Nvidia Pascal Announcement

[mczak]

- Which could explain the significantly lower max. boost clocks of 1.4GHz than the other Pascal GPUs.

I don't think 14nm Samsung is really all that much worse. GTX 1060 is a 120W card. Now if you just look at SMs, yes you could (just) do 60% of it within 75W. But (unlike gm107/206) everything else seems to be 100% of GP106 - same amount of ROPs, same memory interface (ok maybe with a bit less max clock), even same number of GPCs/Rasterizers (if you believe anandtech's diagram). Which is also shown by the transistor count, which is 3/4 of GP106. So, if you'd take that number (that's quite simplified calculation, I know), you'd end up with 3/4 the power draw of 120W with the same clocks - IMHO there's no way they could have achieved these clocks on 16nm TSMC within 75W neither.
Only with the cards drawing additional power from a pcie connector can you deduce anything if 14nm Samsung is really worse.

- 3.3 Billion transistors, 135mm^2 (P11 is 3B, 123mm^2)

At least 14nm Samsung is denser according to these numbers by about 10% compared to GP106...

According to performance estimates, it looks like the GTX 1050 Ti will trade blows with the RX 460, with the typical "above in DX11, below in DX12/Vulkan" results we've seen as of late.

I'd expect better, the 1050 <75W should be roughly on par with the RX 460, the 1050Ti should clearly be above. Albeit a >75W 1050 is possibly going to beat a 1050Ti <75W, if nvidia is going to allow that...

 

[Deleted member 13524]

But (unlike gm107/206) everything else seems to be 100% of GP106 - same amount of ROPs, same memory interface (ok maybe with a bit less max clock), even same number of GPCs/Rasterizers (if you believe anandtech's diagram).

I just had to go back to make sure. GP107 has 32 ROPs at 128bit, whereas GP106 has 48 ROPs at 192bit.
The GP107 is 60-66% of GP106 in pretty much every amount of functional units and TDP, but clocks are taking a 15% (base) to 20% (boost) hits.

Though GP107's clocks are still ~20% higher than Polaris 11's.

 

[kalelovil]

Only with the cards drawing additional power from a pcie connector can you deduce anything if 14nm Samsung is really worse.

Gigabyte 1050(ti) cards with 6pin auxillary power have ~300mhz lower boost clocks than the equivalent 1070/1080s.
http://www.gigabyte.co.nz/products/comparison/list.aspx?ck=3&pids=6054,5916,5915,6061
http://www.gigabyte.com/fileupload/product/3/6061/20161017102136_big.png

Although some of that may be down to yields and familiarity, Nvidia hasn't used a non-TSMC source for quite a while.

 

[Razor1]

If the 6 pin connector is optional (and I think it is optional for all 1050's and ti's), then they will limit the boost clocks to maintain the 75 watts over the pci-e connector.

 

[mczak]

I just had to go back to make sure. GP107 has 32 ROPs at 128bit, whereas GP106 has 48 ROPs at 192bit.

Oops yes my bad. For a second I forgot GP106 is 48 ROPs/192bit.

The GP107 is 60-66% of GP106 in pretty much every amount of functional units and TDP, but clocks are taking a 15% (base) to 20% (boost) hits.

Yes, the chip having narrower memory interface changes things quite a lot. The power and die size saved there should be significant. Indeed it's 60% SMs, 66% MC/ROPs. Only GPCs/Rasterizers are the same amount (plus of course the non-duplicated fixed stuff). So I'll reverse my statement and say the opposite - it does look worse.
(Even the transistor density doesn't look all that much more dense now, since MC take considerable amount of area compared to the number of transistors they use. Though even considering this it might still be slightly more dense - it is 75% of the transistors in 66% the area still.)

 

[mczak]

Gigabyte 1050(ti) cards with 6pin auxillary power have ~300mhz lower boost clocks than the equivalent 1070/1080s.
http://www.gigabyte.co.nz/products/comparison/list.aspx?ck=3&pids=6054,5916,5915,6061
http://www.gigabyte.com/fileupload/product/3/6061/20161017102136_big.png

Ok these numbers are lower than what I would have expected for cards with additional connector (just barely more than reference at least in the non-oc state, and oc doesn't add much neither).
Maybe it really is ~15-20% worse then, though I'd like to see the clock/power curve in practice first.

