Nvidia Pascal Announcement

silent_guy said:
Check for yourself.

For a 22% core clock increase, they see a 13% performance increase, or a a scaling of 0.59.
For an 11% memory clock increase, they see a 5.2% performance increase, or a scaling of 0.46.

So core is definitely a bigger limiter than memory.

Edit: bonus comment from Anandtech's gm200 review:
Click to expand...

I'm not so sure about those numbers. One problem I see with the whole thing is the TDP limit. There's no way to know how it is affecting boost clocks in every case. I think that a +X% increase in base clocks from Maswell to Pascal would have a much more linear increase in performance than comparing overclocks on Maxwell. Anyway, when both core and memory were OCed performance increased pretty linearly too in some games, so I don't think "your rule" necessarily applies.

As for the other post. I worded everything wrong. Again. I didn't even mean 384 bit was ever on the table, just that it might have been if GDDR5X wasn't there to take over. Simply forget about every GDDR5(X) comment I made, I just meant they probably desgined GP104 (full die) to work with more than GDDR5 can offer, let's say, 300 GB/s. I'm pretty sure they knew of GDDR5X existence soon enough. I do think they intended GP104 to have more available bandwidth than 256 GB/s tho, which is kinda obvious, considering GDDR5X support.

I know that GDDR5 wouldn't necessarily be a major bottleneck either, at least when looking at average performance, but as 4 GB GTX 960 owner, I can attest that generally it does pretty well with its limited bandwidth, but I can also attest that in certain cases, some games with some textures and shadow settings, performance really really tanks, and it didn't even hit the 4GB limit, sometimes far from reaching it.
 
Kaotik said:
Well, there has been rumours of 3 models, maybe it's x70, x80 and x80 ti, which of only Ti comes with GDDR5X? (and which should be released a bit later, since Micron still hasn't started mass producing those memories)
Also, just getting more bandwidth doesn't automatically do miracles on performance if the chip already had sufficient bandwidth for most scenarios
Click to expand...

Well, i dont really trust much this rumor, untill Nvidia decide to use the " TI " moniker as a performance marketing point ( understand, use the "high performance" image that consumer can have of it after the 780TI - 980TI ). And want to play the "Higher price " game for the GDDR5x version.

I could be wrong but i imagine the full GP104-400 is still the GTX "1080", the GDDR5 version+ SM disabled the 1070, and the lower one the 1060... the same type of lineup released by Nvidia since Kepler.

Ofc, there's allways some decision who can come from the timeline of release.. We know that GP100 will not be available, even for Tesla, for consumers before early 2017. we can imagine the "gaming card" will not be there too before Q1 2017 or who know, maybe even march april 2017.
 
Last edited:
Benetanegia said:
Anyway, when both core and memory were OCed performance increased pretty linearly too in some games, so I don't think "your rule" necessarily applies.
Click to expand...
If you increase of both memory and core, it's impossible to judge the extent by which the GPU is memory or core limited. That's why it's so important to see numbers were only one such parameter is changed.

The fact that performance increases pretty much linearly when you increase both memory and core at the same time shows that TDP is not a major limiter. And if you increase only one clock, you will stress the system even less, so TDP should be even less of an issue as well.
 
silent_guy said:
If you increase of both memory and core, it's impossible to judge the extent by which the GPU is memory or core limited. That's why it's so important to see numbers were only one such parameter is changed.

The fact that performance increases pretty much linearly when you increase both memory and core at the same time shows that TDP is not a major limiter. And if you increase only one clock, you will stress the system even less, so TDP should be even less of an issue as well.
Click to expand...

With some cards, many people have seen increased performance by increasing the TDP limit alone, without touching clocks at all. Unless we can see performance numbers attached to specific boost clocks, there's no way of telling how much it was gained from the OC and how much by increasing the TDP limit. You could easily be increasing the base and max boost clocks, while the actual (average) boost clock under a certain game is staying the same as before the increase. And likewise, by just increasing the TDP limit, aerage clocks could be higher despite not touching the clock settings.

The fact that performance increases almost linearly when both memory and core are overclocked and only by a factor of 0.46-0.59 when only either one of them is OCed, pretty much goes to show that the card is well balanced in its base values and hence only when this balance is somewhat mantained that you can get proportional improvement.

