For more than a year we have been commenting on the strange movements of NVIDIA with TSMC and Samsung, and finally after AMD‘s coup with the Taiwanese the Koreans have been chosen to give life to the chips of the latest green architecture. What are Samsung’s 8 nm like and what advantages and disadvantages do they have? Is it the most optimal for such novel GPUs?
Perhaps the most curious thing seen in Samsung’s leap to accelerate EUV technology is precisely that its latest and most advanced node available to its customers does not obtain this revolutionary technology. In fact, it will be the company’s 7 nm that will debut an ultraviolet-based recording process, so these 8 nm are still in DUV.
A complexity of patterns that allows more flexible designs
The most obvious question would be the most pertinent, how advanced is this 8nm? We will start from the basis that Samsung uses the same patterns as in its 10 nm node, so we are talking about a natural evolution rather than a revolutionary node.
In fact, the end pitch has been reduced to 42 nm, the Gate Pitch to 64 nm and the Mx to 44 nm , therefore, there is only improvement in Gate Pitch and Mx and for the minimum (4 nm in both cases) . Due to these metrics, Samsung is able to insert fewer fins in less space, supposedly achieving up to 15% of the logical area versus 10nm.
With this, the density for this new lithographic process has been set at 61.2 MTr / mm2 , being the cells with the highest density of the company before making the definitive leap to EUV. You may be wondering if this is a lot or a little compared to the 7 nm of TSMC, well, let’s say it is quite less dense than expected, since its competitor achieves 91.20 million transistors per square millimeter, or what is the same, it is 49% less dense and therefore NVIDIA will not be able to unleash its full potential as it can with Ampere GPUs for HPC.
In other words, Ampere for gaming has a much less advanced lithographic process than AMD’s, it is considerably less dense and less energy efficient, so AMD with RDNA2 has an unbeatable opportunity to beat NVIDIA.
So what’s good about Samsung’s 8nm lithographic process?
The first thing to mention here is that, although it is less dense than its rival, Samsung has achieved several milestones to be FinFET transistors. The first is what they call hybrid contact, a method that is based on the use of silicon together with germanium in the purest Intel style. With this and after the optimizations in nFET, it has been possible to reduce the voltage resistance in the gates by approximately 15% .
What does this mean? Well, reducing the resistance to voltage logically influences less energy losses and thus better efficiency and more exhaustive voltage control, as well as lower tolerances. Therefore, these 8 nm manage to reduce the energy consumed by 7-10%.
The company, however, has not released data on the improvement of performance, either in frequencies or in consumption at the same frequency, mainly because this lithographic process has the peculiarity that it was only going to be used in its LOW POWER version, until NVIDIA intervened and the high-performance variant was designed for Ampere Gaming.
It is not surprising that NVIDIA does not want to lose contact with TSMC, since if it were not for its 7 nm the A100 would not have seen the light with the performance they offer, and which at the same time means that Huang’s have a lot room for improvement to find a TOP node in the future and thus leave behind the Samsung 8 nm .
