LLM GPU VRAM Requirements Guide (2026): Sizing & Specs

Calculate exact LLM GPU VRAM requirements for inference. See VRAM math formulas, FP16 vs FP8 vs INT4 trade-offs, and cloud GPU memory selection rules.

On this page

Understanding llm gpu vram requirements is the single most critical step when planning open-weight model deployment. Running out of video memory results in Out-Of-Memory (OOM) crashes or severe performance drops when execution spills over to CPU system RAM.

GPU Picks aggregates hardware specifications, provider pricing, and architectural memory footprints according to our strict editorial methodology. This guide outlines the mathematical formulas used to calculate VRAM capacity for Large Language Models (LLMs), breaks down quantization precision trade-offs, and maps modern open-weight architectures (such as DeepSeek V4/R1, Qwen 3.7, GLM-5.2, and Llama 4) to specific GPU cloud configurations.

Key takeaways

  • Full FP16 precision requires approximately 2GB of VRAM per 1 billion parameters.[source] FP8 precision halves this to 1GB per billion,[source] while 4-bit quantization (INT4/GGUF/AWQ) reduces model weight footprints to ~0.55GB per billion parameters.[source]
  • Memory sizing must account for the Key-Value (KV) cache, which grows linearly with context length and batch size.[source] A 128K context window can require 20GB to 80GB of additional VRAM beyond raw model weights.[source]
  • Mixture of Experts (MoE) models like DeepSeek V3/R1 use active parameter routing per token (37B active out of 671B total), but the entire 671B weight set must remain resident in VRAM to maintain real-time throughput.[source]
  • Single 24GB GPUs run quantized 32B models. Enterprise 70B models require dual 48GB or 80GB cards, while 671B MoE models require multi-node clusters with 8x H100 80GB or 4x H200 141GB GPUs.[source]

How to calculate LLM GPU VRAM requirements

To estimate the minimum VRAM required to load and serve an open-weight LLM for inference, use the three-part calculation:

