVIDIA technical article, Llama 4 Scout 109B, FP16 | About 218 GB | A model-specific FP16 estimate from NVIDIA’s example, not a universal requirement. |
| NVIDIA NIM 1.7.0 rough guidance, Llama 8B | About 15 GB | Approximate guidance for that NIM version; actual needs may be lower or higher with different hardware or configuration. |
| NVIDIA NIM 1.7.0 rough guidance, Llama 70B | About 131 GB | Approximate NIM guidance, not an exact formula for another backend, quantization or context. |
NVIDIA explicitly cautions that its NIM figures can vary with hardware and NIM configuration. Read the complete versioned guidance before treating a number as a deployment target: NVIDIA NIM 1.7.0 getting-started documentation.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Rank #4
- HUGE SPEED BOOST: Get random read/write speeds that are 40%/55% faster than 980 PRO; Experience up to 1400K/1550K IOPS, while sequential read/write speeds up to 7,450/6,900 MB/s reach near the max performance of PCIe 4.0*
- BREAKTHROUGH POWER EFFICIENCY: Use less power and get more performance; Enjoy up to 50% improved performance per watt over 980 PRO, plus optimal power efficiency with max PCIe 4.0 performance**
- SMART THERMAL CONTROL: Samsung's own nickel-coated controller delivers effective thermal control; With its slim size, 990 PRO is a perfect fit for desktops and laptops that meet the PCI-SIG D8 standard***
- THE CHAMPION MAKER: Up to 65% improvement in random performance enables faster loads for an ultimate gaming experience on PS5 and DirectStorage PC games****
- SAMSUNG MAGICIAN SOFTWARE: Get the most out of your SSD with Samsung Magician's advanced yet intuitive optimization tools; Monitor drive health, protect valuable data, and receive important updates for your 990 PRO
Unified memory is a different architecture
Shared address space is not the same as a typical PC
Some systems present CPU and GPU memory through a shared address space or pool. NVIDIA describes the GH200 platform as combining 96 GB of GPU memory with 480 GB of CPU LPDDR memory in one address space through NVLink-C2C. That design can make large datasets accessible to both processors, but its capacity and bandwidth characteristics should not be generalized to a desktop with a discrete GPU and separate system RAM.
Capacity still has to be interpreted by platform
NVIDIA’s current local-AI overview also lists systems it describes as having unified memory, including DGX Spark configurations up to 128 GB. “Up to 128 GB” in that product overview is a platform specification, not evidence that every computer with 128 GB of ordinary RAM offers the same GPU performance or software behavior. Check the architecture, memory bandwidth, driver support and runtime documentation for the exact system.
How to decide whether a local workload will fit
Use a workload description rather than a single parameter count. The following sequence keeps each relevant variable visible.
- Name the model and representation. Record the exact model family, parameter variant, precision or quantization and any mixture-of-experts behavior. An FP16 build and a quantized build of the same family do not have the same weight footprint.
- Set the context and batch target. Write down the maximum context length, number of simultaneous users or sequences and expected batch size. These settings drive KV-cache demand.
- Identify the backend and platform. Confirm the operating system, GPU runtime, driver, inference engine and whether CPU fallback or partial offload is documented. Do not assume a feature from one backend exists in another.
- Budget VRAM for more than weights. Reserve room for KV cache, activations, workspace and other processes. A model whose weight file is just below the card’s advertised capacity can still fail at startup or when the context grows.
- Budget system RAM separately. Account for the OS, desktop applications, CPU-side runtime and any layers or cache the backend may place in host memory. Product guidance such as AI Workbench’s 16 GB minimum and 32 GB recommendation applies only to its stated configurations.
- Budget SSD capacity. Add model files, alternate quantizations, tokenizer data, application installations and container images. The AI Workbench example of roughly 41 GB for one Windows Docker Desktop PyTorch container shows why software storage can become a separate constraint.
- Test the real maximum. Load the intended context and concurrency, not only a short prompt. Monitor VRAM and RAM while the model is generating; a configuration that starts successfully may exceed memory when the KV cache expands.
Choosing between common system designs
| Design | Strength | Constraint to verify | Best fit |
|---|---|---|---|
| Discrete GPU with ample VRAM | Keeps more weights and runtime state in local GPU memory. | VRAM must cover representation, KV cache and workspace at the target context and batch. | GPU inference where responsiveness or throughput matters. |
| Discrete GPU plus host-RAM offload | Can run workloads larger than the card’s VRAM when the backend supports partitioning or fallback. | Support is backend-specific; host-memory traffic can change responsiveness and throughput. | Making a larger model feasible when “runs” matters more than all-GPU performance. |
| CPU-only system | Uses system RAM rather than requiring a discrete GPU. | RAM capacity and CPU compute become the limits; GPU acceleration is unavailable. | Smaller or quantized models, development and workloads without a suitable GPU. |
| Unified-memory platform | CPU and GPU can access a shared pool or address space designed by the platform vendor. | Capacity, bandwidth and supported software are specific to that architecture. | Systems built expressly for shared-memory AI workloads. |
| Large SSD with limited RAM/VRAM | Stores many models and containers locally. | Storage does not solve an inference-memory shortfall. | Model-library capacity, not larger active workloads. |
Practical fallback setup and its limits
Linux with llama.cpp and CUDA unified memory
If you are using a CUDA build of llama.cpp on Linux, set the documented environment variable before launching the program:
Free tools Windows power users keep installed
One-click scans. No signup required.
export GGML_CUDA_ENABLE_UNIFIED_MEMORY=1Then start the model with the normal command for your build. The setting permits swapping to system RAM when VRAM is exhausted instead of immediately failing. It does not increase physical VRAM, and the resulting performance depends on the system’s memory path and workload.
