Nvidia GF110?

[Ravynmagi]

I believe Nvidia could be making a GF100b with all 512 cores unlocked and launched by December.

But I'm having a hard time believe Nvidia could make a new GF110 GPU with 512 cores, 128 TMUs, and 512 bit memory interface and have that ready by December. This just says too over the top to be real.

 

[RussianSensation]

I believe Nvidia could be making a GF100b with all 512 cores unlocked and launched by December.

"In the immediate future, there are prospects of a more realistic-sounding GF100b"

Should we use Valve Time or NV time or AMD or Intel time to define Immediate Future? :biggrin:

WTH is "immediate future"? To me that's like in 2 min from now. To JHH it might be 6 months.

 

[Lonyo]

I believe Nvidia could be making a GF100b with all 512 cores unlocked and launched by December.

But I'm having a hard time believe Nvidia could make a new GF110 GPU with 512 cores, 128 TMUs, and 512 bit memory interface and have that ready by December. This just says too over the top to be real.

I thought (although it's purely from memory) that while the GTX4xx cards scaled very well with cores, they didn't benefit all that much from increased memory clocks, so a 512-bit bus would seem unnecessary, especially when they aren't using particularly fast RAM (relative to what is now available). They could just up the RAM speed rather than making a 512-bit memory controller (since either chance would require a decent amount of work).

 

[RussianSensation]

They could just up the RAM speed rather than making a 512-bit memory controller (since either chance would require a decent amount of work).

No they can't. GF100 chip doesn't work with fast RAM speeds. It was documented that the controller becomes unstable with faster GDDR5 chips. That's why they can't just add faster GDDR5 chips without a GF100 tweak/redesign.

 

[MutantGith]

No they can't. GF100 chip doesn't work with fast RAM speeds. It was documented that the controller becomes unstable with faster GDDR5 chips. That's why they can't just add faster GDDR5 chips without a GF100 tweak/redesign.

As I recall, they had mentioned specifically that they (NV engineers) had had issues with incorporating the DDR5 memory controller with the efficiency that they had thought they would be able to acheive. As I recall, it was one of the things they had mentioned in relation to the cascading issues that they'd had trying to hop so many architecture milestones in one product iteration, ie 40 nm, new shader/compute model, GDDR3-5. Definitely one of the reasons that fill rate and memory clock were somewhat limited with the first run of GF100 parts, but also a potential reason that a mid-cycle retuning (whatever they/we decide to name it) could allow for more breathing room. It's likely that more knowledge of the process, as well as correlating the feedback they've gotten from the foundry about variations from their modeling would allow for memory scaling to improve.

The return on disabling some of the compute performance features in terms of TDP and therefore clock scaling was impressive from 465 -> 460. It's possible that in addition to whatever optimizations they may have been able to impart due to process understanding in this theoretical part, increasing the efficiency of the memory controller would help increase the part's overall performance, and limit some of the scaling flattening that happens at higher resolutions. Might be especially important, depending on how they want to sell the 3D/Surround performance vs. Eyefinity.

 

[Silverforce11]

As for the unlocked gf100, with the extra shaders, there's only a minor perf increase. They'll need to OC it a fair amount to match Cayman, but then it'll be >300W.

Just make a dual-gtx460 card, with lowered clocks to meet official TDP, but 2x8pin for enthusiasts to OC past the 300W limit, price it at $449 and call it done. I mean if they can sell gxt460s at $229 msrp, putting two together shouldn't be a problem.

 

[Wag]

So does this mean the end of a dual-GPU GF104 based card? Is Nvidia going to do a dual-GPU GF110 card instead or is just skipping the whole dual-GPU route entirely?

 

[Skurge]

http://www.techpowerup.com/132832/NVIDIA-to-Counter-Radeon-HD-6970-Cayman-with-GeForce-GTX-580.html

"It's back to its old tactics of talking about GPUs that haven't even taken shape, to try and water down AMD's launch."

Pretty much sums it up.

 

[Sp12]

And here I was thinking GF110 would follow the naming scheme of every other Fermi chip out so far and be an extreme low-end variant.

 

[netxzero64]

You realize your 875 overclocked GTX460 consumes MORE power than a GTX470 (GF100), and you are only 30-40 Watts away from a GTX480??? :whiste: I do agree that they need to improve performance/watt so that they can make faster chips.

As far as this rumour goes, it's too hard to digest it right now. They could slap 2 x GF104s together to come up with a 512-bit card. Or the rumour can be talking about a new larger GF104 chip. The "renaming" of GF1xx to GTX580.....

yeah I know about my overclocked GF104 and i will only try it for a week ^^, coil noise is just what's bothering me...

a respin could be good if their 40nm process has matured, and when you mean a larger chip it means a more expensive GPU... well, talk about price per performance phew

 

[GodisanAtheist]

And here I was thinking GF110 would follow the naming scheme of every other Fermi chip out so far and be an extreme low-end variant.

Yeah no kidding I skipped over this topic more than a few times thinking "K so we're looking at something lower end than a GT430?"

Very little about the speculation fueling this topic makes much sense. Not even the courtesy of a poorly thrown together power point slide with NVIDIA CONFIDENTIAL watermarked all over it.

I can see Nvidia expanding on its superscalar GF104 arch and making an asskicking machine out of it, though. Imagine a 768sp Monster, GF104 doubled up...