  1. Model Weight Memory: Multiply model parameters by the bytes per parameter based on precision format.[source]
  2. KV Cache Overhead: Calculate memory allocated to maintain user conversation history across active context windows.[source]
  3. CUDA and Framework Overhead: Add a 20 percent buffer for PyTorch/CUDA runtime allocations, activation memory, and serving engine buffers.[source]
Total VRAM (GB)=[(Params in Billions×Bits per Weight8)+KV Cache (GB)]×1.20text{Total VRAM (GB)} = \left[ left( \text{Params in Billions} times \frac{text{Bits per Weight}}{8} \right) + text{KV Cache (GB)} \right] times 1.20</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="katex-base"><span class="katex-strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord text"><span class="mord">Total VRAM (GB)</span></span><span class="mspace" style="margin-right:0.2778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2778em;"></span></span><span class="katex-base"><span class="katex-strut" style="height:2.4em;vertical-align:-0.95em;"></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size3">[</span></span><span class="minner"><span class="mopen delimcenter" style="top:0em;"><span class="delimsizing size3">(</span></span><span class="mord text"><span class="mord">Params in Billions</span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord"><span class="mopen nulldelimiter"></span><span class="mfrac"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:1.3714em;"><span style="top:-2.314em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord">8</span></span></span><span style="top:-3.23em;"><span class="pstrut" style="height:3em;"></span><span class="frac-line" style="border-bottom-width:0.04em;"></span></span><span style="top:-3.677em;"><span class="pstrut" style="height:3em;"></span><span class="mord"><span class="mord text"><span class="mord">Bits per Weight</span></span></span></span></span><span class="vlist-s">​</span></span><span class="vlist-r"><span class="vlist" style="height:0.686em;"><span></span></span></span></span></span><span class="mclose nulldelimiter"></span></span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size3">)</span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mord text"><span class="mord">KV Cache (GB)</span></span><span class="mclose delimcenter" style="top:0em;"><span class="delimsizing size3">]</span></span></span><span class="mspace" style="margin-right:0.2222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222em;"></span></span><span class="katex-base"><span class="katex-strut" style="height:0.6444em;"></span><span class="mord">1.20</span></span></span></span></span></div> <hr /> <h2>Core components of LLM memory allocation</h2> <p>GPU video RAM is split into three main allocations during inference:</p> <h3>1. Model weights memory</h3> <p>Model weights occupy a static footprint in VRAM as soon as the model initializes.<sup class="citation"><a href="#source-huggingface-memory-calc" aria-label="Citation: Accelerate Model Memory Usage Estimator by Hugging Face">[source]</a></sup> The precision format determines how many bytes each weight parameter consumes:</p> <ul> <li><strong>FP16 / BF16 (16-bit)</strong>: Consumes 2 bytes per parameter (16 bits / 8 bits = 2 bytes).<sup class="citation"><a href="#source-huggingface-memory-calc" aria-label="Citation: Accelerate Model Memory Usage Estimator by Hugging Face">[source]</a></sup> A 70B parameter model in FP16 consumes 140GB of VRAM for weights alone.<sup class="citation"><a href="#source-qwen25-huggingface-docs" aria-label="Citation: Qwen 2.5 Technical Specification and Model Cards by Alibaba Cloud / Hugging Face">[source]</a></sup></li> <li><strong>FP8 / INT8 (8-bit)</strong>: Consumes 1 byte per parameter.<sup class="citation"><a href="#source-deepseek-v3-tech-report" aria-label="Citation: DeepSeek-V3 &amp; R1 Technical Report by DeepSeek AI">[source]</a></sup> The same 70B model in FP8 consumes 70GB of VRAM.<sup class="citation"><a href="#source-deepseek-v3-tech-report" aria-label="Citation: DeepSeek-V3 &amp; R1 Technical Report by DeepSeek AI">[source]</a></sup></li> <li><strong>INT4 / AWQ / GGUF (4-bit)</strong>: Consumes approximately 0.5 to 0.55 bytes per parameter after accounting for quantization scales.<sup class="citation"><a href="#source-huggingface-memory-calc" aria-label="Citation: Accelerate Model Memory Usage Estimator by Hugging Face">[source]</a></sup> A 70B model in INT4 consumes ~38.5GB of VRAM.<sup class="citation"><a href="#source-huggingface-memory-calc" aria-label="Citation: Accelerate Model Memory Usage Estimator by Hugging Face">[source]</a></sup></li> </ul> <h3>2. Key-Value (KV) cache memory</h3> <p>The KV cache stores key and value attention vectors for generated and prompt tokens so the model does not recalculate previous tokens at every step.<sup class="citation"><a href="#source-vllm-memory-docs" aria-label="Citation: vLLM Engine Arguments and Memory Management by vLLM Project Documentation">[source]</a></sup> KV cache memory is calculated as:</p> <div class="math-display my-6 overflow-x-auto text-center p-4 rounded-lg bg-surface-card border border-line"><span class="katex-display"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML" display="block"><semantics><mrow><mtext>KV Cache Bytes</mtext><mo>=</mo><mn>2</mn><mo>×</mo><mtext>Layers</mtext><mo>×</mo><mtext>Attention Heads</mtext><mo>×</mo><mtext>Head Dimension</mtext><mo>×</mo><mtext>Context Length</mtext><mo>×</mo><mtext>Batch Size</mtext><mo>×</mo><mtext>Bytes per Element</mtext></mrow><annotation encoding="application/x-tex">\text{KV Cache Bytes} = 2 times \text{Layers} times \text{Attention Heads} times \text{Head Dimension} times \text{Context Length} times \text{Batch Size} times \text{Bytes per Element}

For long-context models supporting 128K to 1M tokens, KV cache usage often equals or exceeds the memory required for the model weights.[source] Serving engines like vLLM optimize this using PagedAttention to prevent memory fragmentation.[source]

3. Activation and CUDA runtime overhead

When serving inputs, activation tensors created between transformer layers require additional memory.[source] PyTorch and CUDA runtime context buffers allocate an additional 1GB to 3GB of VRAM upon initialization.[source] Adding a 20 percent buffer onto raw weight and cache totals prevents sudden OOM crashes under peak concurrency.[source]

To optimize deployment costs across different hardware setups, review our strategies for how to optimize GPU cloud costs.


2026 SOTA model VRAM requirements matrix

The table below outlines real-world VRAM requirements for prominent open-weight models across FP16, FP8, and INT4 precision levels.

