Nvidia Pascal Announcement

Looking at EVGA is interesting with the screenshot of Precision within the GPU card product page.
One of their upper 1080 models is showing 2012MHz with 1050mV, while for their upper 1050ti models is showing 1797MHz with 1093mV.

Down on the 1080, but too early to tell if power constraint (no idea what they are doing with 6-pin), thermal related to the Samsung node and maybe density, or possibly the Samsung silicon-node voltage-frequency-performance envelope.

Cheers
 
Wasn't one of the iphones SoCs A# during what some dubbed the "chipgate" manufactured by both Samsung 14nm finfet and TSMC 16nm finfet actually faster and superior performance and batterlife on TSMC 16nm processes?
 
Looking at EVGA is interesting with the screenshot of Precision within the GPU card product page.
One of their upper 1080 models is showing 2012MHz with 1050mV, while for their upper 1050ti models is showing 1797MHz with 1093mV.

Down on the 1080, but too early to tell if power constraint (no idea what they are doing with 6-pin), thermal related to the Samsung node and maybe density, or possibly the Samsung silicon-node voltage-frequency-performance envelope.

Cheers

Good find. Lets see what the overclocking results on the cards with 6 pin PCIE power connectors are.
Wasn't one of the iphones SoCs A# during what some dubbed the "chipgate" manufactured by both Samsung 14nm finfet and TSMC 16nm finfet actually faster and superior performance and batterlife on TSMC 16nm processes?

Yeah that is correct, TSMC's version had the lead in both but it was like marginal at best in both.

I dont think it was conclusively proved that one process was better than the other. Those tests basically compared one of each and there is basically no meaningful conclusion you can draw from a single sample of each.
 
Ah I just realised all of us forgot one thing with regards to the comparison between Samsung and TSMC and the lower frequency/spec on the 1050/1050ti.
Is it possible that the performance and spec (reduced reduced cores/TMU relative to performance) of the 1050ti is close to the real-world limits of its 128-bit bus and so having the same clocks as other Pascal is pointless.
The 950 consistently managed to get more relative performance out of its OC than the higher TFLOPs 960, when both are overclocked over 1450MHz, furthermore the 960 really does not have the performance gap one would expect for custom cards even before the OC (see below links but need to change to specific games), which reduces even more at the max OC 1450-1500MHz.
Both these cards were 128-bit bus like the 1050/1050ti, with the 1050ti (2.1 TFLOPs) closer to the 960 (2.3 TFLOPs) than 950 (1.57TFLOPs) in terms of FP32 compute.

Trying to find a benchmark was a bit tricky that had comparable test and OC for the 950 and 960 but Guru3d have both running around core 1475-1500MHz, can be seen when looking at the max OC results to each card and also their performance in games (not all games are used for each); best not to try and calculate gain from the custom card standard clocks to custom card max OC (both shown on the OC page for 950) as that is a skew data point, would need to be standard reference clock instead.
Also worth noting these are just shy of the best cards benchmarked in each range..
http://www.guru3d.com/articles_pages/asus_geforce_gtx_950_strix_review,30.html
http://www.guru3d.com/articles_pages/msi_geforce_gtx_960_gaming_oc_review,26.html

Anyway just throwing this out there, which is probably a factor along with the other points raised as potentials regarding clocks to power demand and unknown specifics about the reported clocks, or density/thermal constraints of the node, or voltage-performance-frequency envelope of the silicon-node.

Cheers
 
Last edited:
true that's why I said marginal at best :)

Yep but I wouldn't even call it that :)
Ah I just realised all of us forgot one thing with regards to the comparison between Samsung and TSMC and the lower frequency/spec on the 1050/1050ti.
Is it possible that the performance and spec (reduced reduced cores/TMU relative to performance) of the 1050ti is close to the real-world limits of its 128-bit bus and so having the same clocks as other Pascal is pointless.

If that was the case they could have very easily gone with 8 Gbps GDDR5 instead of 7 Gbps. And if we compare 1050Ti to 1060, which has 10 SMs (66% more) and a 192 bit bus, taking into account the higher clocks, 1060 has about 80% more shader/TMU performance for only 70% more bandwidth.
 
....
If that was the case they could have very easily gone with 8 Gbps GDDR5 instead of 7 Gbps. And if we compare 1050Ti to 1060, which has 10 SMs (66% more) and a 192 bit bus, taking into account the higher clocks, 1060 has about 80% more shader/TMU performance for only 70% more bandwidth.
My point is that it may be pointless for the 1050ti to have core clocks anywhere like the other Pascal models.
It is best to not bring the 1060 into this because it is not comparable in any way as it has much higher Cuda cores/TMUs/ROPs/Mem clocks along with a 192-bit bus; the closest comparable trend is the 960 and 950, even then the 960 has more Cuda cores/TMUs than the 1050ti but the 950 is pretty comparable apart from TFLOPs and that is closer to the 960 for the 1050ti.

