AMD's Export-Friendly Radeon AI PRO R9700 GPU Prepares for China Debut

[AleksandarK]

According to the latest rumor mill, AMD is preparing Radeon AI PRO R9700, a new GPU designed specifically for the Chinese market. The new Radeon AI PRO R9700 GPU will feature deliberately handicapped performance parameters that comply with the most recent US export regulations while still supporting the local industry's needs for tasks such as on-device inference and model fine-tuning. To adhere to Washington's restrictions, AMD has reduced the chip's compute throughput to remain below the critical threshold for controlled exports, yet it retains 32 GB of high-bandwidth frame buffer memory and full PCIe Gen 5 connectivity. These specifications are intended to ensure efficient data transfer, and when multiple R9700 cards are deployed together, they can deliver substantial processing capability despite the lower power per card.

This China-specific strategy follows a challenging period for AMD in that region, including an $800 million charge related to its MI308 accelerator after it was blocked under earlier rules. NVIDIA has also suffered, losing an estimated $5.5 billion in revenue when its H20 series was barred, and seeing its Chinese market share drop from roughly 90 percent in 2021 to about 50 percent today. These developments have created an opening for AMD to compete on both price and supply reliability. AMD's approach mirrors NVIDIA's release of a restricted-performance version of its Blackwell-architecture B20 GPU for China. Priced more competitively than previous China-compliant offerings, the Radeon AI PRO R9700 will arrive in the third quarter of 2025 alongside NVIDIA's B20. AMD intends to present its full AI silicon roadmap at the Advancing AI Summit on June 12, positioning the R9700 as a targeted solution for enterprises and research institutions rather than for hyperscale cloud environments.



View at TechPowerUp Main Site | Source

 

[john_]

I would never imagine that such a card will need a cut down version for China. I doubt it's in the level of performance of a 5090D or even a 5080, so why does AMD needs a cut down version?

 

[Luisds]

I would never imagine that such a card will need a cut down version for China. I doubt it's in the level of performance of a 5090D or even a 5080, so why does AMD needs a cut down version?

Because China does have an alternative, MTT S4000, which has no quantity limit for a single device.

 

[Red Hood]

The 9700 has about 1500 AI TOPS

I assume the new regulation for China compliant chips is lower than that

 

[Daven]

It’s almost as if they take whatever the top SKU from whatever US company and tell them to cut it down regardless of specs. The 9700 Pro is not even close to the 4090D.

 

[Red Hood]

It’s almost as if they take whatever the top SKU from whatever US company and tell them to cut it down regardless of specs. The 9700 Pro is not even close to the 4090D.

The 4090 doesn't support FP4 so the 9700 Pro is faster in that workload compared to the 4090 running FP8 (about 1300 AI TOPS)

 

[Dr. Dro]

It’s almost as if they take whatever the top SKU from whatever US company and tell them to cut it down regardless of specs. The 9700 Pro is not even close to the 4090D.

I mean, if true, it's a politically motivated move. But I don't see much or any circumstantial evidence regarding this. I doubt there will be a China-specific nerf because the R9700 is not powerful enough to be subject to export controls to begin with. I'm writing this one down in the rumor category.

The R9700 has roughly the same AI TOPS rating of the RX 9070 XT, 1531 FP4 and 766 FP8 TOPS, substantially weaker compared to Nvidia's offerings on the upper gaming/pro-viz and data center segments, so it's really a 9070 XT with twice the memory and access to pro drivers. Speaking of pro drivers, they were released alongside the GPU 9 days ago, so it's way past the "preparing" stage. It's a real product already in the market. However, Windows drivers are not currently available, and reference design GPUs will, at least for the time being, not be sold. It's an OEM-only product release.

 

[bug]

Radeon 9700... Stuff of legends: https://www.techpowerup.com/gpu-specs/radeon-9700.c49

 

[blinnbanir]

I mean, if true, it's a politically motivated move. But I don't see much or any circumstantial evidence regarding this. I doubt there will be a China-specific nerf because the R9700 is not powerful enough to be subject to export controls to begin with. I'm writing this one down in the rumor category.

