Nvidia Pascal Announcement


So anyone else gone through those tests and notice in quite a few of the 2fs the TitanX is faster than the 1080, albeit marginally at times.
Cheers

Edit:
Thanks to Razor for the link can say the following related to the test setup: Precision Model = SPFP (GPU), Double Precision (CPU)
 
Last edited:
Looking into some local shops, all 1080 sell higher as the FE aka reference design.
The FE is 800 euro, all 5 other's from Asus and MSI sell in the range 800-900 euro. (prices include Tax).
For a reference 900 euro buys you 3 Xbox Ones here.

It's interesting to say the FE will be $700 and the OEMs will sell for $600 but that doesn't reflect what is happening.
 
Looking into some local shops, all 1080 sell higher as the FE aka reference design.
The FE is 800 euro, all 5 other's from Asus and MSI sell in the range 800-900 euro. (prices include Tax).
For a reference 900 euro buys you 3 Xbox Ones here.

It's interesting to say the FE will be $700 and the OEMs will sell for $600 but that doesn't reflect what is happening.
Give it some time. These things are always more expensive right at launch (supply and demand).
 
So they say it is great FP32 is mixed-precision, but would be cooler if the mixed-precision was also applied to FP64 cuda cores as well.
I would say it is a fair possiblity this is where Volta may be heading.
Cheers
Yeah, I figure the fp16 mods were practice for the fp32/fp64 integration. Failure there was an option, it won't be for fp32/64, though. If the fp16 mods aren't in the P0x cores, that might say something interesting about hedging. But, we'll see -- path and energy optimization might trump other interests. It is called "Volta" and not "Legoland" after all....

Also, as long as we're opening the silly season -- I remember how fp24 used to be a thing. fp64 is 2xfp24 + fp16 ;>
 
Looking into some local shops, all 1080 sell higher as the FE aka reference design.
The FE is 800 euro, all 5 other's from Asus and MSI sell in the range 800-900 euro. (prices include Tax).
For a reference 900 euro buys you 3 Xbox Ones here.

It's interesting to say the FE will be $700 and the OEMs will sell for $600 but that doesn't reflect what is happening.
Interesting observation. Here in germany, nothing else than the FE is actually in stock, but maybe that's just us. You can, however, order for around 80-100 EUR below FE's price (and if the shops accept your order, they have to deliver at the ordered price).
 
Looking at those AMBER benchmarks.
Still too early to say but there is possibly a trend between with the 1080FE relative to the P100, Titan X, and 980.

When the 1080FE performance is marginally slower than the Titan X, the performance gap between 1080FE and the 980 drops from 43-45% to 33-35% faster as well.
At the same time the performance gap widens between 1080FE and the Tesla P100, from P100 being 21-25% faster to 47-50% faster.

So even the 980 is behaving more consistent than 1080 for some of those 2fs SPFP benchmarks, while the 1080 also loses relative performance to both P100 and previous gen Titan X in same ones.

Cheers
 
Interesting observation. Here in germany, nothing else than the FE is actually in stock, but maybe that's just us. You can, however, order for around 80-100 EUR below FE's price (and if the shops accept your order, they have to deliver at the ordered price).
It's not just you, there's only FE's for sale right now (meaning are or at least have been in stock), even though customs are already listed.
 
Regarding the fan issue and its inconsistent behaviour with spinning up.

ManualGuzman at Nvidia seems to have confirmed they identified the problem in the driver and a new one will be released:
Update 6/1/16 - We are testing a driver fix to address the random spin up/down fan issue.
Update 6/2/16 - Driver fix so far has passed internal testing. Fix will be part of our next driver release.
https://forums.geforce.com/default/...n-issue-updated-6-2-16-/post/4891787/#4891787
So maybe this issue did not happen with Beta drivers used by reviewers *shrug*.
Cheers
 
Looking into some local shops, all 1080 sell higher as the FE aka reference design.
The FE is 800 euro, all 5 other's from Asus and MSI sell in the range 800-900 euro. (prices include Tax).
For a reference 900 euro buys you 3 Xbox Ones here.

It's interesting to say the FE will be $700 and the OEMs will sell for $600 but that doesn't reflect what is happening.
On Newegg they had the Gigabyte Gaming 8GD for $649.99 today, but a few minutes later was sold out.
 
Regarding the fan issue and its inconsistent behaviour with spinning up.

ManualGuzman at Nvidia seems to have confirmed they identified the problem in the driver and a new one will be released:

https://forums.geforce.com/default/...n-issue-updated-6-2-16-/post/4891787/#4891787
So maybe this issue did not happen with Beta drivers used by reviewers *shrug*.
Cheers
While I think this sheds a telling light on how the Founders Edition is "superior crafted" [...], I can tell you, this has not happened with (our) review card - with neither the launch nor the officially release .25 driver.
 
While I think this sheds a telling light on how the Founders Edition is "superior crafted" [...], I can tell you, this has not happened with (our) review card - with neither the launch nor the officially release .25 driver.
On the plus side Nvidia state they can replicate the issue, and ManuelGuzman is the customer representative of Geforce/Nvidia.

I would assume the reviewer driver did not suffer this issue, but then not everyone is experiencing this on the production drivers *shrug*.

Cheers
 
Nominal voltage is 0.8V for crying out loud!
Yeah I think many forget that 1.25 is a big voltage and the cards are currently locked/limited much lower.
I thought I read somewhere EVGA will be raising this some cards closer to the limit, but that is a big 'thought I read it on one of the OC type forums'
Those who picked up on the 1.25 limit (but this can be overcome with electrical skills) were testing LN2 OC, but well that is rather niche.

Edit:
This is more of a HW design limit as some have already tried this by firmware.
Also to say the url I included below details exactly the requirements to bypass those HW restrictions...
Yeah really not a good idea for most OCers :)
Cheers
 
Last edited:
Doing this as a separate post as it is quite lengthy answer on the 1.25V, which can be overcome.
This is from one of the top OCers, I think works for EVGA:
Voltage scaling and “1.25V limit”
There were some rumors spreading wildly these days regarding “1.25V limitation” or whatever on modified GTX 1080 cards, which requires here few words to explain.

Hardware itself is well capable of getting to that and above voltage output for GPU core, but GP104 GPU itself is more sensitive to voltage than previous generations. Part of it due to thinner physical process, other part due to difficulty of removing heat quick enough from all those tightly packed 7.2B transistors from 21% less surface area. Those overclockers who did 2.1-2.2GHz on GTX 980 Ti’s are aware of all the things required to achieve those high clocks. Same here applies to Pascal generation. So if you can manage to keep GPU cooled well and have good voltage delivery to it, you indeed can push higher voltages. Cards cooled by liquid nitrogen during this guide testwork were able to run 1.35-1.4V, reaching speeds over 2500MHz.

This also brings and answer to the question if overvolting can help OC on aircooling or watercooling. It does not help, due to thermal, which get only worse. Higher temperature render stability and performance decrease. GPU literally overheats and cannot run high frequency anymore, even though temperature is below specified maximum temperature. So just like in 980/980Ti/TitanX case, overvoltage is not recommended, as it gains no performance improvement.
https://xdevs.com/guide/pascal_oc/
Cheers
 
Last edited:
Back
Top