Nice, you've just put me off upgrading to a 1080... Much appreciated from ye old wallet
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
NVIDIA Volta with GDDR6 in early 2018?
- Thread starter Kaapstad
- Start date
Nice, you've just put me off upgrading to a 1080... Much appreciated from ye old wallet
Certainly not, I had a evga 980ti, the one below the kingpin. It sat around 1.4 - 1.5ghz and my 1080 has a healthy performance delta.
At least 15 - 20 percent.
I am aiming personally at an early 2019 new build, Icelake is exciting as is volta.
I may be wrong I just thought that was the case. If the "2080" is truly 20% faster tan the 1080Ti then gimmie gimmie gimmie.Nice, you've just put me off upgrading to a 1080... Much appreciated from ye old wallet
Soldato
- Joined
- 26 Aug 2004
- Posts
- 5,162
- Location
- South Wales
15-20% is not much if you're already getting low minimums at 4k though.. but I'm @ 1440p so i don't care about 4k performance just yet.
1070 was closest to the 980ti.
The volta 80 card will be at least that much more than the 1080ti.
Depends how low the minimums are, a 1080ti copes with 4k quite well if you can play without AA and knock a setting down that has little to no impact on IQ but takes away a few frames. With new games pushing newer hardware though it's always going to be cat and mouse, we're still a number of years away from 4k fully maxed on a single GPU without frames dropping to levels where we notice the stutter and then we have the onset of 4k 120hz+ with the new HDMI standard and DP 1.3 being utilised.
Agreed. That'll go nicely with my planned monitor upgrade!
It's complicated.
If you overclock both the 980 Ti and 1080 the Pascal card is about 25% faster.
If you run a Kingpin 980 Ti on LN2 it will beat any 1080 on LN2.
Maxwell is more efficient per clockcycle than Pascal.
For normal use though a 1080 is about 25% faster than a GTX 980 Ti.
Soldato
- Joined
- 26 Aug 2004
- Posts
- 5,162
- Location
- South Wales
2+ years is when i plan to upgrade to 4k anyway, and I've only had my 1440p 165hz Gsync screen around 6 months or so.. so yeah GPU's should be nice for 4k high hz by then.
Will it? how do you know?
Should see the 2080 in April at the VERY latest, and possibly January at the earliest. A timescale other than that would be very abnormal for Nvidia.
Also the 2070 should match the 1080 Ti, with the 2080 being 20-30% faster than it.
Then the 2080 Ti's launch is partly dependent on what AMD are doing. It could come as early as Q4 2018, but if AMD's Navi is bad and/or only a mid-range one comes out first, it could be as late as Q2 2019.
How do you know? I am just curious as to what information you are basing your performance figures on?
A 1080ti is 55% faster than a 1070 and 35% faster than a 1080 roughly.
There is no real die shrink this time. Whereas the jump from Maxwell to Pascal was nearly a double die shrink. Isn't a 30% jump more than likely? Wouldn't that be really good in fact?
Are you really expecting a performance jump bigger than the one between Maxwell and Pascal?
I am not saying you are wrong or right, just wondering how you are estimating the performance.
Well firstly it'd be unprecedented in medium-term history for that not to be the case. It's certainly the way things have been since the 680 and 670 launch.
Also the 1080 Ti is more like 62% faster than the 1070. FE to FE. But then it can be overclocked relatively more as the base clocks on the 1080 Ti are lower, so it ends up ~65-70% faster comparing good aftermarkets. And as another note, your figures there would put the 1080 as only 14.8% faster than the 1070, whereas it's ~20-25% faster.
Also also 28nm to 16nm was certainly not anywhere near a double die-shrink. It is just 20-22nm with FinFets. So it's a normal 1 node shrink, with some slightly better power characteristics. Otherwise you'd expect the 1080 Ti to be ~3x the performance of a 980 Ti while also being a little smaller.
For reference, 16/14nm to 7nm will be slightly more than 1 node jump. About a 1.5x jump particularly when EUV is included. Interestingly though the 7nm node is kind of 2 nodes in 1. 1 size for low-power (mobile phone chips) and 1 size for high-performance. The high performance is a 1.5 node jump (i.e. greater than 2x density), and the low-power is a 1.75-2 node jump (i.e. 3-4x density.
Then Kepler to Pascal involved no process improvement, let alone a shrink. But 780 Ti to 980 Ti was a ~43% increase, not taking into account the 980 Ti being cut-down compared to the 780 Ti being full, and also that the 980 Ti overclocks like a monster. The overclocked gap was over 70%.
