Thursday, August 3rd 2023
AMD Announces Radeon PRO W7600 and W7500 Graphics Cards
AMD today announced the Radeon PRO W7600 and W7500 graphics cards for the professional-visualization (pro-vis) market segment. These cards target the mid-range of the pro-vis segment, with segment price-band ranging between $350-950. The two are hence positioned below the W7800 and W7900 that the company launched in April. The W7600 and W7500 are based on the same RDNA3 graphics architecture as those two, and the client-segment RX 7000 series. AMD is pricing the the two new cards aggressively compared to NVIDIA. Both the W7500 and W7600 are based on the 6 nm "Navi 33" silicon.
The Radeon PRO W7600 leads today's launch, maxing out the silicon it is based on—you get 32 RDNA3 compute units, or 2,048 stream processors; 64 AI Accelerators, 32 Ray Accelerators; 128 TMUs, and 64 ROPs. The card comes with 8 GB of 18 Gbps GDDR6 memory across a 128-bit wide memory bus. The memory does not feature ECC. The card comes with a 130 W typical power draw, with a single 6-pin PCIe power connector. It uses a slick single-slot lateral-airflow cooling solution. AMD claims 20 TFLOPs peak FP32 performance.
The Radeon PRO W7500 is cut-down from the same "Navi 33" silicon as the W7600. It is equipped with 28 RDNA3 compute units that make up 1,792 stream processors, 56 AI accelerators, 28 Ray accelerators, 112 TMUs, and 64 ROPs. While the same 8 GB of GDDR6 memory is on offer, across the same 128-bit memory bus, it is down-clocked to 11 Gbps (the slowest for any GDDR6-based graphics card). AMD claims a peak performance of 12 TFLOPs, which is still higher than the 10 TFLOPs put out by the previous-generation W6600. While the older W6600 had a typical power draw of 130 W, the new W7500 is able to match its performance at just 70 W. It lacks any additional power connectors, and comes with a more compact single-slot cooler.
Both the W7500 and W7600 get the latest AMD Radiance Display Engine, with support for DisplayPort 2.1 with UHBR 10 (38.7 Gbps), which supports up to 10K @ 60 Hz with DSC. The GPUs also feature hardware-acceleration for AV1 encode and decode, besides HEVC.
The main play the Radeon PRO W7600 and W7500 have over their client-segment Radeon RX counterparts is the software stack, with industrial certification for dozens of 3D CAD and visualization applications; dozens more digital photography and video-editing applications, and further dozens of 3D content-creation and animation applications. Thanks to the AI accelerators hardware components, and support for ROCm 5.6, the GPUs also get hardware support for a large number of generative AI applications, and should accelerate features of existing content-creation suits that leverage generative AI.
The Radeon PRO family of GPUs also get a more prioritized support channel from AMD, quarterly driver updates under the AMD Software PRO Edition branch, and support for remote-deployable driver updates in an enterprise environment.
AMD is pricing the Radeon PRO W7600 at USD $599, at which price the company claims it outperforms the $646 NVIDIA A2000. The Radeon PRO W7500, on the other hand, is priced at $429, offering with a SPECviewperf score that's nearly double that of the similarly priced NVIDIA T1000.
The complete slide-deck follows.
The Radeon PRO W7600 leads today's launch, maxing out the silicon it is based on—you get 32 RDNA3 compute units, or 2,048 stream processors; 64 AI Accelerators, 32 Ray Accelerators; 128 TMUs, and 64 ROPs. The card comes with 8 GB of 18 Gbps GDDR6 memory across a 128-bit wide memory bus. The memory does not feature ECC. The card comes with a 130 W typical power draw, with a single 6-pin PCIe power connector. It uses a slick single-slot lateral-airflow cooling solution. AMD claims 20 TFLOPs peak FP32 performance.
18 Comments on AMD Announces Radeon PRO W7600 and W7500 Graphics Cards
Contrary to what they want you to believe, these 8GB/128bit are totally useless for any REAL professional uses like CAD/3D/animation ect....
Yea, I know REAL pro cards cost alot more and take more slots & power, but when "time is money", the increased productivity/output pays for them in a very short time :)
Been there, done that, 2 years ago....neva again !
For productivity, what is the value of homologous graphics cards
Was wandering about what core clock speeds these will have is how I found this.
Aside: I hate the name stream processors and CUDA cores. That is like saying each Zen 4 core has 16 AVX-512 stream processors.
/sigh...
Double the VRAM or go home, there's zero reason to buy these over the Radeon equivalents at the moment. Back in the day the single reason these things commanded a premium was the GPU acceleration for Remote Desktop Connection - entirely a driver restriction to paywall that feature for business users.
The thing is, RDP is obsolete in the zero-trust WFH era, and modern solutions don't care about what GPU you're using. So, the reason to buy a Pro card over a consumer Radeon used to be VRAM, and this brings no more VRAM to the table, meaning it's 100% useless in the market, though they'll probably sell some to PHB-types who just say yes to what a dishonest salesman tells them they need.
With Nvidia Studio drivers and AMD Pro drivers I think the biggest use for this cards is to allow big OEMs like Dell, HP and Lenovo to upcharge their lower end workstations.
It's a pity, the single slot cards are pretty sweet but it makes very little sense for a regular user to pay double for more or less the same cardISV Certification, they spend extra money making sure professional software from the likes of Siemens, Autodesk or Dassault etc. work without issues
It looks like calculation of Processing Power of recent AMD GPUs, including AMD Radeon Pro W7x00-series, is incorrect. A multiplier 4 is used instead of 2, for a number of FP-operations per clock.
I'm a C/C++ Software Engineer with a significant GPU programming experience. I always calculate a Peak Processing Power ( PPP ) for a GPU as follows:
PPP = Boost Clock ( in GHz ) * Number of Streaming Processors ( Shading Units ) * Number of FP-operations per Clock
For example, for AMD Radeon Pro W7600 GPU:
- I calculate : PPP = 2.440 * 2,048 * 2 = 9,994.24 GFLOPs = 9.99424 TFLOPs
- You calculate: PPP = 2.440 * 2,048 * 4 = 19,988.48 GFLOPs = 19.98848 TFLOPs
Why do you use 4 FP-operations per clock instead of 2?
I've bought hundreds of Quadros, FirePros, A-series cards etc and outside of the intentional feature-disabling on non-pro drivers, the bugs and fixes for professional CADCAM and simulation tools are, IME, always rolled out to pro and consumer GPUs at the same time. If they find a driver bug, they fix the bug. It's extra effort to fix it for the pro cards and leave it unfixed for the consumer cards, and driver teams always follow the path of least resistance (which often means just adding it to the "known issues" list and ignoring it until enough people rant about it online).
Honestly, you if the pro software version works on the RX card..... use the RX card.
the W cards generally have ECC ram, more conservative clocks for lower tdp/stablity focused.
and software support/testing is what you are paying for.