Multi-GPU LLM Setup: Running 70B Models Across Multiple Cards

Multi-GPU LLM Setup: Running 70B Models Across Multiple Cards

Running a 70B model on a single consumer GPU isn't feasible — even an RTX 4090 with 24GB falls short of the 40GB needed for Q4_K_M. Multi-GPU setups distribute model layers across multiple cards, making large models accessible without professional hardware. Here's how to set one up and what to expect.

Best setup for 70B: Two RTX 3090s (24GB each = 48GB total) at ~$1400 combined runs Llama 3.3 70B Q4_K_M at 12-15 tokens/second. For full multi-GPU build guides, see LLMHardware.io's multi-GPU section.

Tensor Parallelism vs Pipeline Parallelism

Tensor parallelism splits individual matrix operations across GPUs — every GPU participates in every layer, requiring high-bandwidth GPU interconnect (NVLink). Pipeline parallelism assigns different layers to different GPUs — each GPU completes its layers before passing to the next. For consumer setups without NVLink, pipeline parallelism via PCIe is more practical and what llama.cpp uses by default with multiple GPUs.

llama.cpp Multi-GPU Setup

llama.cpp automatically distributes layers across all detected GPUs. With two GPUs of different sizes, use --tensor-split 24,16 to allocate proportionally (24GB GPU gets 60%, 16GB GPU gets 40%). This ensures neither GPU runs out of VRAM. The --override-tensor flag gives fine-grained control over which layers land on which device.

PCIe Bandwidth Considerations

Without NVLink, GPUs communicate via PCIe — typically ~16 GB/s for a PCIe 4.0 x16 slot. This creates a bottleneck at layer boundaries. For inference, PCIe bandwidth is usually sufficient — most of the time is spent on memory-bandwidth-limited matrix operations within each GPU. The overhead is typically 15-25% compared to NVLink. Avoid running three or more GPUs in PCIe x8 slots without testing bandwidth first.

Comments

Post a Comment

Popular posts from this blog

AMD RX 9070 XT for Local LLMs: ROCm Performance Guide (2026)

The AMD RX 9070 XT launched in early 2026 at $599 with 16GB GDDR6 and 648 GB/s memory bandwidth. For local LLM users who want a mid-range GPU without NVIDIA prices, this changes the calculus significantly. Here is what it actually delivers. RX 9070 XT Specs vs Competition GPU VRAM Bandwidth Price LLM Score RX 9070 XT 16 GB 648 GB/s $599 ~60 tok/s RTX 4070 Ti Super 16 GB 672 GB/s $799 ~62 tok/s RTX 4080 Super 16 GB 736 GB/s $999 ~68 tok/s RTX 4090 24 GB 1,008 GB/s ~$1,100 ~90 tok/s At $599 with 16GB VRAM, the RX 9070 XT is the best value GPU for local LLMs in the mid-range tier in 2026. See the full comparison at LLMHardware.io's AMD GPU guide . ROCm Setup on Linux (2026) # Ubuntu 22.04/24.04 sudo apt install rocm-hip-libraries hip-base # or use the AMD installer: wget https://repo.radeon.com/amdgpu-install/6.3/ubuntu/jammy/amdgpu-install_6.3.60300-1_all.deb sudo dpkg -i amdgpu-install*.deb sudo amdgpu-install --usecase=rocm,lrt,openclsdk # Verify ROCm sees your GPU r...
Read more

llama.cpp flash attention: how to enable and benchmark

Image
Flash attention computes attention with far less memory traffic. The math is identical; the memory movement is not. In llama.cpp it matters most in two situations: long context windows and tight video memory (VRAM). With flash attention enabled, the key-value cache shrinks substantially. A 32k-token context window becomes realistic on a 16GB card, and the prompt-processing phase speeds up, so long documents get ingested faster. The catch is compatibility: not every model and quantization combination supports it, and on some setups the speedup is modest. If you run short chats on a strong GPU, you may barely notice. Where it shines is retrieval-augmented generation (RAG) and long-document work on mid-range cards. Enable it, compare your tokens per second, and keep it if your workload benefits. For the common case of pasting a long PDF and asking questions, it is close to free performance. See how context length changes VRAM needs .
Read more

SillyTavern with Local LLMs: Complete Setup for AI Companions

Image
SillyTavern is the most feature-rich frontend for character-based AI conversations. What began as an AI roleplay tool has evolved into a powerful platform with persona management, memory systems, extensive prompt engineering tools, and connections to every major local inference backend. Here's how to set it up with local models. Quick connect: SillyTavern connects to Ollama, KoboldCpp, llama.cpp server, and TabbyAPI (ExLlamaV2) out of the box. Any of the hardware setups at LLMHardware.io provides a solid backend for SillyTavern. Installation Install via git: git clone https://github.com/SillyTavern/SillyTavern && cd SillyTavern && npm install then node server.js . Access at localhost:8000. For persistent installation, use the included start.sh script. SillyTavern uses Node.js and runs on any platform. The companion app (STMP) enables mobile access to your SillyTavern instance via Tailscale or local network. Character Cards and the TavernAI Format Chara...
Read more