I'm curious if anyone looks at this similarly as myself.
I see Volta (likely 896/1792/3584sp @ higher clocks) as one option, but where is nVIDIA going to go after that? I think it's a serpentine line to more-or-less exactly what AMD is doing with more costly R&D in-between. It's great that they can afford that, and that consumers can reap those specialized benefits, but I don't see it being the end-all of the matter.
Clock scaling is going to end for nvidia's current methodology (think essentially half that of a typical CPU; an 1800-1850mhz GPU being similar efficiency as a 3600-3700mhz CPU...with clock scaling up to ~2.34ghz [similar to ~4.6-4.7ghz CPU]), just as it has with countless other architectures. We have seen it, and will likely continue to see it (just as Intel has been stuck at ~4.3-4.7 for generations). We've seen Apple nary increase the clock of the A9x to A10x...Heck, we've even seen ARM completely revamp A53 to A55...EXCEPT IN THAT CASE it uses pretty much the exact same clocks but more transistors for greater power efficiency. The difference for the last being that it's a more high-density arch that's optimized closer to ~1/3 ratio (just like Polaris or Vega).
When the point arrives that nvidia transitions to 7nm, odds IMHO are pretty safe ~1792/3584sp designs will be shrunk down in an obvious manner, while what we've seen on V100 will evolve into something twice that of GP102/GV104. The problem there-in lies for them to make sense for each market (~75/150/225/300w) core architecture (more xtors) use will have to play a larger role than clockspeed as power consumption for higher clocks likely won't fit into those parameters because it doesn't scale nearly as well.
Hence, we see something like the Snapdragon 820/821->835 (although the same things could be said of the Apple A series).
High power cores ran at 1800mhz for power efficiency (like Pascal). The tuned version ran at 2.15ghz (like Pascal overclocking or probably like Volta stock). The maximum performance version was 2.34ghz (which is probably where Volta will overclock).
On 10nm, that clock was only raised to 2.45ghz (2.38 for Apple), implying a fairly hard wall was hit in terms of scaling at efficient perf/w. It makes sense; it's not far off half clock of the exact same wall Intel hit.
OTOH, low power cores are a different story. While the power-efficient cores on 820 were 1363mhz (like Polaris), and tuned to 1593mhz on 821 (Vega), ON 10nm WE SEE THEM SCALED TO 1900MHZ. A much lower ratio between the two...and I imagine that will shrink even further on 7nm as higher-density cores may clock even slightly better while larger cores stay more-or-less stagnant.
I don't know how else to demonstrate this should be one of those 'a-ha' moments for why high density/more fixed units/parallelism is key, and preparation for it's inevitability isn't a terrible idea. AMD has done it with Zen, and is doing it again with Vega.
In the same way one can look at the efficient ~3.5-3.8ghz all-core aim for Zen (or 1/3 of that for Polaris/Vega) and clearly see a way forward to (at least boost clocks of) ~5ghz on 7nm for the CPUs, one could look at Vega and see a way forward to maximizing the use of 1ghz (say equal to a 1080/1170 and good for 4k30) or 1.2ghz (4k60) HBM2 in perhaps 150/225w power envelopes.
I'm not going to sit here and defend Vega as it sits; obviously nVIDIA made the best call for THIS node size (and will likely continue it to 12nm). The shader/rop/voltage/clock capability/power consumption/transistor library etc ratios just make more sense at this juncture. AMD went all-out on density and high compute and it straight-up didn't work because they were some combination of too poor to completely revamp the design or too early for it to make sense.
That said, and obviously this is simply only my opinion, AMD has laid a groundwork for both CPUs and GPU designs that should both transition incredibly easily to 7nm while exploiting them to the fullest (without requiring a completely new arch right away)...from the fab that is cranking out high performance parts on that process first...and that ain't for nothing. Sure, they could have made Zen with less cores and higher clock (and perhaps IPC) on 14nm and straight up lost to Intel on Intel's ground. Instead, they did something different that capitalized on the current process' strengths with an eye toward the very near future not only for application (of more cores) but scalability (7nm). Sure, they could have made Vega with less CUs and higher clocks and lost to nvidia because 14nm is less performative (but inherently more dense) process than TSMC's 16nm...but they didn't. Instead they buckled down, got their architecture advancements in place, and made the best of the current situation with a distinct way forward in the next 6 months to a year that will likely only be equaled from nVIDIA two generations from now.
I don't mean to sound apologist, but when I see Intel (with their higher IPC/clockspeed potential designs) scaling clocks back on Coffee Lake to EXACTLY where AMD aimed their core CPU design on a similar but less performative process, I laugh. When I read stories that Intel is having problems scaling clocks past the exact point they've been stuck (for max 'normal overclocking' clocks) forever on 10nm, I laugh.
When I foresee nVIDIA essentially recycling their arch to save them money (can't we all see a ~200mm2 32 ROP 1792sp/~2100+mhz design replacing 1070 and a 400mm2 64 ROP 3584sp/2100mhz+ design being sold at WAYYY too much money to get those last extra 4k60 frames) and calling it ingenious circumvention of Moore's Law instead of planned obsolescence of their former architecture to perpetuate constant sales, I cry.
When I look at Zen and Vega, I see potential. I can understand where they're trying to go with this and how they currently had to make the best of a less-than-ideal situation (with regards to process and likely R&D budget), even if the first iteration isn't perfect. Maybe Zen 2, even if it clocks well, will only catch up to the overall performance of Skylake per core (after adjusted clocks)...but that's okay. Their R&D can take the extra time they need to pump up the arch over the long road of 7nm DUV and EUV for Zen 3, allowing strengths in different areas than their competition on the road to perhaps eventual parity, while in the meantime adding healthy competition or better pricing. Likewise, maybe a shrunken Vega only at most competes with GV104 (at half die size and probably more than half power), or even it's castrated version (2688sp?), but that too is okay. The potential is there for awesome 4k30/4k60 parts that are better/cheaper than 1080, more performative than any potential 1792sp design that would likely be of similar size (and if nvidia has their way, cost) and could perhaps even scale to 2x within 300/375w for our eventual 4k60+/4k120 hdmi 2.1 future.
In short, AMD's designs make me hopeful.
And really....Is that so bad?
