AMD Takes Aim for Workstation AI Market With Radeon AI Pro R9700

In what has quickly become a busy week for workstation and server class video cards at this year’s Computex trade show, AMD has become the latest video card vendor to join the fray with the announcement of their flagship Radeon AI PRO R9700 video card. Based on AMD’s newest RDNA 4 GPU architecture, AMD is looking to break into a new segment of the AI market with their latest generation of professional video cards, offering a do-it-all video card that can be used for everything from professional visualization and CAD/CAM to inference of the latest AI models.

Announced as part of AMD’s Computex 2025 keynote, the Radeon AI Pro R9700 is the first member of AMD’s revised lineup of workstation-class video cards. AMD’s long-lived Radeon Pro W brand has effectively been given a facelift here, with AMD promoting the AI inference potential for the new cards front-and-center via the new Radeon AI Pro branding.

But the significance of the new cards is more than just a branding exercise: thanks to AMD’s RDNA 4 architecture and its vastly improved matrix (tensor) math capabilities, AMD finally has a workstation card performant enough at matrix operations to be competitive with arch-rival NVIDIA, as well as more purpose-built AI inference accelerators. Thus far AMD has tackled the professional AI market from the top with the Instinct series, and recently, from the bottom with the Ryzen AI Max CPUs. So the Radeon AI Pro series offers a third rail of sorts for the company, with a fast PCIe form factor AI accelerator based on their mainstream graphics hardware.

RDNA 4 Architecture For Professionals

Under the hood, the new card is essentially the professional doppelganger of AMD’s Radeon RX 9070 XT, the first RDNA 4-based product from the company. Like the consumer Radeon, the R9700 is based on AMD’s Navi 48 GPU, a monolithic chip built on TSMC’s 4nm process. And also like the consumer Radeon it occupies an important, if somewhat unusual spot as a mid-to-high-end card, following AMD’s decision not to try to chase the high-end GPU market this generation. The net result being that AMD isn’t in a race to offer the fastest card, but rather is in a race to offer the best value proposition – a recurring strategy for AMD’s Radeon Pro lineup.

Besides incorporating AMD’s top GPU for this generation, the R9700’s other big claim to fame is its 32GB of GDDR6 VRAM, twice that of the consumer Radeon card. With AMD pushing AI inference use cases hard – and with those workloads often being memory capacity-bound due to their large parameter counts – the extra memory is a critical feature differentiator between workstation and consumer cards in this generation, even more than it normally is. With Navi 48’s 256-bit memory bus, this means that AMD is using 2GB GDDR6 memory chips in a clamshell configuration, making it the highest amount of memory that Navi 48 can support.

Due to the lack of a true high-end GPU from AMD this generation, this does lead to some awkwardness when comparing the new R9700 to AMD’s previous generation of Radeon Pro W7000 series of cards. Strictly speaking, the outgoing W7900 supported more memory at 48GB, but its lack of high-performance matrix accelerators makes it a far poorer choice overall for AI workloads. Which is why the R9700 is being setup to excel in that space. In practice, then, the R9700 is more of a direct successor to the W7800 than it is the W7900, and AMD’s promotional materials reflect that.

While the RDNA 4 architecture bring several important advancements to the table, including much improved ray tracing performance, for the R9700 AMD is drawing heavily on its AI improvements. The matrix accelerators within RDNA 4 allows for at least twice as many matrix operation FLOPS per compute unit at all but the highest precisions. And, critically, RDNA 4 introduces support for FP8 and BF8 precisions, both of which have become increasingly popular in the AI inference space as a means to pack in more parameters to memory-limited AI inference accelerators. Coupled with support for sparsity – essentially tossing out half of the elements in a matrix – and RDNA 4 can be upwards of 8x faster per CU per clock than RDNA 3 chips were even in formats the latter did support. Real-world performance isn’t quite so simple, of course, but that kind of 4x to 8x improvement in throughput is a huge part of what is finally making this generation of Radeon Pro cards competitive in the AI inference market.

In practice, AMD is promoting an up to 2x performance improvement over their last-gen W7800 card under DeepSeek’s R1 Distill Llama 8B model.

AMD is also banking on a limited degree of multi-GPU scalability here for AI workloads. With the R9700 being a dual-slot PCIe card, it’s possible to pack up to 4 of them in a full-size chassis – and, if you’ll let AMD talk to you into it, pairing well with their recent Ryzen Threadripper 9000 workstation CPUs. This would give a single system 128GB of VRAM to load an AI model in, spread over the 4 video cards. Notably, however, AMD doesn’t have a dedicated, cache-coherent interconnect to band the cards together (ala NVLink), so the scalability of an AI workload will be heavily influenced by how well it runs over PCI Express-connected cards.

AI aspirations aside, the R9700 is also a full-fledged graphics card in the traditional sense as well, supplanting the W7800 in that role. As with its predecessor, AMD has equipped the card with 4 DisplayPort 2.1a outputs, so this isn’t a dead-head AI accelerator card. With AMD opting to focus on the AI capabilities of the card, the company hasn’t published any performance claims for graphics performance over the W7800, but given what we’ve seen with the consumer Radeon cards, it should offer a respectable performance uplift. Professional users may be especially happy with RDNA 4’s ray tracing improvements, though as a lot of workstation software makes more bespoke use of that hardware, software vendors may need to catch up to take full advantage of the hardware.

Competition & Availability

Once AMD’s latest professional Radeon cards hit the market, they’ll be going up against cards from both NVIDIA and Intel, the latter of whom announced their own Arc Pro B-series video cards just this week at Computex. With neither card shipping, AMD isn’t in a position to predict how the W9700 will fare against Intel’s latest offering. But with AMD offering the larger memory capacity at 32GB – The Arc Pro B60 tops out at 24GB – the W7800 is essentially a higher-tier offering than Intel’s new card.

Conversely, AMD will be going up against NVIDIA’s full stack, full court press of RTX Pro 6000 series cards, based on the Blackwell architecture. The R9700 is not intended to go up against NVIDIA’s high-end cards, so from a pure performance perspective, NVIDIA is not under threat there. But where AMD can needle NVIDIA is on costs, especially as AMD has an edge there by using cheaper GDDR6 memory. Based on memory capacity, the R9700 is likely going to eventually be put up against the RTX Pro 4500, though by the same token look for AMD to price it lower.

But how much lower? Well that’s the big question for the moment. For this week’s launch, AMD isn’t announcing pricing for the Radeon AI Pro R9700 – or any of their other new Pro hardware, for that matter – so we don’t know where it will be priced at. The W7800 launched at $2499, so that may be a good starting point, but with the RTX Pro 4500 listing for roughly the same price, more likely AMD will need to aim a bit lower than that here.

In any case, we should have the answer to that in the next month or so. As with the forthcoming Threadripper CPUs, AMD is planning to launch their new Radeon AI Pro video card in July.