# GPU guides and engineering notes

> Guides on GPU memory, LLM inference and fine-tuning, GPU pricing and hardware.

Source: https://overgpu.com/blog  
Publisher: OverGPU (https://overgpu.com/)

## Articles

- [AMD MI355X for AI: specs, ROCm and the workloads it suits](https://overgpu.com/blog/amd-mi355x-for-ai): AMD MI355X for AI, explained: 288 GB of HBM3e at 8 TB/s, FP8 and MXFP4, the real state of ROCm, and the workloads that benefit most. Read before you buy.
- [Best GPU for LLM inference in 2026: a sizing guide by model size](https://overgpu.com/blog/best-gpu-for-llm-inference): The best GPU for LLM inference depends on model size, precision, KV cache and bandwidth. Use our 2026 sizing table, from RTX 5090 to 8× B300, to choose.
- [Dedicated vs shared GPU: what changes when you share](https://overgpu.com/blog/dedicated-vs-shared-gpus): Dedicated vs shared GPU: how vGPU, MIG, time-slicing and shared hosts affect memory, consistency, security and debugging, and when sharing is fine.
- [DeepSeek R1 hardware requirements: GPUs and memory for R1 and V3](https://overgpu.com/blog/deepseek-hardware-requirements): DeepSeek R1 hardware requirements: about 671 GB of FP8 weights plus KV cache. See which 8-GPU servers fit R1 and V3, and the GPUs for distilled models.
- [FP8 vs FP4 vs BF16: number formats for AI, explained](https://overgpu.com/blog/fp8-fp4-precision-explained): FP8 vs FP4 vs BF16 explained: bit layouts, NVFP4 and MXFP4 block scaling, which GPUs support each format, the memory saved and the accuracy you trade.
- [GPU shortage 2026: what buyers face and how to secure GPUs](https://overgpu.com/blog/gpu-shortage-2026): GPU shortage 2026: on-demand stock sells out, providers push 1-year contracts and prices are up. Why it happens, and how teams still secure GPUs.
- [H100 rental cost in 2026: what you actually pay](https://overgpu.com/blog/h100-rental-cost-2026): H100 rental cost in 2026: SXM vs PCIe, monthly vs hourly billing, egress and storage, plus a break-even formula to compare any quote before you sign.
- [How much VRAM do I need for an LLM?](https://overgpu.com/blog/vram-for-llms): How much VRAM does an LLM need? Quick answers for 7B, 14B, 32B, 70B and 405B models at BF16, FP8 and 4-bit, plus KV cache by context and the GPU that fits.
- [How to choose a GPU server for AI](https://overgpu.com/blog/how-to-choose-a-gpu-server): How to choose a GPU server for AI: a framework for memory, bandwidth, precision, interconnect, CPU and storage, software and billing, plus a checklist.
- [How to fine-tune a 70B model: GPU requirements and workflow](https://overgpu.com/blog/fine-tune-70b-model): How many GPUs to fine-tune a 70B model: QLoRA fits one 80 GB GPU, BF16 LoRA about two, full fine-tuning over 1 TB (8x H200 class). Memory math and setup.
- [How to self-host LLM inference with vLLM](https://overgpu.com/blog/self-host-llm-inference-vllm): Self-host LLM inference with vLLM: size the GPU, run an OpenAI-compatible server, tune FP8, prefix caching and parallelism, and secure it for production.
- [NVLink vs PCIe: bandwidth, topology and when it matters](https://overgpu.com/blog/nvlink-vs-pcie): NVLink vs PCIe compared: per-GPU bandwidth, NVSwitch and NVL72, which multi-GPU workloads need NVLink and which run fine on PCIe, with worked numbers.