Model Architecture & Scale Total Parameters Precision Format Base Weight Memory Recommended VRAM (with 32K Context) Minimum GPU Hardware Needed
DeepSeek R1 671B MoE (37B active) FP8 (Native) ~671 GB ~780 GB 8x H100 80GB or 4x H200 141GB
DeepSeek V3 671B MoE (37B active) FP8 (Native) ~671 GB ~780 GB 8x H100 80GB or 4x H200 141GB
DeepSeek-V4-Flash 284B MoE (13B active) FP8 / INT4 ~160-284 GB ~175-310 GB 2x H200 141GB or 2x A100 80GB
Llama 3.1 405B Dense FP16 ~810 GB ~920 GB 8x H100 80GB or 8x A100 80GB
Qwen 2.5-72B 72.7B Dense FP16 ~145 GB ~165 GB 2x A100 80GB or 1x H200 141GB
Qwen 2.5-72B 72.7B Dense INT4 / AWQ ~39 GB ~48 GB 2x RTX 4090 24GB or 1x L40S 48GB
Qwen 2.5-Coder-32B 32.5B Dense FP16 ~65 GB ~75 GB 1x A100 80GB or 1x H100 80GB
Qwen 2.5-Coder-32B 32.5B Dense INT4 / GGUF ~20 GB ~24 GB 1x RTX 4090 24GB
GLM-4.5 / GLM-5.1 130B MoE FP8 ~130 GB ~155 GB 2x A100 80GB or 2x H100 80GB
Llama 3.3 / DeepSeek R1-Distill 70B Dense FP16 ~140 GB ~160 GB 2x A100 80GB or 1x H200 141GB
Llama 3.3 / DeepSeek R1-Distill 70B Dense INT4 / GGUF ~38 GB ~45 GB 2x RTX 4090 24GB or 1x L40S 48GB
Gemma 2 / GLM-4-9B 9B-14B Dense FP16 ~18-28 GB ~22-32 GB 1x RTX 4090 24GB or 1x A10G 24GB

Mixture of Experts (MoE) VRAM specifics: DeepSeek V4 & R1

A common point of confusion when sizing Mixture of Experts (MoE) models is the difference between total parameter count and active parameter count.[source]

For example, DeepSeek V3 and R1 contain 671 billion total parameters, but route only 37 billion active parameters per token through specialized expert sub-networks.[source]

MoE Memory Golden Rule

Active parameters dictate compute latency, but total parameters dictate VRAM footprint.[source] Because any expert can be selected for any token, all 671 billion weights must remain loaded in VRAM. Fetching un-allocated weights over PCIe buses during token generation slows inference significantly.

To serve DeepSeek V3/R1 in native FP8 precision, hosting environments require at least 700GB of VRAM.[source] Teams targeting enterprise inference workloads can read our roundup of the best GPU cloud for inference or evaluate multi-GPU setups on the best H200 GPU cloud.

  • 8x NVIDIA H100 80GB: Provides 640GB of raw VRAM. Running in FP8 requires light 4-bit quantization or tensor parallelism across 8 cards to fit extended context windows.[source]
  • 4x NVIDIA H200 141GB: Provides 564GB of HBM3e VRAM, allowing FP8 deployment across 4 GPUs with reduced inter-node communication.[source]
  • AMD MI300X Clusters: Featuring 192GB HBM3 memory per card, a 4x MI300X node provides 768GB of VRAM, hosting the full FP8 model on a single node.[source]

Quantization trade-offs: FP16 vs FP8 vs INT4

Quantization compresses model weight precision from 16-bit floating point down to 8-bit or 4-bit representations.[source]

FP16 / BF16 (Full precision)

  • Memory Footprint: 2.0GB per 1B parameters.[source]
  • Accuracy Retention: 100 percent baseline accuracy.
  • Best For: Enterprise production deployments, model fine-tuning, and task-critical reasoning where zero perplexity degradation is allowed.[source] Teams building training clusters can consult our guide on the best GPU cloud for ML training.

FP8 (8-bit floating point)

  • Memory Footprint: 1.0GB per 1B parameters.[source]
  • Accuracy Retention: Greater than 99.5 percent baseline accuracy. Supported natively by NVIDIA Hopper (H100/H200) and Ada Lovelace architectures.[source]
  • Best For: Native precision serving for large MoE models like DeepSeek V3/R1.[source]

INT4 / AWQ / GGUF (4-bit quantization)

  • Memory Footprint: Approximately 0.55GB per 1B parameters.[source]
  • Accuracy Retention: 95 to 98 percent baseline accuracy depending on quantization scheme (AWQ, GPTQ, or GGUF K-quants).[source]
  • Best For: Fitting 32B and 70B parameter models onto single or dual consumer GPUs on the best RTX 4090 cloud.[source]

Matching workloads to cloud GPU instances

Selecting the appropriate GPU cloud instance depends on model size, quantization precision, and expected concurrency.