Best Value
- This product has been replaced by our latest generation. Please search for the SANDISK Optimus GX 7100 NVMe SSD
- HIGH-OCTANE GAMING. Experience speeds up to 7,250MB/s read and 6,900MB/s write (1-2TB models), with up to 35% faster performance than previous generation.
- PURPOSE-BUILT. Designed for serious on-the-go gamers, with a PCIe Gen4 interface and SANDISK’s next generation TLC 3D NAND.
- MORE TIME TO CLEAR THAT CHECKPOINT. Built with laptops and handheld gaming devices in mind, with up to 100% more power efficiency over the previous generation.
- DO MORE WITH DASHBOARD. Ensure your drive is optimized for prime performance with the downloadable WD_BLACK Dashboard (Windows only).
Windows NVIDIA Control Panel
The same llama.cpp documentation identifies a corresponding System Memory Fallback setting in NVIDIA Control Panel. The exact availability and behavior depend on the driver, GPU and application. Enable it only when your backend and workload are known to use that path; a generic Windows setting cannot make an unsupported runtime offload safely.
Partial offload versus complete offload
Some runtimes let you choose how many layers reside on the GPU. More GPU-resident layers can reduce host-memory traffic, while moving more layers to RAM can make a larger model load. The useful comparison is not “RAM or VRAM” in the abstract, but the same model, representation, context and backend at different offload splits.
Common failure symptoms and fixes
| Symptom | Likely cause | What to check or change |
|---|---|---|
| Model fails before generation with an out-of-memory error | Weights plus runtime allocations exceed available VRAM, or another process is using the card. | Close GPU-using applications, reduce context or batch, use a smaller/quantized representation, or select a documented offload path. |
| Model loads, then fails on long prompts | KV-cache growth at the requested context exceeds VRAM or host-memory headroom. | Lower maximum context, reduce concurrency, use a cache format supported by the backend, or provide more appropriate memory. |
| Generation becomes markedly less responsive after fallback | Layers or cache are moving through system memory rather than staying in VRAM. | Reduce the offloaded portion, lower context or batch, or use a GPU with more VRAM. Do not expect SSD capacity to fix this. |
| Operating system starts swapping | System RAM is being consumed by the model, applications and the OS. | Leave headroom, close applications, reduce CPU offload or choose a smaller representation. |
| Model or container cannot be downloaded | SSD capacity is exhausted even though RAM and VRAM are adequate. | Remove unused versions, relocate model directories, or expand storage with an interface and capacity supported by the computer. |
| A unified-memory claim seems inconsistent with a desktop GPU | A platform-specific shared-memory design is being compared with separate VRAM and RAM. | Check the exact architecture, address-space design, bandwidth and software support rather than total gigabytes alone. |
What to upgrade first
Upgrade the SSD for library and software capacity
Choose more SSD capacity when the problem is keeping multiple model files, quantizations, applications or containers. Match the drive’s physical format and interface to the computer. This upgrade helps storage management; it does not increase available VRAM or guarantee faster generation.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows 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 reinstallUpgrade system RAM for CPU work or supported offload
More RAM is useful when the operating system is under pressure, when you run CPU inference, or when your backend explicitly uses host memory. The required amount depends on model representation, context, other applications and the runtime. No source here establishes a universal memory kit, form factor or capacity for every desktop or laptop.
Upgrade VRAM for GPU-resident workloads
More VRAM directly increases the GPU-local budget for weights, KV cache and execution buffers. Select it against the exact model representation and context target, while also checking compute capability, driver support and backend compatibility. A card with more gigabytes is not automatically faster if the workload is limited by another component.
Bottom line for local AI planning
Keep three budgets on your worksheet: SSD space for files, system RAM for the operating system and any CPU-side or offloaded work, and VRAM for GPU-resident weights and runtime state. Then add the architecture question: is this a normal discrete-GPU system, a documented fallback configuration, or a platform with genuine unified memory?
Use model-specific figures only with their precision, context, batch and software qualifications. NVIDIA’s examples show why a 70B model’s weights are not the whole story, while llama.cpp documents one way to use host memory when VRAM runs out. Those facts help you test a real configuration; they do not support a universal minimum or a promise that an SSD, RAM upgrade or fallback setting will deliver all-GPU performance.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
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.
