TechPowerUp's Titan X (Pascal) Review

[Genx87]

Makes me wonder with the die sizes increasing so quickly if we are not in for a LONG stagnation period in graphics improvements if we are stuck on 16/14 as long as we were on 28nm. The 1070/1080/titan x are a HUGE jump in performance, did they use up most of the die shrink advantage already?

I think they are just getting started. Titan is topping out at 205 watts on a first run on this process. That performance\watt is pretty impressive imo.

 

[Arachnotronic]

Yes, a die shrink and some changes. Super easy. Just right click on the die in the "die builder" app, choose "resize" then click "send to fab" so lazy, right??

This was good for a hearty laugh out loud. Thank you.

 

[Arachnotronic]

Funny quote nerp. Appreciate the cynicism.

I recently read an post about understanding the big change coming from 28nm gpus to 16/14 (Nvidia/AMD).

The poster indicated that 28 nm has been around for awhile so both companies had plenty of time to refine it.

This is true. TSMC continually refined the 28nm process, so the process had better electrical/performance/yield characteristics in 2014/2015 than it did in 2011/2012. However, the key thing is that both companies -- NVIDIA more aggressively -- made improvements to the underlying architectures which are much bigger drivers of perf/watt than process is.

In this first generation of 16/14 you don't have the depth of refinement yet. You have higher prices to recoup the investment costs etc.

Partially true

Pascal is a "tick" to use Intel parlance, or perhaps a "tick+." The micro-architecture saw some enhancements (better color compression, dynamic load balancing, etc.) but the real work went into doing a darn good implementation of the architecture in 16FF+. The high level architectural features are important, but you have to remember that the low-level circuit implementation is also critical and NVIDIA's work there is why we saw Pascal clocking at such obscenely high levels.

The rumors say Volta is on 16nm and while I have not been able to confirm this independently I think it makes the most sense by far. As we saw in Kepler -> Maxwell, NVIDIA was able to wring out a 2x perf/watt jump and a massive area efficiency improvement on largely the same process (maybe 5-10% better) through a substantially changed architecture.

I think Volta will be a "tock" and it will deliver significant architectural refinements/performance-per-watt enhancements on the same (perhaps slightly improved) 16-nanometer node.

 

[Bacon1]

http://wccftech.com/nvidia-tesla-p100-gp100-june-2016/
https://www.semiwiki.com/forum/cont...erformance-lead-neural-network-computing.html

Says they are coming in Q1 2017 not currently shipping?

 

[96Firebird]

Says they are coming in Q1 2017 not currently shipping?

For OEM supplier shipments, meaning bundled workcenters. It even says it right above the image in the article.

 

[Keysplayr]

Says they are coming in Q1 2017 not currently shipping?

And did you notice what for? Where is your attention to detail all of a sudden? :biggrin:

 

[Keysplayr]

For OEM supplier shipments, meaning bundled workcenters. It even says it right above the image in the article.

I mean it even says shipping started in June right in the headline.

 

[eddman]

That's some healthy performance increase over titan X (M), and to think that its boost is actually limited by that not-so-good cooler, the upcoming 1080 Ti cards with proper 3rd party coolers are going to be something.

More reason to skip titan P, even if you're richie rich.

 

[Deders]

And fan at 100%, and this is a fan that is already noisy at stock settings.

The noise is one of the biggest reasons I'm not too jealous of TitanX owners. That and I'm pretty sure a single 1080 will do nicely when I eventually get a 34" ultrawide. Only half the pixels of 4K.

 

[Arachnotronic]

Says they are coming in Q1 2017 not currently shipping?

I'm going to assume you are genuinely interested in an explanation, so here it goes.

Many of the large hyper-scale data center players (think Facebook, Google, etc.), at which the Tesla P100 is aimed, don't buy servers from the likes of Dell, HP Enterprise, and so on. They design and build their own proprietary systems for their own consumption, and they usually adopt next generation CPUs, GPUs, interconnect fabrics, etc. well before the "off-the-shelf" server guys have machines ready for the small-fries to go and buy.

Right now, the Tesla P100 in Mezzanine form is shipping to those hyper scale customers.

 

[Bacon1]

I mean it even says shipping started in June right in the headline.

Which references this:

Supercomputers and cloud servers are NVIDIA’s first priority which is further proven by the fact that the NVIDIA DGX-1 supercomputing platform will be shipping in June 2016 and other server partners will be able to sell it through their channels in Q1 2017 which is the date when NVIDIA will plan to ship GP100 to OEMs.

from WCCFTech but the newer article

There’s a very good chance that many of NVIDIA’s customers for Pascal may end up designing their own solutions and NVIDIA is simply enabling early adopters to buy a DGX-1 to get ready for when OEM solutions are available at large scale. NVIDIA is selling one DGX-1 for $129,000 and will be delivering them this summer.

Considering that newer article is from the end of June and the wccftech one is from April, are they actually shipping now and in use in datacenters or still coming soon?

Why all the hostility?

 

[Bacon1]

I'm going to assume you are genuinely interested in an explanation, so here it goes.

Many of the large hyper-scale data center players (think Facebook, Google, etc.), at which the Tesla P100 is aimed, don't buy servers from the likes of Dell, HP Enterprise, and so on. They design and build their own proprietary systems for their own consumption, and they usually adopt next generation CPUs, GPUs, interconnect fabrics, etc. well before the "off-the-shelf" server guys have machines ready for the small-fries to go and buy.

Right now, the Tesla P100 in Mezzanine form is shipping to those hyper scale customers.

Well the articles linked make it sound like the only way to get them early was the DGX-1 which was going to be shipped, but no date listed just "this summer".

I am interested because I like big data and want to see if there is any more info about these being used.

 

[Arachnotronic]

Well the articles linked make it sound like the only way to get them early was the DGX-1 which was going to be shipped, but no date listed just "this summer".

I am interested because I like big data and want to see if there is any more info about these being used.

I strongly suggest you don't use WCCFTech as a source for news. Even in the cases in which they don't outright fabricate things, they sometimes misinterpret actual news, ultimately creating noise and confusion for readers.

 

[Bacon1]

I strongly suggest you don't use WCCFTech as a source for news. Even in the cases in which they don't outright fabricate things, they sometimes misinterpret actual news, ultimately creating noise and confusion for readers.

I asked for a source and that was the one provided to me :\

Also that info is from semiwiki not wccftech

 

[Keysplayr]

Well the articles linked make it sound like the only way to get them early was the DGX-1 which was going to be shipped, but no date listed just "this summer".

I am interested because I like big data and want to see if there is any more info about these being used.

The articlw also made it sound like P100 is shipping to super computer builders like what oak ridge does. But OEM 129,000 super computers will be available in the summer.

 

[OVerLoRDI]

I'd love to see how it OC's with some watercooling.

 

[alcoholbob]

Air cooler if you leave it at 70%+ it's stable at about 1.85-1.9GHz. With water you should be able to hit the typical Pascal limit of 2.05-2.1GHz.

It's not an insane jump, maybe 10% faster on water, although it should leave you room to overclock the memory more with less power going to the fan, so maybe up to 15% faster with a hefty memory overclock.