The R9700 has roughly the same AI TOPS rating of the RX 9070 XT, 1531 FP4 and 766 FP8 TOPS, substantially weaker compared to Nvidia's offerings on the upper gaming/pro-viz and data center segments, so it's really a 9070 XT with twice the memory and access to pro drivers. Speaking of pro drivers, they were released alongside the GPU 9 days ago, so it's way past the "preparing" stage. It's a real product already in the market. However, Windows drivers are not currently available, and reference design GPUs will, at least for the time being, not be sold. It's an OEM-only product release.

View attachment 401856

Any chance you get

 

[Dr. Dro]

Any chance you get

The horror, I actually linked to AMD's website and pointed out to their official product specifications page! Unforgivable, a personal vendetta.

 

[LabRat 891]

It's an OEM-only product release

AIB partner only, right? (like the RX 9070s were?)

So far we've seen, Powercolor, Sapphire, Asrock, and Gigabyte (Aorus Gaming)

Thread 'PowerColor Radeon AI PRO R9700 Card Design Reminiscent of AMD's Reference Renders'

Friday at 3:15 PM

Last week, AMD introduced its Radeon AI PRO R9700 32 GB professional graphics card—representing the RDNA 4 generation's first foray into non-gaming territories. The TechPowerUp team encountered board partner variants on the showroom floor, with GIGABYTE and ASRock exhibiting samples with blower-style cooling solutions. At the time, Team Red's reference design was only visible in promotional/artistic renders. According to the company's official product page, this particular model is "not available for purchase." AMD made similar claims during the early days...

• T0@st
• Replies: 3
• Forum: News

• 

Thread 'Sapphire at Computex 2025: Edge AI Mini PCs, NITRO+ PhantomLink Motherboard, RX 9060 XT'

May 22, 2025

Sapphire brought unique new hardware to Computex 2025. We begin our tour with the new Sapphire Edge AI line of mini PCs. First up, is the Edge AI 370, a mini PC measuring 117 mm x 111 mm x 30 mm (WxDxH). It is powered by AMD Ryzen AI 9 HX 370 "Strix Point" mobile processor, with two DDR5 SO-DIMM slots supporting up to 96 GB of memory. Storage comes from a 1 TB Gen 4 NVMe SSD. The CPU is configured with 12 cores (4x Zen 5 + 8x Zen 5c), while the maxed out RDNA 3.5 iGPU has 16 CU. The XDNA 2 NPU is clocked for 50 AI TOPS, and meets Microsoft Copilot+ requirements.

Next up, is the Sapphire...

• btarunr
• Replies: 12
• Forum: News

• 

G

Thread 'ASRock Unveils New Radeon AI PRO R9700, RX 9060 XT, and ARC Pro B60 Graphics Cards at Computex 2025'

May 21, 2025

ASRock had plenty of new graphics cards at the Computex 2025 show, with a few interesting pieces in its Creator Series, including the new AMD Radeon AI PRO R9700 and Arc Pro B60 graphics cards. In addition, ASRock showcased the newly announced AMD Radeon RX 9060 XT graphics cards that will be available in both Challenger and Steel Legend series.

The ASRock Creator Series is an interesting one as this is aimed at "multi-GPU collaborative computing," meaning it is designed for multi-card computing. This is the reason why the ASRock Creator Series uses a blower-style dual-slot cooler design...

• GFreeman
• Replies: 5
• Forum: News

https://www.aorus.com/en-us/graphics-cards/gv-r9700ai-top-32gd/Key-Features

-with XFX, Asus, and Yeston models remaining to be seen.

Computex 2025: AMD reveals RX 9060 XT, AMD AI PRO GPU and Threadripper 9000 - KitGuru

AMD was saving quite a few announcements for Computex after all. At this year's event, the company r

www.kitguru.net

The final graphics card introduced during today's keynote is the AMD Radeon AI PRO R9700. This card will be available starting in July and board partners like Sapphire, XFX, PowerColor, Gigabyte, ASRock, Asus and Yeston are lined up to release versions of it.

 

[Dr. Dro]

AIB partner only, right? (like the RX 9070s were?)

