Before starting and as usual, let's take this data with tweezers and also add a little salt to it, since it comes in a different shape. First and foremost to create our own, let's take a little pre-filtered:
- 7 different chips: GA100 and GA101 manufactured by TSMC internally 7 nm EUV, GA102, GA103, GA104, GA106 and GA107 made by Samsung in 8nm EUV.
- GDDR6 a 18 Gbps, GDDR6X is possible on high performance chips, HBM2E for HPC.
- Dimensions 5376 Shaders in GA102
With this clear premise, let's see if the user is KH Chia mature, author of a reward like 64 EPYC and related drawings.
The NVIDIA RTX GA100 will be a chip with more than 30,000 million transistors
The table shown above is all the leaks that have been revealed and have been abruptly deleted from your Twitter account. This table does not leave a headache, starting with the fact that the GA102 does not present itself as such and will not give the card the largest gaming card under Ampere, something that can be seen from generation to generation.
Curious details that mark the future of all distance. The GA100 will be included in the HPC high-performance HPC concept with HBM2E and will 8192 Shaders, which is 7 nm will be more or less the figure 30 billion transistors.
By comparing the tort, the TU102 carries 18.6 billion, that is, the GA100 will receive a + 61.29% many transistors are grateful for the reduction of the lithographic process. Although we know that the ratio of the final power to the transistors is not 1: 1, it is a suggestion we can expect.
So, I would find out 128 SM with 256 RT Cores, 1024 Tensor Cores and a minimum of 1750 MHz clock
High serial speed, but without GDDR6 at 18 Gbps
Connecting to the lithographic process can mean an increase in 200 MHz vs. Exercise, so we would be approaching series prices over 2 GHz on the Founders models, but instead, we weren't getting Samsung's latest pieces, much less the next generation of GDDR6X memories.
In return, they would arrive faster 16 Gbps, something we've already seen in several other specialty models and that could make the NVIDIA RTX less AMD rival with bandwidth compared to HBM2E. You may not need it, or you can use a larger bus that allows for greater efficiency.
For example, if NVIDIA uses a 512-bit bus in the GA100, we'll be talking about it 1,024 GB / s, a figure that puts AMD cards at the top of the list. As for dates, the first to arrive will be the RTX 3080 and RTX 3070, likely in Q3 this year, when the RTX 3080 Ti will arrive in Q4 and RTX 3060 in Q1 in early 2021, a similar move happening with Turing.
Logically, the problem we have right now is simple: who's right? Everything that has been seen so far is happening at a technical level, some attributes are diminished by saying what the case is, some are completely inconsistent when we look back, but still how can we know whether, or not, AMD surprises with RDNA 2, or NVIDIA will return to lead and surprisingly, we prefer to reward something or others.
