About a generation Lovelace wedged between Ampere and originally featured Hopper, we learned last year before Christmas. In February this year, there were reports of a probable configuration of the chip that will be used for the top model. The AD102 core is to be equipped with up to 18432 stream processors, which is 71% / 1.71 × more than the current GA102 from the generation Ampere with its 10752 stream-processors.
Intergenerational power shift
In the February article, however, we discussed that the value of “1.71 ×” will not illustrate the difference in performance. While the current top GeForce RTX 3090 model does not carry a partially truncated GA102 core, for the GeForce RTX 4090 it may (but need not) use a less truncated to fully active core, resulting in up to 1.76 times the active GeForce RTX 3090 stream processors. this can lead to a shift in the clock frequencies (with the 5nm TSMC process compared to the 8nm Samsung process, which is a derivative of 10nm production) in the range of 1.1-1.15 × and further a shift of the “IPC” architecture around 1.1-1.15 × we get at least twice the total power (1.71 × 1.1 × 1.1).
Our February estimate is confirmed by the Greymon55 leak report, according to which the GeForce RTX 4000 generation is to double the performance of the current generation. The RedGamingTech website then has the information that the intergenerational shift can be up to 2.2 × (which is still within the range of options described in the previous paragraph), so it can be assumed that the performance of GeForce RTX 4090 will be 2.0-2.2 × higher than the performance of GeForce RTX 3090.
Manufacturing process
In connection with Lovelace, there was originally talk of a 5nm TSMC process, later there was a clarification that it would be N5P (P = performance), which is quite logical (it makes sense to use a better version of the process, if available). Recently, there were reports that Nvidia could eventually reach for Samsung’s process, and there were even reports that it was specifically the 4nm generation.
We know that the 4nm LPE process is renamed 5nm LPA, so a lower number would not be an advantage over Samsung’s 5nm process, but rather a cheaper alternative.
However, after reports of a not very high yield of 5nm generation, Samsung’s processes are in connection with Nvidia Lovelace (at least high-end) stops talking about the Samsung process and Greymon55 talks about the 5nm TSMC process, specifically the N5. Why not N5P anymore, it is not clear – it may be inaccurate information, as well as, for example, the consequence of the fact that TSMC is no longer free on N5P lines.
Date of publication
When speculation about the time of the release of this architecture began to emerge this spring, a reliable leaker kopite7kimi responded to one of the estimates that the first quarter of 2022 would actually be “much, much, much later.”
This is basically confirmed by Greymon55, according to which to release Lovelace it will occur no earlier than in the fourth quarter of 2022, ie not earlier than a year and a quarter to a year and a half. The relatively distant time window gives room for a refresh of the current generation of Super series models, which would make sense in time at the turn of this and next year. If it happened to them.
The next generation of GPU = response to the crash of mining?
To sum it all up, we have information that Nvidia Lovelace will bring 2-2.2 times higher performance than Ampere. We have information that AMD RDNA 3 will bring to 2.5-2.7 times higher performance than RDNA 2. These huge shifts, whether the manufacturers have decided to achieve them purely on the basis of competition, or due to some higher providence, can have a very important application related to something other than simple competition.
Both manufacturers have practical experience with how the sale of new graphics cards negatively affected the end of the first and second waves of cryptocurrency mining on the GPU. The bazaar market was flooded with cheap cards, many users preferred to buy cheaper used hardware over more expensive new ones, and manufacturers had no way to deal with the situation. Nvidia, which also had a more powerful high-end on offer, whose sales were significantly less affected by the flooding of the market, mostly by older (hence slower) cards, came out a bit better (though far from ideally).
The large next-generation performance shift can thus serve as a tool for both manufacturers to make de facto low-end from current mining hardware, if needed. They can’t eliminate the risk of flooding the market with cheap graphics hardware next year. But they can turn this bazaar hardware low-end to lower mainstream in the eyes of players by releasing products many times faster. In short, these second-hand cards can make significantly less interesting simply by the fact that low cost will be associated with (compared to new hardware) low performance.
–