So far we've seen, Powercolor, Sapphire, Asrock, and Gigabyte

Thread 'PowerColor Radeon AI PRO R9700 Card Design Reminiscent of AMD's Reference Renders'

Friday at 3:15 PM

Last week, AMD introduced its Radeon AI PRO R9700 32 GB professional graphics card—representing the RDNA 4 generation's first foray into non-gaming territories. The TechPowerUp team encountered board partner variants on the showroom floor, with GIGABYTE and ASRock exhibiting samples with blower-style cooling solutions. At the time, Team Red's reference design was only visible in promotional/artistic renders. According to the company's official product page, this particular model is "not available for purchase." AMD made similar claims during the early days...

• T0@st
• Replies: 3
• Forum: News

• 

Thread 'Sapphire at Computex 2025: Edge AI Mini PCs, NITRO+ PhantomLink Motherboard, RX 9060 XT'

May 22, 2025

Sapphire brought unique new hardware to Computex 2025. We begin our tour with the new Sapphire Edge AI line of mini PCs. First up, is the Edge AI 370, a mini PC measuring 117 mm x 111 mm x 30 mm (WxDxH). It is powered by AMD Ryzen AI 9 HX 370 "Strix Point" mobile processor, with two DDR5 SO-DIMM slots supporting up to 96 GB of memory. Storage comes from a 1 TB Gen 4 NVMe SSD. The CPU is configured with 12 cores (4x Zen 5 + 8x Zen 5c), while the maxed out RDNA 3.5 iGPU has 16 CU. The XDNA 2 NPU is clocked for 50 AI TOPS, and meets Microsoft Copilot+ requirements.

Next up, is the Sapphire...

• btarunr
• Replies: 12
• Forum: News

• 

G

Thread 'ASRock Unveils New Radeon AI PRO R9700, RX 9060 XT, and ARC Pro B60 Graphics Cards at Computex 2025'

May 21, 2025

ASRock had plenty of new graphics cards at the Computex 2025 show, with a few interesting pieces in its Creator Series, including the new AMD Radeon AI PRO R9700 and Arc Pro B60 graphics cards. In addition, ASRock showcased the newly announced AMD Radeon RX 9060 XT graphics cards that will be available in both Challenger and Steel Legend series.

The ASRock Creator Series is an interesting one as this is aimed at "multi-GPU collaborative computing," meaning it is designed for multi-card computing. This is the reason why the ASRock Creator Series uses a blower-style dual-slot cooler design...

• GFreeman
• Replies: 5
• Forum: News

https://www.aorus.com/en-us/graphics-cards/gv-r9700ai-top-32gd/Key-Features

-with XFX, Asus, and Yeston models remaining to be seen.

Computex 2025: AMD reveals RX 9060 XT, AMD AI PRO GPU and Threadripper 9000 - KitGuru

AMD was saving quite a few announcements for Computex after all. At this year's event, the company r

www.kitguru.net

AMD's site states that no reference design board is being provided and that the imaging are artistic renders only, but I believe it will be launched on the retail channel sometime in July, that's what their presentation at Computex said anyway. There's no specific date or confirmed pricing yet, though with drivers, official specifications and related documentation provided, to me at least, is enough to count a product as "launched". If it's priced at around the RTX 5080 (up to $1,2k IMHO, maybe even 1,5k for enthusiasts and AMD diehards - as a former Vega Frontier owner, miss me with getting bit with that again), this will be an amazing card to have, even for gamers. Anything more and you should just buy an RTX 5090 which is twice as powerful at compute/AI and 90% faster at gaming.

I'd be surprised if XFX, Yeston and ASUS actually release AIB versions of Pro cards, it'd be a first (and a quite welcome first so far). AsRock and Gigabyte had AI market products from the last generation (like the AsRock Creator with 12VHPWR connector and the W7900 AI TOP) and I think PowerColor is just entering that market, makes sense they'd use a reference-ish board if they are targeting exclusively the AI/edge market.

 

[igormp]

The 9700 has about 1500 AI TOPS

I assume the new regulation for China compliant chips is lower than that

The 4090 doesn't support FP4 so the 9700 Pro is faster in that workload compared to the 4090 running FP8 (about 1300 AI TOPS)

Theoretically, given that no one has managed to achieve that raw perf number given the poor software stack.

