NVIDIA Readies Three GP104 "Pascal" Based SKUs for June 2016

[the54thvoid]

Only in America. To the rest of the world maximise is correct.

Beat me to it. I am British.

You're reading that wrong.
980Ti beats the Fury X at 1080, but Fury X scales better towards 4k. And that is using today's hardware. More powerful hardware will only make this difference more obvious.
Also, GDDR5X having more bandwidth, means 384 bit memory interfaces are no longer a must-have. That's money saved on the PCB.
But where I agree with you is that if GDDR5 is enough to (mostly) keep the 980Ti competitive, is highly unlikely mid-range Pascal cards will need GDDR5X. Then again, using HBM2, GDDR5X and GDDR5 at the same time can turn into a supply/inventory nightmare.

Didn't read it wrong and although you are correct in the bulk of your post and I do agree.... The reason Fiji is not that hot at 1080p has something to do with its ROP count (some sort of back end architecture issue). It means as the fps drops, the difficulty's faced by its ROP count are mitigated and therefore it appears to scale better at 4K, when in fact it 'retrogrades' performance as the resolution goes down and the back end is flooded with fps processing. I'm sure @HumanSmoke could explain in technical terms.

 

[Vayra86]

Perhaps instead of nitpicking bits and pieces to make yourself sound clever, maybe you should read everything? No? Thought not. You also love to make assumptions. Have fun duckling.

Wait what.

You started the very old argument of companies and maximizing profits. Any company that doesn't maximize profits is basically tearing down its own progress, so "Double - You - Tee - Eff'' ?? That's not bits and pieces, it is the core of your argument. Also, I want to point out that you started your first post in this topic with OFFtopic nonsense about some storage solution that will come out. If we are speaking of 'clever'...

Either way if you want to see how it goes with companies that do NOT maximize profits, have a good long look at the great AMD and how their market share has evolved over the past ten years. Have fun

The actual process of each new GPU release actually goes a bit different from your line of thinking:

- Current release/gen has a series of price points and performance levels
- Next gen gets announced for the first time, and great improvements are promised
- Many enthusiasts (read: you) jump on the marketing bandwagon and tout all these improvements as 100% factual and overestimate what it means for the next gen
- Next gen gets the first die shots and real information: we see the regular performance increase that we see every gen, which is just about 30% give or take.
- Pricing schemes are discussed. Your argument pops up, and relates it to the previous points.

Do you see where it goes wrong? We over assume what the next gen will bring, we overestimate its performance jump, and then we see that prices are going to be exactly similar like it happens every gen and the performance jump will be quite similar to every previous gen as well.

The bottom line: nothing really changes, but the SAME performance does become cheaper every new release.

Perspective.

 

[Parn]

Will need to see the performance figure and power efficiency before making my decision. Also if the first batch comes with only GDDR5 instead of GDDR5X, then I'll wait.

 

[Octavean]

Possibly 1080 1070 and 1060 Ti

That was my thinking exactly.

If that were the case and if a GTX 1060 performs equal to or better then a GTX 970 (while being cheaper at GTX 960 prices) I would seriously consider buying one. Especially so if there was full VR on Oculus and Vive,...

 

[Moofachuka]

I have a feeling it will turn out like this

• GP 104-400 GPU - Titan
• GP 104-200 GPU - 1080
• GP 104-150 GPU - 1070

1080 Ti TBD to counter AMD.

 

[Fx]

I'll be waiting for GP100 so I can switch over from dual to single GPU. I've had my fun

Hehe. I am most sure you have.

 

[rtwjunkie]

I have a feeling it will turn out like this

• GP 104-400 GPU - Titan
• GP 104-200 GPU - 1080
• GP 104-150 GPU - 1070

1080 Ti TBD to counter AMD.

Except, GP104 is a mid-tier chip, not top end. So it certainly won't be Titan. Otherwise your list looks good....just move the other two up a notch.

 

[bug]

Beat me to it. I am British.