The trend is that the 960 on spec has much higher capabilities than the 950 while both being 128-bit bus, but the real-world does not translate to those differences and when both are OC to 1500MHz the gains become closer.
In other words the 960 has marginal gains when it comes to performance, especially when it is custom vs custom cards as seen with the two reviews I linked earlier.
You cut out the rest of my post that made the point.
If you check the links earlier you would see at max OC around 1475MHz - 1500MHz:
Bioshock
960: 78 fps
950: 75 fps

Tomb Raider
960: 90 fps
950: 83 fps

Hitman Absolution
960: 57 fps
950: 56 fps

3dMark Firestrike
960: 7461
950: 7003

Apart from Tomb Raider they are moderately close (although that is still only 8.5% faster), considering their specs:
960: 1024 cores, 64 TMUs, 32 ROPs, 128-bit bus, 7000MHz effective mem clock, 2.3TFLOPs
950: 768 cores, 48 TMUs, 32 ROPs, 128-bit bus, 6600MHz effective mem clock, 1.57 TFLOPs.
Hence why I raised is there any point for the 1050ti to have clocks any higher than potentially 1700MHz core clock OC that was seen within an EVGA precision screenshot for one of their 1050ti, and why the official clocks are much lower.
However to re-iterate I do mention it could possibly be one of the other factors mentioned, but the trend of 950 and 960 does raise this as a pragmatic real-world reason.
Just as a reference the spec of the 1050ti seems to be 768 cores, 48 TMUs, 32 ROPS, just like the 950, while the memory clock is near identical to that of the 960 at 7012MHz and also with 2.1 TFLOPs FP32 that is not too far behind the 960.

And the caveat before anyone comments about the BGT G1, as I mentioned the two cards I am using are just below the top ones for both, so the % trend does not change.
Cheers
 
Last edited:
Rant incoming, skip this post if you don't want to see me rant.

AAAUUUUGGGHHHH. OK, I'm almost ready to sell my 1070. Nvidia's drivers are so F-ing pathetic if the game isn't either just released or being used for benchmarks. I just had to disable all lighting, shaders, and post processing on the 1070 in order to do one of the puzzles in Guild Wars 2 because having those on meant I couldn't see the dots on a stone face in order to solve the puzzle. F-U Nvidia.

Regards,
SB
 
Rant incoming, skip this post if you don't want to see me rant.

AAAUUUUGGGHHHH. OK, I'm almost ready to sell my 1070. Nvidia's drivers are so F-ing pathetic if the game isn't either just released or being used for benchmarks. I just had to disable all lighting, shaders, and post processing on the 1070 in order to do one of the puzzles in Guild Wars 2 because having those on meant I couldn't see the dots on a stone face in order to solve the puzzle. F-U Nvidia.

Regards,
SB

Read 'nvidia implements a driver feature to make Guild Wars 2 more challenging' :p

All jokes aside, that sucks. Are you able to identify which driver update brought this change/bug?

I haven't play guild wars 2 in a while (don't get me started on that because my rant will kill yours) but I never had any graphical issues like you describe.

Oh and to be fair I play tons of games that aren't just released or used for benchmarks and I've never had any issues, at least not that I remember.
 
Read 'nvidia implements a driver feature to make Guild Wars 2 more challenging' :p

All jokes aside, that sucks. Are you able to identify which driver update brought this change/bug?

I haven't play guild wars 2 in a while (don't get me started on that because my rant will kill yours) but I never had any graphical issues like you describe.

Oh and to be fair I play tons of games that aren't just released or used for benchmarks and I've never had any issues, at least not that I remember.

Not sure which driver update. The last time I did that puzzle was on an AMD card. At the time my friend that was with me with a GTX 970 was wondering how I was solving the puzzle. I tried explaining it multiple times, but they couldn't see it. I wonder if they were having the same issue on their 970.

Regards,
SB
 
guild wars 2 has a history of bad, bad issues with nvidia. Not sure who's the culprit here, but if it don't work, chunk the card and get AMD. Only real solution for the consumer.
 
guild wars 2 has a history of bad, bad issues with nvidia. Not sure who's the culprit here, but if it don't work, chunk the card and get AMD. Only real solution for the consumer.
the other solution is not to play guild wars 2 :p for some reason the giant glass dome structure in the Human capital simply would not render entirely for me in my nehalem i7, once I upgraded to haswell I finally saw it in its entirety. I wish I had taken screenshots, for years I thought that was simply how it looked lol
 
Nvidia to sell Titan X via channel partners instead of graphics card players
Wednesday 26 October 2016
Nvidia so far only plans to release its Founders Edition model, leaving Taiwan-based graphics card players such as Asustek computer, Gigabyte Technology and Micro-Star International (MSI) on the bench over the business opportunity, according to sources from the upstream supply chain.

Taiwan-based graphics card players have accused Nvidia of trying to enter the high-end branded graphics card market with its Founder Edition products, but company CEO Jen-Hsun Huang said that the Founder Edition products are meant to help resolve graphics cards' design problems and are not meant to compete against its partners.

