The era of FinFET based GPUs is finally drawing closer and as far as Nvidia is concerned, the relevant architecture is Pascal (and later, Volta). Today, we will be looking at some basics of the chip lineup of Nvidia’s upcoming 16nm FinFET+ graphic cards. Before we begin the editorial, I would like to point out that while most of this is technically educated guesswork – it is based on the historical nomenclature used by Nvidia – and is as such more or less factually accurate. That said, I have marked the dubious areas with tags to help speculation-averse readers in digesting this peace.

Mulling over Nvidia's chip lineup for the next generation: GP100 through GP108

As every graphic card enthusiast knows, there are two kinds of nomenclature: internal-chip-nomenclature and the commercial nomenclature. While the commercial nomenclature of the upcoming lineup can be anything (for eg Geforce GTX 10xx) the chip nomenclature remains the same as ever – and therefore predictable. Although most of our readers would be well aware of how Nvidia’s internal nomenclature works, here is a breakdown for those not in the fold:

Nomenclature example: GM204

Gx xxx: The first letter is a constant which stands for Graphics (Technology)
xM xxx: The second letter is a variable which stands for the architecture of the chip. For Kepler it was K, for Maxwell it was M, for Pascal it was P and for Volta it will be V and so on.
xx 2xx: This numerical stands for the generation of the architecture. First generation Maxwell for eg,  had a 1 over here.
xx x0x: This numerical has a variable meaning and can stand for everything from a better binned chip to a  refresh. This digit is usually ignored when interpreting chips.
xx xx4: The last digit is the actual performance indicator of the chip and follows an inverse approach. Lower means higher performance so a 0 here would mean the flagship die and an 8 would mean the weakest die.

Now that we have set aside, lets talk about the actual lineup. Maxwell currently has a total lineup of 6 dies. These are: GM 200, GM204, GM206, GM107 and GM 108. Traditionally, the flagship die lands first. However, Nvidia has been mixing up this formula from time to time. In Maxwell’s case for eg, the GM107 and GM108 came first. One can argue that they were, technically, the flagship chips of their own generation (1^st Generation Maxwell) but then Nvidia also introduced the GM204 before the GM200.

The lineup speculation was originally done by 3DCenter – I will just be building upon the same. Now, since the 1^st generation and 2^nd generation are both active the moment, [caution: speculation] we can safely assume that they will be replaced by 16FF+ parts sometime in the future. By 2017, we can assume that the entire lineup should be present and that Volta should be visible on the horizon as well. The first lineup that is slated to appear is based entirely on Pascal and the chip nomenclature should be as follows:

• GP100: First generation Pascal flagship which will replace the GM200 die.

• GP104: First generation high end Pascal GPU which will replace the GM204.

• GP106: First generation mid end Pascal GPU which will replace the GM206

• GP107/108: First generation low end Pascal GPU which will replace the GM107/GM108 chips respectively.

Previous reports have stated that the GP100 ‘big’ pascal chip will hit the professional market first, giving time for green to roll out the GP104 chip to the mainstream consumer segment (followed by GP100 at a later date).  The GM200 taped out in June and hit the shelves in March. That is a total time of approximately 9 months. The GM204 on the other hand took only five months to hit the shelves from its tape out date. At this schedule we can tentatively expect the Pascal architecture GPUs next year in late second quarter. Pascal flagships will also have HBM2 memory, which unlike the HBM1 specification can extend upto 16 GB of total vRAM size.

Nvidia’s Volta series of GPUs will eventually replace the Pascal however, at this point there is no clear indication of what process the chip will be at. If the GP100 die is over or around the 550mm^2 limit then it is very likely that the GV100 chip will be on the 10nm FinFET process based on the 14nm backbone (the 16nmFF+ process is based on the 22nm backbone). However, if the GP100 is around the 500mm^2 mark, then there will be atleast one more generation on the 16nm FinFET+ process since the usual limit at TSMC is 600mm^2 – be that a Pascal GPU (GP200) or a Volta GPU (GV100).

Keep in mind that reports have alleged that the GP100 will have approximately 17 billion transistors. Take a look at the official statement about the 16FF+ node from TSMC:

TSMC’s 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology. Comparing with 20SoC technology, 16FF+ provides extra 40% higher speed and 60% power saving.

As you can see, there is an obvious conflict present in our available information. Two times 8 Billion (the transistor count of the GM200) is still around a Billion shy of the alleged transistor count - and this is at the huge die size of 601 mm^2, something which is highly improbable if not impossible for the first batch. Its possible that generational differences between the architectures make it possible to fit 17 Billion transistors on a ~520-550mm^2 die, but it is more probable that the transistor count is an exaggeration.

Assuming [caution: speculation] a die near the 550mm² mark initially, you are looking at around 5000-6000 CUDA cores. With the architecture improvement and and the process upgrade – an improvement of around 60% – 80% is on the cards depending on how well Nvidia handles it. There isn’t a single GPU card in existence that can handle 4K@60fps on its own. First or second generation Pascal however, should most definitely be able to hit the mark with the technological upgrades heading its way.