I know what you're saying, for a relative X% percent performance increase a X% increase in memory bandwidth is not required. However I think that in this particular case we are talking about an specific case where 256 GB/s is simply not enough for the full die, if the full die is supposed to "handily" beat the 980 Ti as some people are suggesting. Even your rule of thumb fails here. I mean, a 50-60% performance increase (over GM204) from a 15% memory OC, still requires another 20% to come from improved bandwidth efficiency, and while not imposible, it seems rather unlikely (like how inefficient were pre-Maxwell cards, if they can improve another 20% on top of Maxwell?). IMO GDDR5X support on GP104 speaks for itself, it was deemed necessary, and I don't think it comes without costs, as to implementing it just for the sake of it. It doesn't even mean they need all th BW that GDDR5X could provide, it simply means that more than 15% increase was necessary. For all we know these cards could come with just 10 Gbps GDDR5X.
 
The third (lowest SKU of gp104) is most likely the mobile version of the x70 chip? That will fit in line with how they launched the gtx 980 and 970 and month later the gtx 970m.
 
Rumor: NVIDIA GeForce GTX 1060 Ti Would get GP104 ASIC

Let's call it rumor Monday today ? It looks like there will be another SKU, or let me rephrase that, there's rumor of another ASIC from NVIDIA. It would be the GP104. Earlier on we indicated the GeForce GTX 1060 to get a GP106 GPU, now there's talk about a Ti model with a GP104.

The rectangular die of the GP104 was measured at 15.35 mm x 19.18 mm which should house (very speculative) a transistor-count of 7.4~7.9 billion. Interesting to see is that this chip is tied towards Samsung 8-gigabit GDDR5 memory chips. These ICs run an effective speed of 8 GHz (GDDR5 and thus not GDDR5X). At 256-bit you'd be looking at 256 GB/s of bandwidth. But this is likely being cut-down to 192-bit for the 1060 Ti model. The shader processor core count of the GP104 should be closer to 2,560, opposed to the 4,096 indicated from an older report.

It is ChipHell that spreads this new rumor, it could be a 6 GB memory card over a capped 192-bit GDDR5 memory bus. The GTX 1060 Ti would feature the ASIC code "GP104-150-A1." After the release of the 1070 and 1080, this one could launch in the November 2016 timeframe.
Click to expand...
http://www.guru3d.com/news-story/rumor-nvidia-geforce-gtx-1060-ti-would-get-gp104-asic.html
 
https://twitter.com/AMDGaming/status/727226204839923712

pharma said:
Rumor: NVIDIA GeForce GTX 1060 Ti Would get GP104 ASIC


http://www.guru3d.com/news-story/rumor-nvidia-geforce-gtx-1060-ti-would-get-gp104-asic.html
Click to expand...

I don't buy it. x80 x70 and x60?! Isn't x60 a large volume SKU? Unless gp104 yields are so bad they are swimming in salvaged parts.

Are this the actual specs or just some wild guess fillers?

http://www.techpowerup.com/gpudb/2839/geforce-gtx-1080
http://www.techpowerup.com/gpudb/2840/geforce-gtx-1070
 
Last edited by a moderator:
SimBy said:
https://twitter.com/AMDGaming/status/727226204839923712



I don't buy it. x80 x70 and x60?! Isn't x60 a large volume SKU? Unless gp104 yields are so bad they are swimming in salvaged parts.

Are this the actual specs or just some wild guess fillers?

http://www.techpowerup.com/gpudb/2839/geforce-gtx-1080
http://www.techpowerup.com/gpudb/2840/geforce-gtx-1070
Click to expand...

Could be guess fillers or ties into the other rumours of the 1070 and 1080 being higher in price compared to the 970 and 980 - I think this will see their sales tank though if they did this as it felt like they were near the limit on what many would pay and indeed can be seen in the number of 980s sold compared to 970s (substantially more).
Consumers really need a pretty good Polaris card to come out and be competitively priced, even if it is only around 1070 performance or close to it but with good price/perf ratio.
That could lead to the speculation that maybe they see the x60 competing with current Polaris....

Cheers
 
You must log in or register to reply here.

Similar threads

F
RDNA4
21 22 23
Replies
448
Views
81K
Dangerman
D
D
NVIDIA celebrates 25 years since the GeForce 256
Replies
14
Views
2K
Roly
R
Shortbread
The PC Gaming Show 2023 (June 11 / 1pm PT / 4pm ET)
Replies
12
Views
932
Silent_Buddha
S
F
ARM Speculation Thread
Replies
8
Views
3K
Arun
A
R
9800X3D releases Nov 7th, review embargo on Nov 6th
4 5 6
Replies
100
Views
7K
Daozang
D
Back
Top