 

[Picao84]

Looks very disappointing indeed. It looks like GM107 with additional features that were missing from Maxwell 1 and more ROPS. I have a feeling that its performance will be all over the place varying wildly with the game engine employed. Since an OC RX460 is already at GTX950 performance, if GP107 does not manage to beat its immediate predecessor (in product line up since that card uses a cut down GM206), this card is DOA, especially if it predictably follows the trend of better DX12 / Vulkan performance set out by RX480 against GTX1060. Lets hope benchmarks show a better light, otherwise this generation's low end GPU offers are truly terrible.

https://www.techpowerup.com/reviews/ASUS/RX_460_STRIX_OC/24.html

 

[Deleted member 13524]

That 6pin connector in the G1 models is giving them a 100MHz boost, at least in factory clocks. GP107 definitely has some more overclocking headroom compared to P11 (and so did Maxwell compared to GCN), but it's a lot lower comparable to the TSMC Pascal chips.

This leads us to the question of why nvidia chose to go with 14nm Samsung instead of 16nm TSMC.
Given TSMC's seemingly better performance, maybe the price per die area is substantially higher and nvidia wasn't willing to have their lower-margin product using a more expensive process.

AMD's willingness to pay GF to use other foundries was already a sign that TSMC's 16nm was getting better results overall, but now we actually have an idea of how much.



EDIT: worthless

 

[Picao84]

Now you know why we should wait for the cards to come out before singing the demise of X or Y GPU?

Just as you should before talking about the end of dedicated GPUs caused by APUs. Plus, hold your horses since we still don't know its real performance yet.

This generation's low-end $110 GPUs are offering the performance of last-gen's mid-end ~$150/170.
How is 30% less money for same performance a truly terrible thing? You just have to acknowledge that priorities are very different in this market segment.

I'm not talking about price, but the chip itself. GM107 aka GTX750Ti was a huge leap (especially so considering it was done in the same 28nm node) for the low end in terms of performance for cards under 75W. The GTX950 the GTX1050Ti is looking to replace has a TDP of 90W, not that far from the latter 75W. If they have the same performance then I believe this to be worst jump in perf/watt between GPUs done in different process nodes in quite a while.

But even if we talk about price, if all the GTX1050Ti can do is match a GTX950 (at the same price point!), then there is no way the vanilla GTX1050 has increased performance for less money. In fact there would not have been any change at all in perf/[insert you currency here]!

Are you looking for me to retract my statement that RX460 is a terrible chip? I won't, since it is as bad as GP107, they are both horrible. On the other hand, is it a good buy? Sure it is, but that is more because of nVIDIA's apparently spectacular failure (which is not even worse because this is not a halo product), than any merit from AMD.

 

[ieldra]

Just as you should before talking about the end of dedicated GPUs caused by APUs. Plus, hold your horses since we still don't know its real performance yet.



I'm not talking about price, but the chip itself. GM107 aka GTX750Ti was a huge leap (especially so considering it was done in the same 28nm node) for the low end in terms of performance for cards under 75W. The GTX950 the GTX1050Ti is looking to replace has a TDP of 90W, not that far from the latter 75W. If they have the same performance then I believe this to be worst jump in perf/watt between GPUs done in different process nodes in quite a while.

But even if we talk about price, if all the GTX1050Ti can do is match a GTX950 (at the same price point!), then there is no way the vanilla GTX1050 has increased performance for less money. In fact there would not have been any change at all in perf/[insert you currency here]!

Are you looking for me to retract my statement that RX460 is a terrible chip? I won't, since it is as bad as GP107, they are both horrible. On the other hand, is it a good buy? Sure it is, but that is more because of nVIDIA's apparently spectacular failure (which is not even worse because this is not a halo product), than any merit from AMD.

I imagine this chip will behave like all the other nvidia ones. Boost clock on the box says 1455, I expect you to see higher clocks in reality.

I sincerely doubt this is only as fast as a 950

 

[Picao84]

I imagine this chip will behave like all the other nvidia ones. Boost clock on the box says 1455, I expect you to see higher clocks in reality.