Didn't read it wrong and although you are correct in the bulk of your post and I do agree.... The reason Fiji is not that hot at 1080p has something to do with its ROP count (some sort of back end architecture issue). It means as the fps drops, the difficulty's faced by its ROP count are mitigated and therefore it appears to scale better at 4K, when in fact it 'retrogrades' performance as the resolution goes down and the back end is flooded with fps processing. I'm sure @HumanSmoke could explain in technical terms.

I know what you mean, but AMD never admitted Fiji being ROP limited. In fact, the only statement I got from someone working at AMD is "we have seen no evidence of Fiji being ROP limited". Though all evidence seems to say otherwise.

 

[rruff]

If that were the case and if a GTX 1060 performs equal to or better then a GTX 970 (while being cheaper at GTX 960 prices) I would seriously consider buying one.

In that case I don't think your wallet will be in danger of losing weight...

 

[rruff]

There is no problem in making profits, however, there used to be limits; the guts get ripped out purposefully and then sold.

Blame AMD. Competition is what gives us the maximum performance and minimum prices. There is no reason for Nvidia sell anything cheaper than they have to. It ain't a charity organization.

 

[rruff]

980 and 970 replacements and a new mobile chip makes the most sense imo too.

More like 980, 970 (980m), and 960 (970m) replacements. Do they ever make chips for mobile exclusively?

 

[EarthDog]

Brutal triple post... please edit that up man...

 

[ppn]

"200" denotes 3/4 enabled chip. "150" could be 192 bit.

 

[efikkan]

I have a feeling it will turn out like this

GP 104-400 GPU - Titan
GP 104-200 GPU - 1080
GP 104-150 GPU - 1070

1080 Ti TBD to counter AMD.

I'm not claiming it's certain, but this surely sounds much more like:
GP104-400 GPU - GTX 980 replacement
GP104-200 GPU - GTX 970 replacement
GP104-150 GPU - GTX 980M or GTX 960 replacement.

GP104 will be the first chip to launch, together with GP106, but GP104 will not be a high-end chip (especially given it's size). GP102 will relese before the end of the year, probably as a Titan X replacement.

I guess GP100 becomes the 980ti and Titan replacements then?

No, wait for GP102. GP104 will probaly outperform today's high end but will position itself as the new (upper) mid-range.

GP104 will have to be a good performer or I'm skipping this one, even though when Maxwell released 104 I was really rooting for Pascal. It better bring GDDR5X, for one, or I'm instantly turned off. That memory looks promising, the yields are reportedly very good, so there will be some good OC potential to be had, something GDDR5 can't really offer anymore.

If you're waiting for a new high-end product then GP102 will be your bet.

It's not impossible that the GP104-400 could use GDDR5X, but that would mean the supply will be very limited until September. Shipments of GDDR5X has started, but the supply is very limited. But if it turns out that Polaris can't compete with GP104-400, then perhaps it's a smart gamble.

But I still don't get why you insist on faster memory. Which games do actually need more than 8 GHz GDDR5 on a 256-bit bus? (for a GP104 class GPU)

I'm waiting for the GP102, but I don't know if it will use HBM2 or GDDR5X. But either way the memory will be more than fast enough, I'm more curious about how many CUDA cores it will get.

You're reading that wrong.
980Ti beats the Fury X at 1080, but Fury X scales better towards 4k. And that is using today's hardware. More powerful hardware will only make this difference more obvious.

"Better scaling" doesn't really matter when GTX 980 Ti still beats it.
Keep in mind that the fastest GDDR5 now is 8 GHz, up from 7 GHz used in the GTX 900 series.

Also, GDDR5X having more bandwidth, means 384 bit memory interfaces are no longer a must-have. That's money saved on the PCB.

The money saved on the memory controller and interface is probably more significant than the PCB, which in turn could give room for more CUDA cores.

The reason Fiji is not that hot at 1080p has something to do with its ROP count (some sort of back end architecture issue). It means as the fps drops, the difficulty's faced by its ROP count are mitigated and therefore it appears to scale better at 4K, when in fact it 'retrogrades' performance as the resolution goes down and the back end is flooded with fps processing.

If you knew what the ROPs do then you'll see it works the other way around. Higher resolution, higher samples(AA), etc. increases the load on the ROP, so that should punsh high resolutions.

