AMD this week issued a teaser post in Twitter that pre-announces formal launch of the company's Radeon Pro graphics card based on the RDNA2 architecture on June 8. While the company did not officially disclose specifications of the Radeon Pro W6800, an unknown leaker disclosed a Geekbench 5 entry early on Friday June 4. As this is a leak the information should be taken with a grain of salt.
AMD's upcoming workstation-grade Radeon Pro W6800 reportedly carries AMD's Navi 21 graphics processor in its full configuration with 5120 stream processors as well as 32GB of GDDR6 memory connected to the GPU using a 256-bit bus, according to previous leaks and a Geekbench 5 database entry. So far, AMD has not offered an RDNA2-powered graphics card with 32GB of memory onboard, so this workstation-grade product will be the first one in the family to feature 32GB of GDDR6 SDRAM.
The maximum GPU frequency was reportedly set to 2,554 MHz, though whether this was final or not is unknown. AMD's Radeon RX 6900 XT graphics board has an official Game Clock frequency of 2,250 MHz, but its actual maximum clockspeed is much higher. Typically ProViz cards are clocked lower than their gaming counterparts, but RDNA2 cards have proven very capable of high clocks.
AMD itself confirmed that the Radeon Pro W6800 graphics card supports up to six 4Kp60 monitors (or three 5Kp60 or one 8Kp60 monitors) using six mini-DisplayPort 1.4 connectors as well as AMD's Eyefinity technology, which will be another unique feature of the product.
The board uses a blower-type cooler, which is common for workstation and datacenter cards as it ensures longevity. Speaking of datacenters, it should be noted that AMD's Big Navi GPU supports GPU virtualization (including SR-IOV), so the upcoming board could be used in datacenters to power multiple remote workstations.
As far as OpenCL performance of the Radeon Pro W6800 is concerned, the card scores 137,230 points in Geekbench 5, which is in line with that of the Radeon RX 6800. Geekbench 5 isn't a very good benchmark since its results vary greatly, because the tests are too short for a GPU to set itself to a stable frequency.
AMD will officially unveil its Radeon Pro W6800 32GB GDDR6 graphics card on June 8, so before that date any information about the product should be taken with a grain of salt.
And it's about time ... AMDs portfolio is completely outdated. Currently the most "up to date model" is the Radeon Pro VII, a chip from the end of 2018.
And the validation topic is the same for Nvidia, that's no excuse. AMD is simply not putting enougth efford into their Pro series, therefore they have to compete mainly via a cheaper price.
And current gen Ampere-chips will for example be still be faster in RadeonPro Renderer, because they have more horse power (this time the AMD-chip is simply the consumer-chip).
https://www.tomshardware.com/news/AMD-announced-Radeon-Pr... and offers up 6.5GFLOPS for $2.5K (newegg), while the A100 was introduced sometime in the fall of 2020 offering up 9.7GFLOPS of FP64 for about $10K (nextwarehouse).
So (2.5/6.5)/(10/9.7) = 38% (meaning one gets a GFLOP of FP64 compute on the AMD card for 38% of a similar amount of FP64 on the A100). So, if all you care about is FP64, like myself, then the AMD card is currently a no-brainer.
Oh and this W6800 will very likely have crippled FP64 (1/32 or 1/16 vs 1/2). So we wait for a so-called real FP64 update card from AMD sometime this fall which might be the first to break the 10GFLOPS FP64 barrier AFAIK.
And yes, it's cheap, simply because it has to be cheap because otherwise nobody would by it and instead use the market leader's hardware.
And FP64 performance is only a small part of the portion and for example in large scale deployments I have never heard from anyone using Radeon Pro's (maybe in some datacenters). For example there only exists a singular system in the Top500 list, witch uses a (unspecified) Vega20 card, the rest of the list uses nVidia (or is CPU-only).
Additionally your comparison is incomplete, because the (G)A100's FP64 is in the range of about 10 - 20 TFlops. The lower value belongs to the base shader (ALU) performance, the upper value is the preak performance via using the new FP64 Tensor Core v3 functionality with MMA ops, therefore in real world apps, the performace will be somewhere between these two values, depending on the concrete workload.
Additionally the A100 outperforms the Vega20 card in AI workloads by far.
Therefore over all even the latter CDNA hasen't managed to really outperform the A100.
And yes, the Navi21 is crippled according to FP64, because AMD has limited FP64 performance as nVidia does to save transistor space for other purposes (here 1/16 FP32), whereas nVidia even limits it to 1/64 FP32 performance (it is only available for compatibility purposes).
Also, there will be no real FP64 "update" from AMD. AMD has split up its development lines and compute workloads are covered by CDNA (later on this year CDNA2), whereas RDNA will be limited to graphics. And for example a CDNA-chip will never be used in a Radeon Pro because the chip isn't a complete/full-blown GPU any more because AMD has removed certain elements from the render pipeline in the architecture.