[guru3d rumor] GTX 880- 8GB-3200 SP's

[Mondozei]

This is basically the same old rumor as before. GM204 is small Maxwell, too.

 

[OCGuy]

256-bit memory for a successor to the 780? Step backwards?

Successor to the 780 (GM 210) won't be released as a Geforce initially. The GTX680 ended up being the GTX770.

As far as pricing, that will depend on the market at the time, but they certainly can command a higher price if the relative performance/efficiency/OC ability between GK and GM stays near what we got with the 750/750ti and its predecessors

 

[TeknoBug]

lol 8GB on a 256-bit memory pipe? I'd like to see how that scales out.

The price is what has me worried, the 780 and 780Ti were expensive but not THAT expensive, I wouldn't be surprised if Nvidia trolls us with $1000 price tag.

 

[Face2Face]

As far as pricing, that will depend on the market at the time, but they certainly can command a higher price if the relative performance/efficiency/OC ability between GK and GM stays near what we got with the 750/750ti and its predecessors

I'll take a 3200 core part that overclocks to 1400MHz :thumbsup:

 

[Lepton87]

the ignorance surrounding memory interface needs to die already

remember the GTX280? It has a 512bit memory interface. GTX480/580 had 384 bit interface. GTX680 had a 256bit...

just because we see a decrease in memory interface doesn't mean we're taking a step back, as that's only one pieces to what determines the memory bandwidth

the 512bit GTX280 had GDDR3, so that hurt it, going to GDDR5 allowed the 480/580 surpass that memory bandwidth despite 384bit. Going to much faster GDDR5 allowed the 680 to match the 580's memory bandwidth despite 256bit, and improvements in efficiency meant it could do more with that same raw bandwidth.

Granted, I will admit, going back to 256bit might be a step backwards in performance this time around if rumor is true, but we can't really make any snap judgements until we see a more complete picture of the design, and even then we can't really make and true judgements until we see actual hardware in action

All of that doesn't matter, in the past when they narrowed the bus they increased the clocks. They didn't have to sacrifice memory bandwidth but with 256bit and GDDR5 it's inevitable. How much clocking headroom do they have now? GDDR5 is pretty much maxed out on NV cards, it was supposed to go to 7GHz and 780ti is already at 7GHz. What are the fastest GDDR5 chips produced? Even if they go beyond 7GHz it will still be a substantial decrease in bandwidth.

 

[thilanliyan]

This is just midrange guys...just wait for the REAL high end...

 

[Dribble]

This is just midrange guys...just wait for the REAL high end...

That's a lot of transistors - at 28nm this will be the high end.

 

[Keysplayr]

This sounds like another GTX680 (GK104) midrange product. Just like the 6xx and 7xx series, GM104 will kick off the 8xx series and GM110 will arrive on 20nm somtime next year or by years end. GM104 will get promoted to GTX970 just as GK104 got promoted to GTX770. Anyways, thats what I think will go down.

 

[exar333]

This makes we want to wait for 20nm even more. :/

 

[x3sphere]

If it has an HDMI 2.0 port this will be a buy for me, as I'd really like to game at 60 Hz on a 4K TV.

 

[OCGuy]

8GB VRAM + Mid-range doesn't make much sense. Why would anyone want to waste money paying for useless memory?

 

[UaVaj]

memory bandwidth is more that just the memory interface. need to account for memory speed too. some will never learn.

as for rop. 32 is definitley not enough. especially when 4k is on the horizon.

the spec is all wrong.

 

[Vesku]

Doesn't seem like too much more can be squeezed out of GDDR5 in terms of MHz, so Nvidia would have to have some new tricks to make 8GB on a 256 bit bus competitive. I suppose there is the possibility of GDDR6 but this particular rumor has GDDR5 in the description and last I read no one was sure of GDDR6's development status.

 

[Ajay]

That's a lot of transistors - at 28nm this will be the high end.

If true then yes. GK110 had 7.1B on a huge die. Even if this new node is able to cram in a few % more xtors/mm2 - it will still be a big die and hence is a high-end GPU (where 256b mem interface makes no sense - NV would have had to increase mem efficiency by what - 30% for this to make sense?!).

So...the numbers don't add up :thumbsdown:
Doesn't mean there is no truth to the rumor, but it can't all be right; unless NV is just counting on the vast majority of buyers being clueless virtual ATMs dedicated to enhancing NV's profits.

 

[bunnyfubbles]

All of that doesn't matter, in the past when they narrowed the bus they increased the clocks. They didn't have to sacrifice memory bandwidth but with 256bit and GDDR5 it's inevitable. How much clocking headroom do they have now? GDDR5 is pretty much maxed out on NV cards, it was supposed to go to 7GHz and 780ti is already at 7GHz. What are the fastest GDDR5 chips produced? Even if they go beyond 7GHz it will still be a substantial decrease in bandwidth.

except it does matter, because people are judging the product on the wrong specs. ie don't get upset at seeing 256bit, get upset at the 238GB/sec bandwidth (that's a big step back)

there's going to be a time in the not too distant future (2014 is supposed to see coming out party for new stacked memory chips, such as HMC/HBM that will offer a drastic leap in capacity and bandwidth) where we see Volta be nVidia's introduction to stacked memory and 1+TB/sec of memory bandwidth, and that will likely be achievable on no more than 4 memory channels...should we be upset by that when that will likely more than double the memory bandwidth we see with Maxwell?