The problem with Fiji is that it's an outdated architecture with a lot of minor inefficiencies in the design. Just have a look at the "raw" computing power of AMD's chips vs Nvidia, and you'll see they have a problem tweaking their architectures. GCN competed OK with Kepler, but then Nvidia tweaked it and released Maxwell and now is about to release Pascal with a very significant performance increase, while 4th gen GCN is only going to be a "minor" change. (except for the shrink)

 

[Recus]

That's fine (and up to them) level of evilness near zero.

I wonder, though, whether we'll see performance of Maxwell cards suddenly drop in the same benchmark (recall minus 10 fps on 780Ti in reviews on this very site) once new gen is available. That stinks quite a bit.

And whether we'll see more "Project (we didn't get a penny from nVidia) Cars" kind of games, with nVizillaCripplingCompetitorAndPreviousGenWithObscureCodeWorksWondersForOurSales. That's pure evil.

Spoiler

AMD doesn't have money to buy developers.

Spoiler

http://www.hardwareluxx.de/index.ph...x-970-black-exoc-sniper-im-test.html?start=16

 

[Casecutter]

Except, GP104 is a mid-tier chip, not top end. So it certainly won't be Titan. Otherwise your list looks good....just move the other two up a notch.

Yea, some said the "full" GP104-400-A1 would be a 1080Ti. IDK I think this will replay the same as Kepler 680, 670 and 660Ti. Then GP100 will first give a Titan, and then the same 700 Series re-do and those could have GDDR5X. Heck I might go out on a limb and say the 1070 could be as a reference 4Gb of fast GDDR5 on 256-Bit, though will be all available!

 

[mcraygsx]

That's the consensus view I think. Like last time, the 980 matched the 780ti and we had to wait for 980ti for the performance part.
Tell you what though, if I get a whiff of Maxwell driver support slipping, I'm going red team.

Exactly the reason I switched over from 780Ti Lightening to 980 Ti. With each drive and new game it was lagging behind the rest. I hope nVIDIA will treat its Maxwell owners better this time around.

 

[xenocide]

If the hypothetical GTX 1080 performs near a GTX 980 Ti, I will definitely be buying one to replace my aging GTX 670.

 

[medi01]

AMD doesn't have money to buy developers.

Spoiler

http://www.hardwareluxx.de/index.ph...x-970-black-exoc-sniper-im-test.html?start=16

So, not getting the point problems we have. Let me elaborate, you need to accompany these pictures with:
1) Developers claiming they didn't get a penny from AMD, as "Project wedidnt'getapennyfromnvidia Cars" developers did
2) You need to slap someObscureCodeCripplingCompetitorAndPreviousGenWithObscureCodeWorksWondersForOurSales into the mix

And then I'll be generous and won't notice the major difference in ads presence in game.

 

[Vayra86]

I'm not claiming it's certain, but this surely sounds much more like:
GP104-400 GPU - GTX 980 replacement
GP104-200 GPU - GTX 970 replacement
GP104-150 GPU - GTX 980M or GTX 960 replacement.

GP104 will be the first chip to launch, together with GP106, but GP104 will not be a high-end chip (especially given it's size). GP102 will relese before the end of the year, probably as a Titan X replacement.


No, wait for GP102. GP104 will probaly outperform today's high end but will position itself as the new (upper) mid-range.


If you're waiting for a new high-end product then GP102 will be your bet.

It's not impossible that the GP104-400 could use GDDR5X, but that would mean the supply will be very limited until September. Shipments of GDDR5X has started, but the supply is very limited. But if it turns out that Polaris can't compete with GP104-400, then perhaps it's a smart gamble.

But I still don't get why you insist on faster memory. Which games do actually need more than 8 GHz GDDR5 on a 256-bit bus? (for a GP104 class GPU)

I'm waiting for the GP102, but I don't know if it will use HBM2 or GDDR5X. But either way the memory will be more than fast enough, I'm more curious about how many CUDA cores it will get.


"Better scaling" doesn't really matter when GTX 980 Ti still beats it.
Keep in mind that the fastest GDDR5 now is 8 GHz, up from 7 GHz used in the GTX 900 series.


