Graphics and memory
GPUs: buying guides for India
Graphics cards for deep learning, local LLMs and Stable Diffusion, compared on VRAM, bandwidth, board power and published benchmarks.
By VRAM7 articles
By use8 articles
Brands3 articles
What should you know about GPUs for AI?
An AI GPU is a graphics card used to run or train machine-learning models. What matters most is video memory (VRAM), which decides which models fit, then memory bandwidth, board power and software support. Gaming cards and workstation cards both qualify, and the right pick depends on the workload.
Where should you learn the basics first?
Which comparisons, reviews and builds are in GPUs?
- RTX 5060 Ti 16GB Review for AI WorkReview
- RTX 5090 vs RTX 5080 for AI: 32 GB or 16 GB?Comparison
- RTX 5080 vs RTX 5070 Ti for AI: Is the Step Up Worth It?Comparison
- RTX 5070 Ti vs RX 9070 XT for AI: CUDA vs ROCmComparison
- INNO3D RTX 5090 X3 OC 32GB Review for AIReview
- RTX 5060 Ti 16GB vs RTX 5070 for AI: VRAM or Speed?Comparison
Which other categories are related?
Common questions from buyers in India
Which GPU should you buy for AI in India?
Start with the models you want to run, because VRAM sets what fits. Then compare memory bandwidth, board power and the power supply the maker recommends. Our guides name the maker-listed specs and the benchmark source for every pick.
AI ke liye GPU kaise chunen?
Pehle kaam tay karein: local LLM, image generation ya training. Phir VRAM, bandwidth aur power supply dekhein. Har guide mein maker ke specs aur benchmark ka source diya hota hai.
GPUs buying guide for India
Short answer
Choose a graphics card by memory first, then bandwidth, software support and power. Start with our best GPU for deep learning guide. Then read how much VRAM for an LLM, CUDA vs ROCm for AI and GPU benchmarks for local LLMs. These four cover most buying decisions.
Features that matter
A graphics card (GPU) does the heavy arithmetic in deep learning, local language models and image generation. Six specifications decide how well it suits that work.
| Feature | Why it matters | What to check |
|---|---|---|
| VRAM size | Decides which models load at all | The largest model you plan to run, plus headroom for context |
| Memory type and bandwidth | Decides how fast a loaded model generates text | Bus width and memory type on the maker's page |
| Board power and PSU | Decides what your power supply and case must handle | The maker's recommended PSU wattage and power connectors |
| Tensor core generation | Decides which low-precision formats run fast | The generation NVIDIA or AMD lists for the chip |
| Software support | Decides whether your tools run without workarounds | CUDA or ROCm support for your exact card and operating system |
| Cooling and size | Decides whether the card fits and stays quiet | Length, thickness in slots and the case clearance |
VRAM size. VRAM is the memory on the card. LLM Hardware estimates 6.8 GB for an 8B model at 4-bit. It puts a 27B model at 18.2 GB and a 70B model at 44 GB. Hardware Corner says 16 GB suits models up to about 20B at 4-bit and 24 GB reaches about 36B. Our how much VRAM for an LLM explainer shows the arithmetic. If a model does not fit, the system can spill into slower system memory, and speed drops sharply.
Memory type and bandwidth. Once a model fits, bandwidth sets the speed. Hardware Corner's llama.cpp table shows two 16 GB cards on Qwen3 8B at Q4. The RTX 5070 Ti, with a 256-bit bus, reached 87.54 tokens per second. The RTX 5060 Ti, with a 128-bit bus, reached 51.41. Same memory size, very different speed. Our GPU benchmarks for local LLMs round-up collects such results.
Board power and PSU. Power figures come from the makers. ASUS recommends a 550 W supply for its Dual RTX 5060 Ti 16GB, 750 W for its Prime RTX 5070 Ti and 850 W for its Prime RTX 5080. Check the connector type as well as the wattage. A card that needs a bigger supply raises the total cost of the build.
Tensor core generation. Tensor cores are units that multiply blocks of numbers at low precision. NVIDIA lists 5th-generation tensor cores on its RTX 50 series. The RX 9070 XT has 128 AI accelerators, according to AMD. Newer generations add lower precisions, but they help only when your software uses them. Read tensor cores explained for the details, and what is TOPS for why a bare rating misleads.
Software support (CUDA vs ROCm). CUDA is NVIDIA's platform and ROCm is AMD's. On the PyTorch install selector, the ROCm build is offered for Linux only. Ollama lists RX 9000 series cards for AMD acceleration on Linux, with Vulkan as an option on both systems. AMD's ROCm 10.0.0 notes list Windows 11 support for Radeon. Our CUDA vs ROCm for AI guide has the full picture, and best AMD GPU for AI covers the Radeon range.
Cooling and size. Dimensions vary widely. ASUS lists its Prime RTX 5070 Ti at 304 x 126 x 50 mm. Measure your case before you buy. Small cases need a different class of card, which our best low-profile GPU for AI guide covers.
Common mistakes
- Buying on speed and ignoring memory. A fast card that cannot hold the model is slower than a modest card that can. Work out the model size first.
- Assuming all 16 GB cards behave alike. Bus width and chip class change tokens per second, as the Hardware Corner figures above show. See best 16 GB GPU for AI.
- Skipping the PSU and case check. The maker's recommended wattage, connector and card length decide whether the card works in your PC.
- Not checking tool support for your exact card. Support lists differ by card and operating system. Check the tool's own page before you buy.
- Trusting a bare AI TOPS number. TOPS ratings depend on the number format and conditions, so compare them only within one maker's range.
Where to go next
For language models, read best GPU for local LLM and best high-VRAM GPU for AI. For images, see best GPU for Stable Diffusion and best GPU for AI video upscaling. On a tighter budget, read best budget GPU for machine learning and our RTX 5060 Ti 16GB review.
Frequently asked questions
How much VRAM do I need for AI?
It depends on the model. LLM Hardware estimates 6.8 GB for an 8B model at 4-bit and 18.2 GB for a 27B model.
Is NVIDIA or AMD better for AI?
CUDA is NVIDIA's platform and ROCm is AMD's. Ollama and PyTorch list support for both, with different limits by card and operating system, so check each tool's page for your exact card.
Does memory bandwidth matter as much as VRAM?
Memory size decides whether a model loads. Bandwidth then decides how fast it generates text.
Do I need a bigger power supply for an AI GPU?
Check the card maker's recommendation. ASUS, for example, recommends 550 W for its Dual RTX 5060 Ti 16GB, 750 W for its Prime RTX 5070 Ti and 850 W for its Prime RTX 5080.