 

[OCGuy]

I'm not sure I would get upset over theoretical maximum memory bandwidth if the performance is there.

That is like getting upset at the shroud art.

 

[Vesku]

Well, if this rumor is 1. real and 2. not obfuscated (big ifs there) then 8GB of GDDR5 on a 256 bit bus means you are most likely paying for extra RAM slapped on mainly for marketing reasons. With current 4 Gb GDDR5 chips that is 16 in 16 bit mode rather than 8 in 32 bit mode. So can be a little upset if there is no cheaper 4GB option (none on manifest image) even if they've found some tricks to make the memory bus itself less of a handicap.

 

[tviceman]

I'm not sure I would get upset over theoretical maximum memory bandwidth if the performance is there.

That is like getting upset at the shroud art.

Agreed. Nvidia has always capitalized on bandwidth in relation to performance when moving to new architectures. GTX 680 has the same bandwidth as the GTX 580, but is about 25-30% faster. Likewise, the GTX 660 has less bandwidth than the 560 TI 448, but handily outperforms it. GM107 does extremely well given it's available bandwidth.

The big concern I have is with 32 ROP's. I think it's safe to assume clock speeds aren't finalized yet but likely aren't going to be anything astronomical (probably no more than 1150mhz boost). So unless the ROP's have been significantly upgraded from Kepler, then we're looking at a very unbalanced part. Way more shading power with about the same pixel throughput as the GTX 770. That lends to the exact same "issues" people have when scaling current Nvidia graphics up to ultra high resolutions or applying large amounts of MSAA / SSAA.

I'd like to think that the reason GM104 taped out AFTER GM107 was released to retail is because Nvidia engineers were still working on different aspects of Maxwell's odds and ends.

 

[Haserath]

memory bandwidth is more that just the memory interface. need to account for memory speed too. some will never learn.

as for rop. 32 is definitley not enough. especially when 4k is on the horizon.

the spec is all wrong.

Yes, precisely, the speed needs to be counted too. It's not fast enough.()

But anyway, I think Maxwell uses on die caching a lot more than Kepler, so the bandwidth shouldn't be as big of a deal. Much of their performance improvement seemed to be from memory pipeline enhancements.

I like how you go from memory to ROPs but your entire argument changed... Just saying...:\ I think the caching helps the ROPs out too.

I expect it to be the same distance as 580->680. But 4GB, 8 would be a waste.

 

[moonbogg]

Just release the thing already. If the card was 20% faster than 780ti, had 4gb of ram and cost about $500, that would make a lot of people happy enough, including me. I'm just ready to scratch the itch but won't fork over the $700 i'm thinking they might ask for it. Its ok if they do, prices will fall soon enough or 870 will release at better price/perf.

 

[Majcric]

Just release the thing already. If the card was 20% faster than 780ti, had 4gb of ram and cost about $500, that would make a lot of people happy enough, including me. I'm just ready to scratch the itch but won't fork over the $700 i'm thinking they might ask for it. Its ok if they do, prices will fall soon enough or 870 will release at better price/perf.

20% percent faster than ti and I will buy, even if it is so-called midrange.

 

[Silverforce11]

There's alot leaks within the the Taiwanese and Chinese tech forums all pointing to the same specs as these, so I wouldn't be shocked if it was real.

On paper, it would be a killer 1080p to 1600p card, probably 30-40% above 780ti performance.

Also, it's probably likely at this point given all the other rumors, that 20nm TSMC is going to be skipped altogether by AMD/NV. So these 28nm next-gen stuff will carry the flag until 16nm.

 

[3DVagabond]

Well, if this rumor is 1. real and 2. not obfuscated (big ifs there) then 8GB of GDDR5 on a 256 bit bus means you are most likely paying for extra RAM slapped on mainly for marketing reasons. With current 4 Gb GDDR5 chips that is 16 in 16 bit mode rather than 8 in 32 bit mode. So can be a little upset if there is no cheaper 4GB option (none on manifest image) even if they've found some tricks to make the memory bus itself less of a handicap.

It's not unusual for ES models to have more ram than the actual reference will. It would be pretty embarrassing to have the AIB's want to do an 8gig model and have it not work. This doesn't mean the reference model wouldn't be 4gig. It's not like nVidia to go overkill on the amount of ram on their cards. All they care about is that there's enough ram to hold the card until it's replaced by something new.

 

[Keysplayr]

Fun with extrapolation:

 

[omeds]

32 ROP's and 256bit.. son I am disappoint. Numbers speak louder than words though, might be another 680 that by all accounts should have been destroyed by Tahiti based on specs, but never happened.