NVIDIA has yet to launch their entire Pascal graphics card lineup and rumors have already started hinting at their upcoming GPU architecture. Announced more than years ago, the Volta GPU was to be NVIDIA's next greatest chip architecture to replace Maxwell but was pushed back with Pascal taking its place. With Pascal development complete, NVIDIA has set their efforts in the developing their latest GPU project.
NVIDIA Volta Architecture Rumored To Be Featured in High-Performance GV102, GV104 and GV110 GPUs
The rumor on NVIDIA Volta GPU comes straight from the Motley Fool who were able to obtain some big details on the next-gen chip architecture through a Chinese source. According to the rumor, NVIDIA is prepping a range of chips that will utilize the Volta architecture. Not only that, the Volta GPU architecture has been significantly revamped and will offer a serious improvement in terms of performance and power efficiency compared to the Pascal GPUs.
The site mentions development of three GPUs based on the Volta architecture is underway. Last night, NVIDIA confirmed during their earnings call that they have taped out and finished development for all Pascal GPUs but they will be introducing all of them at various time frames. This statement confirms that NVIDIA's development department is now focused on finishing Volta in time. However, I would like to point out that development on these GPUs had begun much earlier as it took NVIDIA three years to to develop the Pascal architecture with a total cost between 2-3 billion dollars.
Following are the chips that are rumored to be introduce as high-end gaming GPUs:
- NVIDIA GV110 GPU
- NVIDIA GV102 GPU
- NVIDIA GV104 GPU
Volta has long been part of NVIDIA's roadmap and there's no doubt that most of the work on these chips will be nearing competition as we enter 2017. During these phase, engineering secrets leak out and we are looking at a similar situation where someone was hinted about Volta's architecture enhancements. According to the rumor, Volta is significantly redesigned compared to the Pascal architecture.
NVIDIA's Pascal architecture has significant gains in almost all departments over Maxwell but it's also fact that Pascal and Maxwell share a common DNA. The SM design for both chips is very same. Aside from the GP100 GPU which is different due to its FP64 hardware blocks, the GP106, GP104 and GP102 share a lot in common with Maxwell designs.
NVIDIA Maxwell (GM204) and Pascal (GP104) SM Designs:
The streaming multiprocessor block for Pascal and Maxwell was similar but was vastly tuned along with the upgrade to the new FinFET node which led to some good increases. Volta on the other hand will feature a completely new design across the board.
NVIDIA GV102, GV104 For High-End Gamers, GV110 Also Aimed at Consumer Market?
Starting with the GV104 GPU, we know from the name that this chip will replace the GP104 GPU and will be the most ideal chip for gamers with a decent price tag. NVIDIA's most successful cards came from the G*104 series of GPUs since the Fermi series. Moving on, we can see vastly improved performance from such chips. The next chip in the stack is GV102. Succeeding the GP102 which is housed inside the GeForce GTX Titan X (P) card, the new high-performance chip will be aimed at enthusiasts who demand serious horsepower.
The third chip is the most interesting as it is also mentioned as a gaming chip. Known as GV110, the GPU will be the ultimate design featuring the Volta architecture. Currently, the GP100 GPU is the top dog of Pascal line and is housed inside the HPC accelerators such as Tesla P100. This chip is built for double precision workloads and designed around NVLINK interconnect.
The GP100 GPU is also the only chip that utilizes the fastest HBM2 memory. Now GV110 is interesting as it recalls the naming scheme of GK110. GK110 was also a chip which featured the double precision capabilities, however they were cut down on the consumer variants.
With GP102 and GP100, NVIDIA has tried to maintain parity in terms of performance while getting rid of the non-essential features that are not required in gaming and professional workloads (FP64/NVLINK). GV110 can be the full fat chip with limited FP64 capabilities designed for the prosumer market.
It will be interesting to see how this works out but we shouldn't expect Volta GPUs this soon. The closest we will hear anything official from NVIDIA on Volta would be their GTC presentation in 2017 but aside from that, much of the talks will be focused towards new Pascal GPU launches as they have yet to introduce the full fat variants of GP102 and GP100 in the market.
|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)|