The money saved on the memory controller and interface is probably more significant than the PCB, which in turn could give room for more CUDA cores.


If you knew what the ROPs do then you'll see it works the other way around. Higher resolution, higher samples(AA), etc. increases the load on the ROP, so that should punsh high resolutions.

The problem with Fiji is that it's an outdated architecture with a lot of minor inefficiencies in the design. Just have a look at the "raw" computing power of AMD's chips vs Nvidia, and you'll see they have a problem tweaking their architectures. GCN competed OK with Kepler, but then Nvidia tweaked it and released Maxwell and now is about to release Pascal with a very significant performance increase, while 4th gen GCN is only going to be a "minor" change. (except for the shrink)

8Ghz GDDR5 on 256 bit with delta compression, is not a huge leap in bandwidth from 7Ghz @ 256 bit with delta compression. Also if they push 8Ghz GDDR5, the OC headroom is limited. GTX 980 is not starving for bandwidth, but with a more powerful core (give or take +30%) I do think we will see a VRAM bottleneck left and right. That is why GDDR5X seems sensible, or at least preferable. Memory OC's are giving performance boosts already, since Kepler...

Also as far as I know there have not been 102 SKU's, only 104, 106, 110/100.

 

[EarthDog]

IMO, I doubt it will be a vRAM bottleneck, at least not at 2560x1440/1600 on down (where 97%+ of people play). Now, 4K, or 2560x1440 x3, I'm with you there. But people don't buy a midrange card to drive 4K monitors. If they went SLI, well the horsepower of the core(s) covers up a potential glass ceiling at those higher resolutions due to a bandwidth limitation.

Memory overclocks have always given performance boosts. The level of boost varies widely on the title, settings, and resolution though. The higher the res and settings, particularly AA, the more bandwidth that is typically needed.

 

[efikkan]

8Ghz GDDR5 on 256 bit with delta compression, is not a huge leap in bandwidth from 7Ghz @ 256 bit with delta compression.
<cut>
GTX 980 is not starving for bandwidth, but with a more powerful core (give or take +30%) I do think we will see a VRAM bottleneck left and right.

Assuming there are no improvements in the compression, this will still provide 14% extra memory bandwidth. And given that GTX 980 is not starving in terms of memory bandwidth, that can easily scale well with an estimated 30-35% performance increase. In resolutions like 1080-1440p the GPU will have no bandwidth problems even without GDDR5X.

Also if they push 8Ghz GDDR5, the OC headroom is limited.

I don't think that will matter at all. Most overclockers don't even care about overclocking the memory, it generally just gives more heat and a marginal increase in performance.

That is why GDDR5X seems sensible, or at least preferable.

GDDR5X is possible and would be nice, but not required for this class of hardware. If Nvidia choose GDDR5X the supply will be limited until September.

Memory OC's are giving performance boosts already, since Kepler...

Overclocking almost always gives some performance increase. In most balanced builds the GPU computing power will be the bottleneck, but still overclocking the CPU and even the GPU memory will give a few percent gain. This happens because during rendering of every frame the system is limited by different factors. ~80-90% of the time it's compute limited, but at times it can be CPU limited, GPU memory limited, cache limited, TMU limited etc. That's the reason why any overclock will give some gain, but generally the GPU clock is the "only" thing to care about.

 

[HumanSmoke]

If you're waiting for a new high-end product then GP102 will be your bet.

I think that will be the case. GP100 has 1,920 large dedicated FP64 units. If Nvidia repurpose a sizable percentage of these to add to the 3840 FP32 cores, the overall gaming performance will provided with a significant uplift.

It's not impossible that the GP104-400 could use GDDR5X, but that would mean the supply will be very limited until September. Shipments of GDDR5X has started, but the supply is very limited. But if it turns out that Polaris can't compete with GP104-400, then perhaps it's a smart gamble.

Memory IC's are small (10mm * 14mm). A 12" wafer produces 400-420 die candidates. Even a pilot line should be churning out a reasonable number of usable silicon. It also seems that GP104-400 will indeed be offered with GDDR5X if this leaked shot is any indication. I'm not a fan of the rumor/shill/clickbait sites but the picture has considerable detail.