NVIDIA Pascal and AMD Polaris GPUs readying for Computex 2016
With more and more “leaks” and “rumors” spreading through the interwebs as we approach Computex 2016, we try our best to compile what we found out about the two upcoming GPU architectures. We are getting onto the hype train right now as we go through as many rumors as we can. We will go alphabetically so let’s start with AMD first.
We know about the two AMD parts, the AMD Polaris 10 and 11 and the fact that no Polaris part will pack HBM. More recent news from videocardz.com connects the marketable name of the GPUs and also offers some insight into the targeted TDPs for the two Polaris parts. AMD Radeon R9 470 would be based on Baffin GPU aka Polaris 11, with a TDP below 50W, while Radeon R9 480/490 is supposedly equipped with Polaris 10 aka Ellesmere with TDP slightly above 100W (110-135W).. A leak involving a certain Polaris 10 GPU with the code name “67DF:C4” is also found on VR-World. Some key specifications of the “67DF:C4” Polaris engineering sample are as follows:
Polaris 10 GPU “67DF:C4” Specifications (according to VR-World)
- 14nm FinFET, GlobalFoundries
- Diffused in USA (New York state)
- Assembled in Taiwan
- 2304 Cores (silicon: 2560)
- 36 Enabled Core Clusters (silicon: 40)
- New GPU Architecture (not GCN “1.5”)
- 256-bit Memory Controller
- 8GB GDDR5/GDDR5X (when available)
This GPU is supposedly not in its final form yet and may see both the core and memory clocked much higher than the engineering sample’s. The core is expected to be clocked somewhere in the range of 1150 MHz while the memory may see speeds in the range of 1750 to 2000 MHz when it enters production.
The takeaway from these rumors are an interestingly low TDP for both Polaris GPUs. The claims are pretty well backed-up by the promise of improved efficiency thanks to the leap from 28nm planar to 14nm FinFET manufacturing process. If these rumors are to be believed, both GPUs may see themselves in gaming notebooks soon. Afterall, we have seen the desktop NVIDIA GeForce GTX 980 with up to 150W TDP crammed into high-end gaming notebooks. True enthusiasts may decide to sit on their hands until the Vega 10 parts come out with HBM2 technology.
Pascal is a much anticipated architecture over in the green camp, as it also involves both a new architecture and process shrink. Things are looking pretty similar over here with the GP104 also reportedly not coming with HBM memory. However GP100 “big Pascal” has already made an appearance in the form of the Tesla P100 and it does pack HBM2 memory. The Tesla P100 is reportedly not using the full GP100 silicon and has a whopping TDP of 300W. TDP figures of the GP104 are yet to be leaked/released though but should be much lower than the Tesla P100’s TDP.
The super-sleuths at videocardz.com referred to an image of the GP104 chip and noticed the VRAM modules looked distorted, meaning that the GPU silicon’s size was also distorted. They then fixed it to obtain a die area of 310mm2, with an estimated transistor count of 8 million, based on the announced transistor count in the GP100 chip used in the Tesla P100. The Samsung memory used is the 8Gb 8Gbps chips, which should mean 8GB of VRAM and more bandwidth than the current 7Gbps chips used.
Three cards are rumored to be based on the GP104 silicon, namely the GTX 1070, GTX 1080 and GTX 1080 Ti. While the GTX 1080 Ti seems unlikely to be baed on the GP104, the GTX 1070 and GTX 1080 comes as no surprise as the GTX 970 and GTX 980 are both based on the GM204 silicon. While we await the actual announcement of these cards, enjoy the above image of the reference GPU shrouds.
A smaller Pascal part is also spotted in the form of the Drive PX 2 module. The guys at videocardz.com once again got to work with the images to bring us a comparison of the GP106 chip against the older GM206 silicon. The GP106 chip powers the Drive PX 2 module. They reasoned that it will measure somewhere between 170mm2 to 200mm2. Based on the rectangular shape they also came to a conclusion that it contains two Graphics Processing Clusters side by side. Based on tradition, the GP106 should be in the GTX 1060 graphics cards. Looking at the VRAM modules, it appears that it will be yet another GDDR5/GDDR5X part.
So what do you think of the upcoming GPUs? Both teams will have similarly powerful GPUs, but with AMD releasing their lower TDP parts and NVIDIA already firing their big guns, are you more attracted to the low TDPs and the promise of great perf/Watt of the AMD Polaris parts or the higher end NVIDIA parts now?