I sincerely doubt this is only as fast as a 950

If we go by the slides shown here, it will be around 25% faster:
http://wccftech.com/nvidia-geforce-gtx-1050-ti-pascal-announce/

But, as usual, slides are made to show the product in the best light possible. Even if it is faster by 15%, that is barely enough to justify an upgrade from GTX950 users.

EDIT - If that is the case then it would offer 15% more performance than RX460 for also a 15% higher asking price.

 

[Deleted member 13524]

I imagine this chip will behave like all the other nvidia ones. Boost clock on the box says 1455, I expect you to see higher clocks in reality.

I sincerely doubt this is only as fast as a 950

Boost clock of 1455 is for the non-Ti version. The GTX 1050 Ti is 1300-1400MHz, whereas the GTX 950 was 1200-1300MHz.
"Typical boost" clocks are still unknown but yeah, we're probably seeing a 100MHz boost which should be good for 8-9% better performance.

The GP107 does have 10% (300M) more transistors than GM206, though one could wonder how many of those are dedicated to Pascal-specific improvements like async compute support.


EDIT: worthless argument is worthless.

 

[homerdog]

EDIT - If that is the case then it would offer 15% more performance than RX460 for also a 15% higher asking price.

That's been the case for a long time now. Hard to find any modern card that stands out in price/perf.

 

[DuckThor Evil]

There were GTX 950 cards without a 6-pin connector, at least from Evga, so these don't look too great compared to that. Perhaps TSMC really is quite a bit better, but also more expensive or has not enough production capacity to satisfy yet another line of cards at this time.

 

[Picao84]

Oh no. [exclamation points, dramatic yelling]
How am I ever going to live without the approval of an internet stranger over a totally irrelevant issue [several interrogation points, dramatic crying]

Hey, I'm not the one attacking people every time someone considers AMD is under delivering, which should also be a total irrelevant issue... and then running from a conversation with childish remarks.

 

[Razor1]

If we go by the slides shown here, it will be around 25% faster:
http://wccftech.com/nvidia-geforce-gtx-1050-ti-pascal-announce/

But, as usual, slides are made to show the product in the best light possible. Even if it is faster by 15%, that is barely enough to justify an upgrade from GTX950 users.

EDIT - If that is the case then it would offer 15% more performance than RX460 for also a 15% higher asking price.

evga has stated 26% faster that a gtx 950 (I'm assuming this is for the 1050ti), so I'm thinking that is what we will be seeing on average. That should put it above the gtx 960, which by all counts is the typical new gen update, one tier difference. Nothing special to see. Still a good card for people that don't want that extra power connector. Better than anything else out there right now.

http://videocardz.com/63898/evga-launches-ten-geforce-gtx-1050-graphics-cards

yep here ya go

The EVGA GeForce GTX 1050 Ti and 1050 may look small, but they offer big performance. In fact the EVGA GeForce GTX 1050 Ti is about 26% faster than a GTX 950, and the GeForce GTX 1050 is about 16% faster.

so what you are getting here is the stock gtx 1050 at the same price as the rx 460 with more performance (or equal performance to the overclocked models and cheaper). And the 1050ti in a bracket of its own, which is going to be uncontested till AMD fills in that gap somehow.

I don't see the 2gb's for the gtx 1050 holding it back at least on at what this card will be playing at certain resolutions and settings.

 

[Deleted member 13524]

There were GTX 950 cards without a 6-pin connector, at least from Evga, so these don't look too great compared to that. Perhaps TSMC really is quite a bit better, but also more expensive or has not enough production capacity to satisfy yet another line of cards at this time.

I would go as far as saying maybe 16FF TSMC 132mm^2 GP107 wasn't worth the effort/money/man-hours compared to the already developed 28nm TSMC 227mm^2 GM206, and Samsung's 14FF was the middle term.
And if it wasn't for the power-critical mobile designs, this could be have been a chance to just rebadge the GM206.

evga has stated 26% faster that a gtx 950, so I'm thinking that is what we will be seeing on average.

http://www.evga.com/articles/01063/evga-geforce-gtx-1050-ti-and-1050/

That comparison is probably between a vanilla GTX 950 and EVGA's 1050 Ti FTW model with 1400MHz base / 1500MHz boost clocks, and a 6-pin connector.
Divide that FTW's 1493MHz boost clocks by the GTX 950's "average boost" clocks of 1188MHz and you get 1.2567
Given how the number of execution units is pretty much the same all around, it's a good (albeit optimistic) estimate.

