NVIDIA Ampere either Hooper? It is not as clear as it seems. Although we always refer to a series called RTX 3000 as Ampere, Hooper's rumors do not end in its efforts and, as we noted with Turing, NVIDIA's steps to rename its premises with RTX are repeated.
In any case, Huang's new construction has a launch date, and although it is not official yet, it is ripe. Although this may be interesting, it may not be so if the circulation of new rumors is correct.
NVIDIA will launch its Ampere launch in the GTC on March 23
The last thing we can hear is that this new build method can be a step past the desktop, introduced as a graduate Volta and not from Turing, which could leave us with professional-level graphics cards. The move that NVIDIA can try to privatize is the introduction of BIG Navi by AMD.
Or March 23 on an important day, it may be inside the event GTC NVIDIA is moving the date to the 26th of the same month, which will be subject to AMD travel.
In any case, sources and rumors say that until September we will not see the cards in stores, so it would be more of a preview than the official presentation itself.
Original details and construction details
As we expected two weeks ago, NVIDIA could offer half of its invisible chips so far, when it looks like it will try to cover the market better to avoid a move like this that has made it to the SUPER range to come out. AMD Navi cards.
So, at least we will have 5 chips in the playlist already included, which could result in 8 or 10 RTX 3000 cards altogether, when it is unclear which one will reach the market before.
As we can see, the amount of SM has almost doubled in many cases, where the GA100 will rule less 128 SM and 8192 Shaders. This leaves us a part 64 Shaders per SM, so NVIDIA would not change its figure again, but what improvements will each Ampere smoker offer compared to Discovery?
As we know, SMs are a low-level unit used by NVIDIA and where all the "magic" happens.
New SM with some important issues
If the filter in the image above is true, we are talking about improvements that are still smaller than expected, but at the same time very useful. The number unit will remain unchanged, keeping the number of units and their value, in the same way, Warp scheduler, Dispatch and log files will remain unchanged and in the same clock per cycle.
In turn FP32 units doubled their number maintain a good balance with shaders, which means a significant increase in performance. Similarly, the performance of Tensor Cores also doubles, something that will improve your performance with DLSS.
L1 cache extends from 96 KB to 128 KB, where it is not specified which part corresponds to the Data and which memory is shared. In any case, it is said to have improved in terms of complex operations.
Finally, RT Cores of Track Ray they will be much better, more efficient and faster when working with BHV algorithms, but it is unclear if these will be upgraded and if NVIDIA has been able to take any further steps to accept the fact that it ceases to be a hybrid and shack process.
In addition, Huang would confirm internally that Samsung was not ready to support the volume of the larger plates as delivered by NVIDIA in the RTX 3000 series, so it could rearrange orders TSMC and Samsung itself simultaneously to ensure continuous chip movement. We're less than two months away from knowing more details, but if the writing is completely true, NVIDIA's performance will be more than noteworthy.
