Earlier this year, graphics specialist NVIDIA (NASDAQ:NVDA) began rolling out graphics processors using its Pascal graphics architecture and manufactured in TSMC's (NYSE:TSM) 16-nanometer technology. The Pascal-based chips managed to deliver significant performance increases over their predecessors based on the company's Maxwell architecture and manufactured in an older 28-nanometer technology.
NVIDIA's next architecture after Pascal is known as Volta, which the company has previously said would bring substantial performance-per-watt gains over Pascal. According to a post on Baidu from user USG Ishimura (I am told that this user is a well-known leaker), NVIDIA will be doing some very interesting things with Volta.
High performance, significantly revamped architecture
According to a translation of USG Ishimura's post, the performance of GV104 -- the Volta-based successor to the recently released GP104 chip (which powers the GeForce GTX 1070 and GeForce GTX 1080 graphics cards) -- will offer "really strong" performance.
This performance enhancement, the user indicates, is due to the fact that the "sm structure [has] changed" relative to the prior-generation Pascal architecture (in NVIDIA's graphics architecture, "SM" is short for "streaming multiprocessor."
It's worth noting that with the Maxwell architecture, NVIDIA made substantial changes to the streaming multiprocessor architecture. Here's what NVIDIA said about Maxwell relative to the prior-generation Kepler architecture in its whitepaper describing Maxwell:
The primary contributor to Maxwell's improved efficiency is the new Maxwell SM architecture, SMM. This new SM architecture achieves much higher power efficiency and delivers 35% more performance per CUDA Core on shader-limited workloads.
In going from Maxwell to Pascal, however, NVIDIA appeared to have kept the streaming multiprocessor structure nearly identical generation on generation. Here's what the Maxwell SM looks like:
And here's the Pascal SM:
NVIDIA tells me that the company did make "general SM architectural optimizations," so they aren't identical. However, it's clear that the company didn't make the kinds of changes in going from Maxwell to Pascal at the SM level as it did in going from Kepler to Maxwell.
In light of this, I think that it makes a lot of sense to expect a dramatically overhauled SM architecture with Volta, driving significant performance-per-watt gains.
More high-end gaming chips this time around?
During the Kepler generation, NVIDIA's highest-end gaming-focused graphics processor was a chip known as GK104. NVIDIA also brought out a chip known as GK110 to serve as higher-end gaming processors (Titan, Titan Black, 780, and 780 Ti), but it was a product aimed mainly at high-performance computing that was forced to do duty as a higher-end gaming product.
Beginning with the Maxwell generation, NVIDIA sold a chip known as GM204 as the GeForce GTX 980, and then rolled out a part known as GM200 -- essentially a GM204 with 50% more everything -- as the GeForce GTX Titan X and a partially disabled variant known as the GeForce GTX 980 Ti.
NVIDIA continued with this practice with Pascal, selling GP104 as the GeForce GTX 1080, and GP102 as the new Titan X.
According to USG Ishimura, NVIDIA is actually planning not two, but three high-end gaming graphics processors based on Volta this time around: GV104, GV110, and GV102.
It's not clear how GV110 and GV102 will be differentiated, but I would imagine that GV110 will be an even higher-performance product than GV102, which could potentially allow NVIDIA to introduce a performance and price tier above the $1,200 level that the GP102-based Titan X sells for.