 

[Razor1]

good eye!

 

[Erinyes]

According to performance estimates, it looks like the GTX 1050 Ti will trade blows with the RX 460, with the typical "above in DX11, below in DX12/Vulkan" results we've seen as of late.

The RX460 currently trades blows with the GTX950. The1050Ti is basically the same but has ~17% higher clocks and ~27% more effective bandwidth so I expect it to beat GTX950 by about 20%. The RX460 and the GTX 1050 should be quite close though.

A 16CU RX465 4GB at $129 should be somewhere between the 1050 and 1050Ti and would complete the lineup quite nicely.

I don't think 14nm Samsung is really all that much worse. GTX 1060 is a 120W card. Now if you just look at SMs, yes you could (just) do 60% of it within 75W. But (unlike gm107/206) everything else seems to be 100% of GP106 - same amount of ROPs, same memory interface (ok maybe with a bit less max clock), even same number of GPCs/Rasterizers (if you believe anandtech's diagram). Which is also shown by the transistor count, which is 3/4 of GP106. So, if you'd take that number (that's quite simplified calculation, I know), you'd end up with 3/4 the power draw of 120W with the same clocks - IMHO there's no way they could have achieved these clocks on 16nm TSMC within 75W neither.

I think its clear that the Samsung process is a bit worse. If we look at GM107 vs GM206..the boost clocks were similar and the power consumption was very similar to the ratio of resources, i.e. ~60% of the SMs for ~60% of the power (75W/120W). With GP107 vs GP106 though, the same 60% ratio gives you ~20% lower boost clocks.

Now this is complicated by the fact that power dosen't really scale linearly as we see with GM206 vs GM204 or GP106 vs GP104, but it gives us somewhat of an idea of the difference.

At least 14nm Samsung is denser according to these numbers by about 10% compared to GP106...

A bit more than that actually since the smaller GPUs have lower density than the larger GPUs. However higher density does not necessarily mean lower per transistor costs (Though considering the performance I expect it is lower)

This leads us to the question of why nvidia chose to go with 14nm Samsung instead of 16nm TSMC.
Given TSMC's seemingly better performance, maybe the price per die area is substantially higher and nvidia wasn't willing to have their lower-margin product using a more expensive process.

Probably a combination of factors including cost per transistor and wafer availability.

This generation's low-end $110 GPUs are offering the performance of last-gen's mid-end ~$150/170.
How is 30% less money for same performance a truly terrible thing? You just have to acknowledge that priorities are very different in this market segment.

But those last gen mid end GPUs are roughly the same price or slightly higher today. So there isnt really significant increase in performance at this price point. However this is to be expected and its pretty much been the norm for low end GPUs for the last few generations (Barring GM107 maybe)

If we go by the slides shown here, it will be around 25% faster:
http://wccftech.com/nvidia-geforce-gtx-1050-ti-pascal-announce/

But, as usual, slides are made to show the product in the best light possible. Even if it is faster by 15%, that is barely enough to justify an upgrade from GTX950 users.

EDIT - If that is the case then it would offer 15% more performance than RX460 for also a 15% higher asking price.

True..it isnt a big upgrade from GTX950..but it isnt meant to be. This is targeted at users upgrading from GTX 650/750. 4K HEVC decode is a good feature to have though.

Boost clock of 1455 is for the non-Ti version. The GTX 1050 Ti is 1300-1400MHz, whereas the GTX 950 was 1200-1300MHz.
"Typical boost" clocks are still unknown but yeah, we're probably seeing a 100MHz boost which should be good for 8-9% better performance.

The GTX 950 was actually 1024-1188 mhz so we're looking at a 200 Mhz or ~17% increase in boost clock. Combined with the ~25% increase in effective bandwidth..should be good for about 20% better performance IMHO.

 

[Deleted member 2197]

MSI announced their 1050 lineup (8 products). Highest factory OC at 1442 base/ 1556 boost.
http://www.guru3d.com/news-story/msi-geforce-gtx-1050-gaming-series-announced.html