The big chip developer has been patiently waiting for its move to rely on 7-nm manufacturing processes - to go all-out with an unprecedented amount of processing units that will jeopardize our familiar levels of performance?
Today, the company NVIDIA Offers some of the largest processing chips seen in the commercial market ever in the top end products in its Turing and Volta architectures - up to 854 square millimeters for the formidable GV100 core that features Titan V and Quadro GV100 (as well as a number of branded accelerators -Tesla). How can you look for a core that will boost this unit's performance by 20 percent or even more? This is one of the questions that pervades hardware enthusiasts now, after some mysterious future-core performances in test pool repositories Performence In the network.
Models NVIDIA Unknowns with 108 processing clusters, 118 processing clusters, and even 124 processing clusters were observed in a well-known Geekbench test repository - also ignited by rumors and rumors of being larger (and probably more powerful) than anything we have known so far, with 80 processing clusters. Only "on the same GV100 core in which 5,120 CUDA units are active.
The product includes 118 processing clusters, with a relatively modest reported operating frequency of 1.11GHz and 24GB of memory Dedicated graphics, did not settle for the potential of paper alone, and even recorded a result that is about 20 percent higher than the GV100's core accelerators Tesla V100 in the OpenCL sub-test which is one of the components of the Geekbench array.
Now comes the big speculation stage: It's likely that what we're seeing here is the next generation of Tesla accelerators designed for the HPC, new 7-nm cores under the code name Ampere or Hopper and with 24GB to 48GB of built-in HBM2E fast memories The chip - all seemingly separate from the next-generation architecture that will be used by the GeForce family models for the gaming world, as we've seen in the current generation of NVIDIA With a relatively clear division between Turing products and products Volta (Which nevertheless also appeared in some products Titan and-Quadro Since they are capable of operating on standard home systems and can be used to run games).
While the 7 nm lithograph should allow for a high tens of percent transistor density, compared to the 12 nm lithography it does NVIDIA Use of its current products - a 55 percent jump from 80 processing clusters to 124 processing clusters and 7,936 active CUDA units is almost unprecedented, and may indicate the use of exotic physical-size chips of more than 1,000 square millimeters. It is possible that this increase in processing hardware implies a much larger and more fundamental change, as part of it NVIDIA Switch to using small Chiplets that are combined into one large chip chassis, much like it does with great success AMD As part of the architectureZen attitude Its 2 in general processors? While this is a far-reaching hypothesis that has no actual evidence at this time, the exciting technical data do not allow us to rule it out completely in the meantime.
We will continue to follow and hope to receive clues or answers soon - perhaps as part of the official GTC 2020 event NVIDIA Recently converted to an online transmitter due to World Corona virus hardships.