Accelerate AI Computing with AMD Radeon™ AI PRO R9700 Creator 32GB
Designed for the next generation of AI computing, the AMD Radeon™ AI PRO R9700 Creator 32GB is purpose-built for professional workstations with 2 to 4 GPU configurations. Its optimized 2-slot blower-type design efficiently expels hot air directly outside the chassis, minimizing thermal interference between adjacent GPUs and ensuring consistent performance under demanding AI workloads.
Powered by 32GB of high-capacity GDDR6 memory per card, the AMD Radeon™ AI PRO R9700 Creator 32GB provides the memory headroom required for running large language models, generative AI applications, and complex machine learning workflows locally. With support for PCIe® 5.0, multiple GPUs can communicate at high bandwidth, enabling faster data movement and improved scalability for model training and AI inference.
Engineered for reliability, the card features an advanced Vapor Chamber cooling system, premium Honeywell PTM7950 thermal interface material, and a reinforced metal construction to deliver exceptional thermal performance and long-term durability—even during continuous, high-intensity computing.
Whether you're building a local LLM platform, an AI development server, or a professional content creation workstation, the AMD Radeon™ AI PRO R9700 Creator 32GB delivers the cooling efficiency, expandability, and stability that modern multi-GPU AI systems demand. Scale confidently from two to four GPUs and unlock the full potential of your AI workstation.
Performance｜Real Throughput, Running Fully Local
Two things decide how a card feels for local AI: how fast it answers a single user, and how many users it can serve at once. With a large model kept fully resident in 32GB of VRAM, the AMD Radeon™ AI PRO R9700 Creator 32GB delivers on both — here running OpenAI's open-weight gpt-oss-20b under vLLM on the AMD ROCm™ stack. 1
Responsive for a single user.
Serving one request at a time, a single AMD Radeon™ AI PRO R9700 Creator 32GB generates ~60 tokens/s — well past the ~30–40 token/s that already reads as instant on screen, with time-to-first-token around 37 ms. Add a second card in tensor-parallel and single-user generation rises to ~117 tokens/s while per-token latency nearly halves (TPOT ~16.5 ms → ~8.4 ms).
Built to serve many.
This is where the card earns its keep. Under vLLM's continuous batching, a single AMD Radeon™ AI PRO R9700 Creator 32GB sustains a peak of ~1494 tokens/s of generation throughput across concurrent requests (~13400 token/s total, including input processing). A second card scales that almost linearly to ~3057 tokens/s generated (~27500 token/s total) — roughly 2x the single-card result.
Every model stays in VRAM.
Throughout these runs the model and its KV cache sit entirely on-card: the AMD Radeon™ AI PRO R9700 Creator 32GB reports ~98% GPU utilization at ~300W with ~30GB of its 32GB in use — no spillover to system memory, no throughput cliff.
Compute｜An RDNA 4 Architecture Built for AI Inference
At the heart of the AMD Radeon™ AI PRO R9700 GPU is AMD's RDNA™ 4 architecture, built on a 4nm process with 64 compute units and 4096 stream processors. What makes it an AI-first card is its 128 second-generation AI Accelerators — dedicated engines for the heavy matrix math of modern inference.
It supports the precision formats today's models actually use — FP16, FP8, INT8, and INT4 — with structured sparsity to push throughput further. In practice, that means real responsiveness when serving quantized LLMs and diffusion models.
Capacity｜How Big a Model Can You Actually Run?
This is where the AMD Radeon™ AI PRO R9700 Creator 32GB pulls ahead. Its 32GB of GDDR6 runs at 20 Gbps on a 256-bit interface, delivering 640 GB/s of bandwidth. That large, fast memory pool is the difference between a model that fits and runs smoothly and one that barely loads and then crawls.
When a model fits fully in VRAM, the GPU can keep accessing every layer of weights at 640 GB/s. But the moment a model exceeds available memory and layers are offloaded to system RAM, access speed drops from VRAM-class down to PCIe and system-memory speed — and throughput falls off a cliff. That 32GB of headroom is precisely what keeps the whole model resident in VRAM and avoids the cliff.
For developers, that headroom means longer context windows, larger batch sizes, and models that simply won't load on smaller cards — all fully local, no cloud required.
Scale｜Turn a Single Workstation into an Inference Array
For a serious AI workstation, one GPU is rarely the end of the story. The AMD Radeon™ AI PRO R9700 Creator 32GB is engineered from the ground up for dense, multi-card deployment. Its compact 2-slot form factor and blower fan cooler drive air front-to-back and straight out of the chassis — unlike consumer axial-fan designs, which recirculate hot air and quickly throttle when cards are stacked close together.
ASRock's blower fan cooling solution of the AMD Radeon™ AI PRO R9700 Creator 32GB backs that airflow with serious hardware: a full vapor chamber for lateral heat spreading, a Honeywell PTM7950 phase-change thermal interface for stable long-run temperatures, a die-cast metal shroud, and a metal backplate built on Super Alloy components. With PCIe® 5.0 enabling fast card-to-card transfers, scaling a single workstation into a 2- or 4-card array is straightforward.
.png)
Ecosystem｜Open AMD ROCm™, Not a Walled Garden
Hardware is only half the equation. The AMD Radeon™ AI PRO R9700 Creator 32GB is fully supported by the AMD ROCm™ open software platform — giving developers a scalable, open environment rather than a single-vendor locked-in ecosystem.
AMD ROCm™ natively supports the frameworks teams already use — PyTorch, ONNX Runtime, and TensorFlow — alongside modern high-throughput inference servers. Because everything runs on-premises, teams gain not just performance and lower latency, but full control over sensitive data that never has to leave the building.
.png)
Who It's For
- • AI developers and data scientists — building local LLM and multimodal pipelines, free from cloud costs and data-exfiltration concerns.
- • System integrators and studios — needing scalable 2–4 card workstations that deliver high-throughput local inference.
- • Engineering teams that value open standards — requiring data to stay in-house and refusing to be locked into a closed ecosystem.
Learn more about ASRock AMD Radeon AI PRO R9700 Creator 32GB:
Learn more about ASRock AMD Radeon AI PRO R9000 Series:
https://www.asrock.com/microsite/AMD_Radeon_AI_PRO_R9000_Series/
¹ Performance measured by ASRock. Model: OpenAI gpt-oss-20b served with vLLM 0.14.0 on AMD ROCm™ 7.2.0; tensor-parallel 1 (single card) and 2 (dual card) across ASRock AMD Radeon AI PRO R9700 Creator 32GB cards (gfx1201 / RDNA™ 4). PCIe® link width: single card at x16; dual-card configuration at x8/x8 (platform lane split when both slots are populated). Workload: 1024-token input / 128-token output, request concurrency swept 1–200 with peak sustained throughput reported; max model length 2048; gpu-memory-utilization 0.9. Test system: ASRock X870E Taichi OCF, AMD Ryzen™ 5 7400, 64 GB RAM, Ubuntu 24.04.4 LTS. Performance varies by model, configuration, and workload.
