Nvidia showcases die shots of their Pascal P100 GPU
Nvidia showcases die shots of their Pascal P100 GPU
Published: 23rd August 2016 | Source: Anandtech |
Nvidia showcases die shots of their Pascal P100 GPU
Nvidia has showcased new die shots of their Pascal P100 GPU at Hot Chips 2016, showing us that Nvidia's GP100 GPU core has a lot of untapped potential.
Below we can see a lot of the GPUs underlying features and the GPUs SM unit, which are in groups of two and are the emerald green rectangular sections in the below die shot, showing 30TPCs and a total of 60 SMs. Here we can also see the GPUs memory bus and other GPU logic.
With 60 total SMs in their GP100 GPU core Nvidia has so far not released a product which uses their full new GP100 GPU core, meaning that Nvidia will likely be releasing a more powerful GP100 Tesla GPU in the future.
It is also expected that Nvidia's GP102 GPU core will have a total of 60 GPU cores, meaning that Nvidia's Pascal GTX Titan X is also expected to have a larger variant released in the future. These 4 additional SMs will add a total of 256 CUDA cores to the GPU, giving the GPU around 7% more GPU performance if run at the same clock speeds.
Nvidia may not be using the full GP100 pascal GPU core in the Tesla P100 for many reasons, but it is very likely that it is due to yield issues with the new GPU design. The Nvidia GP100 core is the largest that I have ever seen from Nvidia, with a die size that is 610mm squared, making it larger than even the Titan X's 601mm squared core design. This GPU is also the first Nvidia GPU that has been shown that was made using the TSMC 16nm FinFET processing node and the first to use HBM 2.0 memory, making this new GPU core design much more complex to manufacture.
| Full GP100 | Tesla P100 (NVLink) | Tesla P100 (PCIe 1) | Tesla P100 (PCIe 2) | GTX Titan X | GTX 1080 | GTX 1070 | GTX 1060 | |
| GPU Architecture | Pascal | Pascal | Pascal | Pascal | Pascal | Pascal | Pascal | Pascal |
| Process node | 16nm | 16nm | 16nm | 16nm | 16nm | 16nm | 16nm | 16nm |
| GPU core | GP100 | GP100 | GP100 | GP100 | GP102 | GP104 | GP104 | GP106 |
| SM Units | 60 | 56 | 56 | 56 | 56 | 40 | 30 | 20 |
| Cores per SM | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 |
| SP FP Performance | - | 10.6TFLOPs | 9.3TFLOPs | 9.3TFLOPs | 11 TFLOPs | 9 TFLOPs | 6.5 TFLOPs | 4.4TFLOPS |
| CUDA Core Count | 3840 | 3584 | 3584 | 3584 | 3584 | 2560 | 1920 | 1280 |
| VRAM Type | HBM2 | HBM2 | HBM2 | HBM2 | GDDR5X | GDDR5X | GDDR5 | GDDR5 |
| VRAM Cappacity | 16GB | 16GB | 16GB | 12GB | 12GB | 8GB | 8GB | 6GB |
| Memory Bus Size | 4096-bit | 4096-bit | 4096-bit | 3072-bit | 384-bit | 256-bit | 256-bit | 192-bit |
| Memory Bandwidth | 720GB/s | 720GB/s | 720GB/s | 540GB/s | 480 GB/s | 320 GB/s | 256 GB/s | 192 GB/s |
| Base clock speed | - | 1328MHz | - | - | 1417MHz | 1607MHz | 1506Mhz | 1506MHz |
| Boost clock speed | - | 1480MHz | - | - | 1531MHz | 1733MHz | 1683MHz | 1708MHz |
| TDP | - | - | - | - | 250W | 180W | 150W | 120W |
| Power Connection | - | - | - | - | 1x 8-pin 1x 6-pin | 1x 8-pin | 1x8-pin | 6-pin |
| NVLink | NVLink | PCIe 3.0 | PCIe 3.0 | PCIe 3.0 | PCIe 3.0 | PCIe 3.0 | PCIe 3.0 |
Right now Nvidia states that the P100 is in volume production and that the chip was first sold to the supercomputing market in June 2016 and will be arriving with OEMs in Q1 2017.
You can join the discussion on Nvidia's full Pascal P100 GPU on the OC3D Forums.
Most Recent Comments
|
Originally Posted by SuB
The signal strength indicators are a nice touch.
|
Quote
Quote
CSS Valid
Nvidia could bring back the 90 series of GPU or it could be the 1080Ti.Quote