Entry Tier (24GB to 48GB VRAM)

  • Suitable Hardware: NVIDIA RTX 4090 (24GB), RTX 6000 Ada (48GB), L40S (48GB).[source]
  • Compatible Models: Qwen 2.5 32B (INT4), DeepSeek-R1-Distill-32B (INT4), Gemma 4 27B (INT4), GLM-4-9B (FP16).
  • Target Use Cases: Single-user development, local RAG pipelines, internal developer testing.

Mid-Range Tier (80GB to 141GB VRAM)

  • Suitable Hardware: NVIDIA A100 80GB, H100 80GB, H200 141GB.[source][source]
  • Compatible Models: Llama 3.3 70B (FP16), Qwen 2.5 72B (FP16), DeepSeek-R1-Distill-70B (FP16).
  • Target Use Cases: Multi-user production APIs, high-throughput batch inference, long-context document processing.

Enterprise Cluster Tier (320GB to 1,500GB+ VRAM)

  • Suitable Hardware: 8x H100 80GB, 4x H200 141GB, 4x AMD MI300X 192GB.[source]
  • Compatible Models: DeepSeek V4 / R1 671B MoE, Llama 3.1 405B, GLM-5.2 130B MoE.
  • Target Use Cases: Enterprise AI infrastructure, multi-tenant API serving, foundation model fine-tuning.

To compare current cloud provider hourly rates, region availability, and spot discounts across these GPU instances, explore our interactive GPU Cloud Lookup tool.


Frequently asked questions

How much VRAM is needed to run DeepSeek R1 locally?

Running the full 671B parameter DeepSeek R1 model in FP8 precision requires at least 700GB of VRAM, which mandates an enterprise multi-GPU server (such as 8x H100 80GB or 4x H200 141GB).[source][source] However, distilled versions of R1 (such as DeepSeek-R1-Distill-32B in 4-bit GGUF format) run locally on a single 24GB GPU like the RTX 4090.[source]

What happens when an LLM runs out of VRAM?

When VRAM capacity is exceeded, the execution engine either crashes with a CUDA Out-Of-Memory (OOM) error or offloads model layers into system RAM (CPU offloading).[source] CPU offloading avoids crashes but reduces generation speed by 90 to 98 percent due to PCIe bandwidth bottlenecks.

Does context length affect VRAM requirements?

Yes. Every token added to the prompt or generated response stores key-value attention pairs in the KV cache.[source] At context lengths of 64K to 128K tokens, the KV cache can consume more VRAM than the model weights themselves.

Is an RTX 4090 sufficient for 70B models?

A single RTX 4090 (24GB VRAM) cannot run a 70B model.[source] A 70B model in 4-bit INT4 precision requires approximately 38.5GB of VRAM for weights alone, plus 6GB for KV cache, requiring dual 24GB GPUs (48GB total) or a single 48GB GPU like the L40S.[source]

To compare live hourly pricing and specs for GPUs capable of running these models, visit the GPU Cloud Lookup tool or read our guide on the Best GPU Cloud for Inference.

Sources

  1. Accelerate Model Memory Usage Estimator (opens in a new tab) , Hugging Face primary Accessed August 2, 2026
  2. DeepSeek-V3 & R1 Technical Report (opens in a new tab) , DeepSeek AI primary Accessed August 2, 2026
  3. NVIDIA H100 Tensor Core GPU Data Sheet (opens in a new tab) , NVIDIA primary Accessed August 2, 2026
  4. NVIDIA H200 Tensor Core GPU Architecture & Memory Specs (opens in a new tab) , NVIDIA primary Accessed August 2, 2026
  5. Qwen 2.5 Technical Specification and Model Cards (opens in a new tab) , Alibaba Cloud / Hugging Face primary Accessed August 2, 2026
  6. vLLM Engine Arguments and Memory Management (opens in a new tab) , vLLM Project Documentation technical Accessed August 2, 2026

Editor Ahmad Nugraha