And lastly there is a clue in the specs for the V100 chip. The V100 is 50% higher performance per watt than the P100. Taken from the FP32 performance on the 250W card of each respectively.
In summary:
- This is like a repeat of Kepler -> Maxwell
- The die sizes will increase (Titan Xp is only 471mm2) and it's a new arch
- This time there is a process improvement, albeit not crazy, whereas before there was none at all
- The V100 is 50% higher perf/w than the P100
- We should expect the same ballpark of 780 Ti -> 980 Ti when comparing the largest Pascal to largest Volta
- xx70 has matched/beaten xx80 Ti from previous generation since the 600 series at least
- And also bear in mind Pascal isn't fully optimised for DX12 and Vulkan. So there are reasonable gains to be made just by adding full support for those, in terms of future games
Last edited:
Kepler to Pascal no process improvement and no die shrink? Maxwell was the process improvement, Pascal was the Die shrink and a double die shrink at that.
You can't base it on overclocks, not all cards overclock the same, and besides the 780TI was a really good overclocker too. As for been cut down, it was cut down to make it perform better in gaming.
I don't know what way you are reading my figures, but, 55%-35% =20%
Anyway, whatever way you look at it, a 50-70% jump in performance to bring the 1070 in line with the 1080ti is going to be asking a lot for Volta.
You can't base it on overclocks, not all cards overclock the same, and besides the 780TI was a really good overclocker too. As for been cut down, it was cut down to make it perform better in gaming.
I don't know what way you are reading my figures, but, 55%-35% =20%
Anyway, whatever way you look at it, a 50-70% jump in performance to bring the 1070 in line with the 1080ti is going to be asking a lot for Volta.
Put it this way, at the moment it is only Nvidia delivering on performance, so I wouldn't write them off yet on Volta. A number of posters on here, who I might add have been pretty arrogant in their opinions, have been proven very wrong in the past.
There was no process improvement with Kepler to Maxwell, other than it being a more mature process. TSMC's 12nm process is an actual improved/tweaked process. It is a very minimal shrink, but has significantly better power performance.
Kepler -> Maxwell was essentially pure architecture improvement, then Maxwell -> Pascal was process + shrink + minor arch tweaks.
Pascal -> Volta is process + very minor shrink + new arch. Therefore it is a greater difference then Kepler -> Maxwell, and that was a big jump on its own.
Also, again, 28nm to 16nm was not a double node shrink. It was a single with some improved power from the FinFets. 16nm is around 20nm or 22nm with FinFets added in real size. The '16nm' is marketing.
If it was a double node shrink then Pascal, Polaris, and Vega are all the worst architectures in history. As you'd expect a ~5x perf/w improvement from a double node shrink and arch tweaks/changes combined. So a 1080 Ti should be close to 5x the performance of a 980 Ti.
Not sure what you mean by this. The 780 Ti was the fully enabled largest Kepler chip, and the 980 Ti was a partially disabled largest Maxwell chip. So a fairer comparison is the original Titan vs the 980 Ti, since they're both partially disabled chips. And that puts the 980 Ti further ahead as the 780 Ti is faster than the original Titan.
And also the 780 Ti (and certainly the Titan) were worse overclockers than the 980 Ti relatively. The 980 Ti could overclock ~25-30%.
Percentages are multiplicative not additive. What I said is correct, you're miscalculating. And this matters as your figures are ~35% off the real figures.
i.e. 20% is 35% more than 14.8%
The 2070 will be a larger die than the 1070, and the P100 is 50% higher perf/w than Pascal according to Nvidia. The combination of 50% higher perf/w and a larger die could easily hit 70%.
Plus the 2080 HAS to be faster than the 1080 Ti, and thus needs to be ~70% faster itself (i.e. 2080 would be ~70% faster than 1080, in order to end up ~25% faster than 1080 Ti). So unless the 2080 isn't faster than the 1080 Ti, the 2070 is automatically going to be ~70% faster than the 1070. And the 2080 will definitely be faster.
Last edited:
I don't get you calculations. GV100 is already a larger die as GP100 and get's 50% perf/w. No idea how you're getting to 70% from that. Other point is, that only the Nvlink GV100 is 50% better in perf/w. PCI-E Version is only 40% better in perf/w. So i'll take this as a starting point and expect 2080 to be 40% faster than 1080, which makes it ~10% faster than 1080Ti. That's way more realistic than your stuff.