For its previous Founder Edition products, Nvidia had its downstream graphics card partners help sell them in the channel; however, for the Titan X, Nvidia has decided to have channel distributers in each region sell the cards. In Taiwan, Nvidia's Titan X graphics card will be distributed by channel distributor Synnex. With the strategy, Nvidia is expected to see strong profits from sales of the Titan X.
http://www.digitimes.com/news/a20161026PD202.html
 
A little birdie tells me the GTX 1080 Ti launch is around the corner...which makes sense since the holiday season is almost upon us.
My point is that it may be pointless for the 1050ti to have core clocks anywhere like the other Pascal models.
It is best to not bring the 1060 into this
because it is not comparable in any way as it has much higher Cuda cores/TMUs/ROPs/Mem clocks along with a 192-bit bus; the closest comparable trend is the 960 and 950, even then the 960 has more Cuda cores/TMUs than the 1050ti but the 950 is pretty comparable apart from TFLOPs and that is closer to the 960 for the 1050ti.

So you claimed maybe the 1050Ti dosent have the perf/resources compared to other Pascal chips..but when I compared it to the 1060 you now want to omit it? :rolleyes: I compared Pascal to Pascal..I dont understand why you have an issue with that.

I reiterate my point..if the 1050Ti was actually limited by its 128 bit bus (To the extent of anywhere close to 20% of its clock speed)..NV would just have put 8 Gbps GDDR5 instead of 7 Gbps for a "free" 14% perf increase.
 
So you claimed maybe the 1050Ti dosent have the perf/resources compared to other Pascal chips..but when I compared it to the 1060 you now want to omit it? :rolleyes: I compared Pascal to Pascal..I dont understand why you have an issue with that.

I reiterate my point..if the 1050Ti was actually limited by its 128 bit bus (To the extent of anywhere close to 20% of its clock speed)..NV would just have put 8 Gbps GDDR5 instead of 7 Gbps for a "free" 14% perf increase.
The memory/bus spec chosen primarily comes down to budget/TDP focus/good performance at this price level rather than all out performance (although it is great for the 1050ti relative to other cards at its tier now and in the past).

You seem to be ignoring the custom 950 and 960 scenario that is applicable to the 1050ti, even though I provided real world results for you showing how close they are in performance compared to their spec and how the gap reduces further when both are OC'd, those 2 cards are the closest to the 1050ti generation and not 1060 you used, professional reviewers are using the same context as me when comparing to other Nvidia cards; you did look at the specs of those cards and the trend when the core clocks are pushed with custom models in the information I provided.
And my point is when you clore clock it closer to other Pascal cards and their upper custom clock ceiling (say 1800MHz and above).
We are going to disagree so lets leave it at that as you keep reducing my point and context, and I guess we will both have to wait for a more in-depth analysis on frequency-volts-performance across the whole envelope.
Cheers
 
Last edited:
Gigabyte launches G250 with 8x GTX 1080 passive server solution

Where previous Gigabyte G250 HPC servers incorporated up to 8x Nvidia Tesla K80 GPU accelerators, the new model can fit up to 8x passive GTX 1080 boards. This configuration is now in mass production and destined for European distribution shortly.

Gigabyte boasts that the new G250 offers the "1st GPGPU server with passive-consumer GPU cards application". It reckons such a system will be picked up to provide 'VR Ready' processing on tap to the server users.

Beyond the specification of up to 8x passive consumer GTX1080 boards, which make use of the Gigabyte G250's optimised built-in cooling solution, the server has the following standard specs:

  • 2U form factor
  • 2 x Intel Xeon E5-2600 V3 / V4 processor slots
  • 24 x DDR4 DIMM slots
  • 2 x GbE LAN ports
  • 8 x 2.5-inch hot-swappable HDD/SSD bays
826102ad-ec93-459b-b7b2-7007d469269e.jpg
 
Notebook 1050 Ti Specs and Benchmarks

Base core clock 1490 MHz
Boost core clock 1624 MHz
CUDA cores 768
Memory bandwidth 112.1 GB/s
Memory type and size GDDR5, 4GB
Memory interface 128-bit
ROPs / TMUs 32 / 64

http://laptopmedia.com/highlights/nvidia-geforce-gtx-1050-ti-laptop-specs-and-benchmarks

7-10% faster than mobile GTX 970 (100W), up to 86% faster than GTX 960. Performance on par with the desktop version based on the benchmark results.
 
Notebook 1050 Ti Specs and Benchmarks

Base core clock 1490 MHz
Boost core clock 1624 MHz
CUDA cores 768
Memory bandwidth 112.1 GB/s
Memory type and size GDDR5, 4GB
Memory interface 128-bit
ROPs / TMUs 32 / 64

http://laptopmedia.com/highlights/nvidia-geforce-gtx-1050-ti-laptop-specs-and-benchmarks

7-10% faster than mobile GTX 970 (100W), up to 86% faster than GTX 960. Performance on par with the desktop version based on the benchmark results.

Huum, so they have set the 1050TI in laptop at 200mhz higher than the desktop counterpart. thats a first.
 
Last edited:
Back
Top