AMD’s Greenland GPU May Feature Up To 32 GB of HBM2 Memory – Around 15-18 Billion Transistors, Two New GPUs Leaked
It has been four years since the GPU industry saw its first 28nm fabricated product in the form of Tahiti from AMD. Since Tahiti, both AMD and NVIDIA have been stuck on the 28nm node offering little architectural and design changes. While NVIDIA managed to release two new families (Kepler and Maxwell) on the 28nm node, AMD had to stick with their GCN architecture and offered just three new chips after their 7000 series cards in the mainstream and performance oriented lineup that include Tonga, Hawaii and Fiji. With the arrival of the latest FinFET node imminent, AMD is ready to introduce their latest and most beefiest design to date in the form of Greenland GPU in 2016, whose alleged details have made their way on the net.
AMD Greenland GPU May Feature Up To 32 GB GB HBM2 Memory and 18 Billion Transistors
Straight from TweakTown, the latest reports mention the Greenland GPU which has been in the talks for a while now. The latest Arctic Islands family of graphics cards will include the Greenland GPU and fall inside AMD's Radeon 400 series graphics lineup (or AMD might just give it a special name like they did with Fiji). Reports of the Greenland GPU first arrived several months back (prior to the launch of Fury X) where it was revealed that AMD might be preparing a new chip codenamed Greenland but never really revealed the true nature of it. The first report also mentioned that the Greenland GPU will be integrated on APUs with HBM2 memory but now it makes sense as to what the chip was made for. Just a few weeks back, we covered in an article that Greenland would not only become the flagship chip of the stack, but also the most scalable chip AMD will ever produce.
Image Credits: Fudzilla
AMD's Fiji GPU has shown what is achievable with HBM as it can be incorporated on designs ranging from Fury X, Fury X2, Radeon R9 Nano and even two Fiji chips inside the Quantum PC (which unfortunately will never be made available to consumers). With Greenland, FinFET and HBM2, AMD is aiming towards both the discrete and APU market. An HPC APU design has already made its round and AMD has officially confirmed that they plan to launch a server class APU featuring TFlops of performance. This HBM powered HPC APU will pack the Greenland GPU, several Zen cores, a 100 GB/s interconnect and several GBs of DDR4 memory. This HPC APU is planned for launch in 2017 timeframe and pack 4.0+ TFlops of performance while AMD's current FirePro S9170 accelerator has 2.62 TFlops of performance. While Fiji isn't geared for HPC purposes, a 4.0 TFlops class APU is indeed a marvel of engineering.
So coming back to the discrete part which is also entirely related to the APU part, the rumors not mention that the Greenland GPU will feature around 15-18 billion transistors which falls in line with NVIDIA's Pascal GPU that is going to end up with more than 17 Billion transistors and is based on the FinFET process to, featuring up to 32 GB of HBM2 memory. Now both NVIDIA and AMD have waited a long time to make transition over to a smaller process node. Their current chips are the largest GPUs ever produced in terms of sheer size and overall transistor count (GM200 and Fiji XT). On their Financial Analyst Day, AMD mentioned that their next GCN based cards will feature 2x the energy efficiency over the current GCN iteration. Since FinFET offers up to 90% more transistor density, the 17-18 billion transistors on NVIDIA's and AMD's chip sound completely possible. AMD is expected to offer a completely new GCN architecture with Greenland with a new Instruction Set Architecture or ISA. It has been mentioned that AMD has spent the last two years developing the Greenland GPU and they just recently taped it out.
The performance of next generation graphics cards is indeed going to shake things up. Currently, we see the integrated graphics processors from Intel and AMD closing up on entry level discrete cards but when FinFET comes, we will once again see entry level cards featuring twice or more performance. AMD is planning to incorporate just as much HBM2 memory as NVIDIA's professional and high-end solutions with up to 32 GB of VRAM. The consumer oriented parts will stick with 8-16 GB but the more high-end professional cards will get insane amounts of HBM2 with speeds of 1 TB/s. Technically speaking, both AMD and NVIDIA GPUs sound same in specifications but the ultimate decider would be the price, performance and total efficiency of these cards.
AMD Arctic Islands (Radeon 400 Series) Baffin and Ellesmere Leaked
Two more chips from the Arctic Islands have been tipped by Kitguru which include Baffin and Ellesmere. Now its kind of funny that although these two Islands are found in the north of Canada, they don't essentially fall in the Arctic Islands but rather in the Canadian Islands. Only Greenland is so far the only confirmed GPU that falls in the Arctic Islands so we aren't exactly sure whether these codenames can be confirmed right now but AMD does choose names from different Islands, close to each other. For instance, we get to see both Caribbean Islands in the Radeon 300 (Antigua, Trinidad, Tobago, Grenada) and then we see Fiji which isn't a part of the same Islands. These two chips are mainstream and entry level GPUs featuring the latest GCN core design. No more details are yet known but we will sure keep you updated as new information arrives.
|GPU Family||AMD Vega||AMD Navi||NVIDIA Pascal||NVIDIA Volta|
|Flagship GPU||Vega 10||Navi 10||NVIDIA GP100||NVIDIA GV100|
|GPU Process||14nm FinFET||7nm FinFET||TSMC 16nm FinFET||TSMC 12nm FinFET|
|GPU Transistors||15-18 Billion||TBC||15.3 Billion||21.1 Billion|
|GPU Cores (Max)||4096 SPs||TBC||3840 CUDA Cores||5376 CUDA Cores|
|Peak FP32 Compute||13.0 TFLOPs||TBC||12.0 TFLOPs||>15.0 TFLOPs (Full Die)|
|Peak FP16 Compute||25.0 TFLOPs||TBC||24.0 TFLOPs||120 Tensor TFLOPs|
|VRAM||16 GB HBM2||TBC||16 GB HBM2||16 GB HBM2|
|Memory (Consumer Cards)||HBM2||HBM3||GDDR5X||GDDR6|
|Memory (Dual-Chip Professional/ HPC)||HBM2||HBM3||HBM2||HBM2|
|HBM2 Bandwidth||484 GB/s (Frontier Edition)||>1 TB/s?||732 GB/s (Peak)||900 GB/s|
|Graphics Architecture||Next Compute Unit (Vega)||Next Compute Unit (Navi)||5th Gen Pascal CUDA||6th Gen Volta CUDA|
|Successor of (GPU)||Radeon RX 500 Series||Radeon RX 600 Series||GM200 (Maxwell)||GP100 (Pascal)|