Also, TOPs as a metric is usually meant for INT4, which the 4090 does support and manages 2600 TOPs with sparsity. Don't conflict that with TFLOPs.

As per this:

The performance metric used is known as Total Processing Power (TPP) which is calculated by the maximum compute for a given bit-length using TFLOPs or TOPS multiplied by the number of bits. The maximum threshold allowed by the U.S. export regulations is 4,800. It just so happened that the RTX 4090's performance level in this benchmark is 10% higher (5,286) than the regulation limit, which is how the 4090 ended up on the ban list.

This is about the dense numbers, which the 4090 does at 1300 TOPs. 1300*4 = 5200, which is above the 4800 limit.
We can then assume that 4090D should have less than 1200 INT4 dense TOPs.

The 9070xt (and by extension the Pro 9700 since they're pretty much the same) achieves 1557 INT4 TOPs with sparsity, so ~780 TOPs dense, that's way less than the 4090D and way bellow the sanction limits, so I don't get why they had to reduce it further.

Could it be not about the ban restriction, but about that limit where companies need to notify such exports instead?

Couldn't be related to memory bandwidth either, given that this GPU has a pretty puny bandwidth and AMD did not change it for this new China-exclusive.

 

[Dr. Dro]

Theoretically, given that no one has managed to achieve that raw perf number given the poor software stack.

Also, TOPs as a metric is usually meant for INT4, which the 4090 does support and manages 2600 TOPs with sparsity. Don't conflict that with TFLOPs.

As per this:

This is about the dense numbers, which the 4090 does at 1300 TOPs. 1300*4 = 5200, which is above the 4800 limit.
We can then assume that 4090D should have less than 1200 INT4 dense TOPs.

The 9070xt (and by extension the Pro 9700 since they're pretty much the same) achieves 1557 INT4 TOPs with sparsity, so ~780 TOPs dense, that's way less than the 4090D and way bellow the sanction limits, so I don't get why they had to reduce it further.

Could it be not about the ban restriction, but about that limit where companies need to notify such exports instead?

Couldn't be related to memory bandwidth either, given that this GPU has a pretty puny bandwidth and AMD did not change it for this new China-exclusive.

Yup. Compute-wise, it has around the same performance of the RTX 4080 non-Super, that card's performance rating being:

No way it's getting sanctioned on a performance basis, but... let me paraphrase my fan club:

Any chance you get

Any chance I get, indeed.

 

[igormp]

Yup. Compute-wise, it has around the same performance of the RTX 4080 non-Super. No way it's getting sanctioned on a performance basis, but...

In theory. In practice it's still behind a 3090 with the current software stack (to the surprise of no one), I verified this with a colleague on a simple mamf-finder run.

No way it's getting sanctioned on a performance basis

Yeah, I don't think this is related to the actual blocks, but rather them being able to avoid the necessity of any kind of notification whatsoever instead, as I had said before.
But that's just me guessing, I'm not really keeping up with whatever rules the US has been throwing around every other week.

Any chance I get, indeed.

Just stop being a fan boy, ez

 

[Dr. Dro]

2025 AI Diffusion Export Controls – Microsoft Regulatory Capture, Oracle Tears, Impacts Quantified, Model Restrictions

The US government lobbed the largest salvo in the new technology cold war with its new Framework for Artificial Intelligence Diffusion. These new export restrictions are completely unprecedented in…

semianalysis.com

Best article I've found, and it's still rather convoluted, but seems hardware at this performance class doesn't need to be issued a special authorization for export, while pretty much everything faster does.

On that note, I actually just installed LM Studio, getting 171 tokens/sec* on deepseek-r1-0528-qwen3-8b, not sure if that is good or bad, but it generates the text much faster than I can read it, and the output is surprisingly coherent. My RTX 5090 whirrs and purrs with a pretty unique coil whine while it's running, I found it adorable lol. Gonna try downloading some other models to play with

*lol literally just realized I had it at 66% power limit when I went to bench Wukong

 

[JustBenching]

If it's priced at around

9070xt price + ~50$ (the cost of 16gb of vram). No?