AMD Radeon AI PRO R9700: New workstation GPU for Local AI and LLMs with 32 GB memory and RDNA4

At Computex 2025, AMD presented the Radeon AI PRO R9700, a new GPU class for professional AI workloads. The card is explicitly aimed at developers, data scientists and companies that rely on local computing power – without dependence on cloud infrastructures or potentially insecure API access. AMD is thus addressing the increasing demand for edge computing for generative models, multimodal AI and specialized inference.

Architecture and performance: RDNA 4 with 128 AI Accelerators

The Radeon AI PRO R9700 is based on the RDNA 4 architecture, but in a version optimized for AI applications. It combines 128 specialized AI accelerators with a raw performance of up to 1531 TOPS (INT4 Sparse) or 96 TFLOPS at FP16. The card uses 32 GB GDDR6 memory and is designed for PCIe 5.0 x16. The TDP is 300 W. In contrast to classic GPUs, AMD dispenses with gaming-oriented rendering functions in the R9700 and focuses entirely on AI calculation and model inference.

Memory: Why 32 GB is becoming the new standard

A key selling point is the generously dimensioned VRAM: 32 GB GDDR6 enables the local execution of demanding AI models such as Mistral, DeepSeek or Qwen in their 24B to 32B configurations. According to AMD, the typical memory requirement of these models is between 24 and 28 GB – larger LLMs such as GPTQ-compressed 70B models even require over 100 GB, which is addressed by multi-GPU setups. Applications such as SD 3.5 or Flux.1 Fast for text-to-image also occupy more than 17-24 GB – and thus significantly exceed the memory frame of common gaming cards such as the RTX 4080/5080 (16 GB).

Scalability: 4× R9700 for 112 GB usable capacity

In combination with Threadripper PRO and a PCIe 5.0 system, up to four R9700 cards can be operated in parallel. This scales the available VRAM to 112-116 GB, which allows the execution of 123B models such as Mistral Large or LLaMA 70B with full KV cache. The card is therefore also suitable for local fine-tuning, model quantization, RAG scenarios or real-time inference with low latency.

Comparison with the Radeon PRO W7800 and NVIDIA RTX 5080

According to AMD, the R9700 delivers up to twice the performance of the Radeon PRO W7800 (also with 32 GB) in AI workloads such as DeepSeek R1 Distill LLaMA 8B. In direct comparison with the RTX 5080 (16 GB), the performance of large models is even estimated at a factor of up to 4.96. The dedicated AI accelerators of the R9700 in particular are said to offer advantages over classic shaders in highly parallel tasks.

Edge performance and data protection: ROCm and on-premise advantages

Another plus point is the support of ROCm on Radeon, which makes it easier to integrate the card into existing Linux-based AI stacks. Windows support is set to follow shortly. AMD is explicitly promoting the card as an alternative to cloud services, where data sovereignty and control over sensitive content are restricted. The R9700 therefore offers decisive added value, particularly for corporate or research environments with regulatory requirements.

Availability and board partners

The Radeon AI PRO R9700 will be available from July 2025. The official partners include ASRock, ASUS, GIGABYTE, PowerColor, Sapphire, XFX and Yeston. It can be assumed that these partners will each offer factory versions with an adapted cooling concept and optional passive ventilation for rack use.

Conclusion: AMD’s attack on NVIDIA’s AI dominance begins at the edge

With the Radeon AI PRO R9700, AMD is making a serious foray into the market for professional, locally deployed AI GPUs. Thanks to its VRAM potential, dedicated AI hardware and ROCm support, the card is ideally positioned for workstations that need to run LLMs or multi-model applications with high requirements.

Especially in times of growing cloud skepticism and increasing latency and data protection concerns, AMD could become a relevant alternative here.

In short, anyone who trains, injects or integrates LLMs locally will find it hard to avoid the Radeon AI PRO R9700. Provided that the software ecosystems do not lag behind AMD’s progress.