Nvidia is reportedly preparing a gigantic Hopper graphics chip with a size of 1,000 mm²
- The new large Hopper core should power high-performance HPC cards.
- Taiwanese TSMC should take care of the production using 5nm technology.
- It is possible that Nvidia is preparing an even more powerful multichip solution.
The graphics card market is about more than just players.
In the development of new solutions, manufacturers are constantly focusing on another, more financially appealing segment. The new Nvidia architecture, dubbed Hopper, is reportedly aimed at HPC (High-Performance Computing), or supercomputers. According to leaker kopite7kimi’s Twitter post, the area of the next most mighty GPU of its kind is 1,000 mm2.
The author of the status is a bit more conservative in the following post, but based on his statements, the size of the GPU with the alleged name GH100 can still be expected to be between 900 and 1,000 mm2.
In comparison, Ampere’s largest gaming GPU, GA102, which powers the GeForce RTX 3090, measures 628 mm2. Nvidia also manufactures it in Samsung’s factories using an 8nm process.
However, the new Hopper architecture will already rely on TSMC’s more advanced 5nm technology, which will enable many more transistors to be packed into a smaller area. A higher proportion of integrated units will result in much higher performance, in addition to an increase in area.
A huge piece of silicon
Importantly, the authors explained that the GH100 core will not be made up of multiple smaller chips, as previously speculated. On the contrary, it is intended to be a monolith or a single piece of silicon. This claim was later backed up by another well-known leaker who goes by the handle Greymon55.
This would imply that Nvidia chose a different path than the competing AMD. Its latest Instinct MI200 computing card is powered by a GPU that combines two interconnected smaller chips into a single package. The manufacturer will use the same approach for MI300 calculation cards, and its upcoming most powerful Radeon game graphics models based on the RDNA 3 architecture should already be multi-chip.
However, although the source states that the GH100 is not made up of multiple chips, it is possible that Nvidia is preparing such a solution, which could arrive a little later.
Hopper is confirmed
The presence of the Hopper architecture has been established to a large extent. According to Tom’s Hardware, there is a dispute between Nvidia and Dish Network, which already uses the brand on its DVRs and satellite receivers.
According to reports, Nvidia intends to unveil its Hopper kernel-based cards in March at the GTC 2022 event.
Let us also mention that Nvidia is working on an entirely new generation of GeForce graphics cards for this year, which will be based on the Lovelace architecture.