The GT300/Fermi Thread

[Solus Corvus]

Except, if Fermi is indeed 40% faster than Cypress (which I don't believe is possible) GTX 360 won't be equal to 5870 but will be a lot faster than it, and GTX 380 will be extremely close to 5970, so price points being the same comparatively to the last gen (as in difference between GTX 275/285 and HD4800's), Nvidia will be in an über condition.

If you look at 40% faster then evergreen with 9.11 drivers then yeah, GTX380 would be pretty close to 5970. But 9.12 drivers are already much better for CF. GTX380 would be closer to 5970 then 5870 but it would still be in between. As for GTX360 beating 5870 by a great margin, I disagree. Not if Fermi is only 40% faster then evergreen.

[~CS~]

This a bit OT , but since there is so much discussion i thought i would add my speculation .If rumor of ATI switching to GF for GFX production is true and the new series will be delayed , they will release a card on same node 40nm just like they did with 3870 (55nm) to 4870 (55nm) , just a bigger chip after the process is butter smooth and yields are plentiful ,after all NV will be making a 500mm2 chip for a while and tcms will have some experience. So add say another 480 Sp's ( 800 would be nice ) some 6 Gbps GDDR and you got nice new card .

On topic .The new cards from Nvidia are nice but i keep wondering , how many cards can they make with one chip ? 3 Tops , if they are to keep 260 (if still under same name,lol ) positioned against 5770 they are going to lose a lot of mainstream market , maybe a bit less performance but overall its a better card .

[annihilat0r]

everything aside, you guys checking the Nvidia Facebook page? 32xAA :S Another good way to kill FPS without making anything look better!

NVIDIA GeForce GF100 Fermi Video Card Facts
GF100 is the codename for the first GeForce GPU based on the Fermi architecture!
The GF100 board is 10.5-inches long -- the same length as GeForce GTX 200 Series graphics cards!
GF100 packs in over 3B (billion!) transistors
The GF100 supports full hardware decode on the GPU for 3D Blu-Ray
GF100 graphics cards will provide hardware support for GPU overvoltaging for extreme overclocking!
GF100 supports a brand new 32x anti-aliasing mode for ultra high-quality gaming!

[Iconyu]

everything aside, you guys checking the Nvidia Facebook page? 32xAA :S Another good way to kill FPS without making anything look better!

NVIDIA GeForce GF100 Fermi Video Card Facts
GF100 is the codename for the first GeForce GPU based on the Fermi architecture!
The GF100 board is 10.5-inches long -- the same length as GeForce GTX 200 Series graphics cards!
GF100 packs in over 3B (billion!) transistors
The GF100 supports full hardware decode on the GPU for 3D Blu-Ray
GF100 graphics cards will provide hardware support for GPU overvoltaging for extreme overclocking!
GF100 supports a brand new 32x anti-aliasing mode for ultra high-quality gaming!

Why would someone want to throw more volts at something already using 225watts at stock?

[annihilat0r]

225w is for Tesla, with 3/6GB of memory. It should be lower for GeForce. Plus, with your line of thinking no one should overclock a GTX 295 or HD5970.

[zalbard]

225w is for Tesla, with 3/6GB of memory. It should be lower for GeForce.

But it might have higher clocks so you never know...

[ajaidev]

But it might have higher clocks so you never know...

That is what i am counting on higher clocks higher performance..

[Iconyu]

225w is for Tesla, with 3/6GB of memory. It should be lower for GeForce. Plus, with your line of thinking no one should overclock a GTX 295 or HD5970.

How much less? I doubt they'll release with less than one 1GB, so we're looking at 1.5GB on a single chip card that will probably still chuck out 200+watts when it's really loaded.

Maybe it won't be that bad, it's possible that number came before the switch to 448 shaders (which they might get back now 40nm yields are better). But I personally wouldn't overclock a dual gpu card unless I was benchmarking. Too many hot parts in too small a area for my liking.

Unless it can manage that magic 40% faster than a 5870, nVidia will be in trouble as they can't make a dual GPU with two of those chips.

[annihilat0r]

Yeah, I don't doubt that GTX 380 will have a TDP of higher than 200. But 5870's TDP is very close to 200 too, and you can bet 5890 will be above 200W when it's released.

Still don't get how a GTX 395 is going to be made though

[trinibwoy]

What would be the point of discussing it instead of sarcastically beating up strawmen?

What strawmen? You're the one making sweeping assumptions. I just chose not to counter them with my own unfounded predictions. You never explained why you think there are big driver improvements to be had on Cypress even though the architecture barely changed.

I'm not even sure why you brought up GTX285 SLI, everyone else was talking about GTX295.

That's because I didn't. Mario did and you did as well.

If that happens performance would probably, IMO, look something like: 5850 < 5870 ~= GTX360 < GTX380 < 5970 < GTX390
That's basically a redux of last generation.

Except the GTX280 wasn't 40% faster than the HD4870 and was quite a bit slower than the 4870X2.

[Tim]

everything aside, you guys checking the Nvidia Facebook page? 32xAA :S Another good way to kill FPS without making anything look better!

I don't understand your reasoning. 32xAA as a feature is pretty cool. Even if it's not possible on Crysis. Older games will fly. 32xAA is still not realistic AA. Anything and everything can be improved in my opinion. I welcome new Antialiasing modes.

DX11 is nice, but it has a reasonably big effect on fps in Dirt2. Same story I guess, it just means more eyecandy, worse performance.

I'm sure Fermi will be worth the wait.

[JohnZS]

Ok people, Quadro NEWS. There will be 2 SKU's on launch. (Early Q2 2010) and then a mega expensive donate your organs and house SKU in Q3 2010, it would not surprise me if we see CUDA cores "cut" in groups of 32, for example 448, 480 and 512 although this is speculation, the only confirmation is that there will be 2 Quadro FX Series GPU's in Q2 2010 and "the big daddy" in Q3 2010.
If (and yes this is a big if) the same is reflected at consumer level my guess would be.

360 = 448 pipes 1.5GB of RAM
380 GTX = 480 pipes 3GB of RAM
380 ULTRA = 512pipes and 6GB of RAM and an expensive electricity bil and large carbon footprint.

Pure speculation at this stage.. all eyes on the 7th of January I guess eh?
John

[ajaidev]

everything aside, you guys checking the Nvidia Facebook page? 32xAA :S Another good way to kill FPS without making anything look better!

Rather would like to use 8x SSAA with old games than 32x CSAA i dont think that 32x will be for MSAA, CSAA is totally lame compared to MSAA....

Ok people, Quadro NEWS. There will be 2 SKU's on launch. (Early Q2 2010) and then a mega expensive donate your organs and house SKU in Q3 2010, it would not surprise me if we see CUDA cores "cut" in groups of 32, for example 448, 480 and 512 although this is speculation, the only confirmation is that there will be 2 Quadro FX Series GPU's in Q2 2010 and "the big daddy" in Q3 2010.
If (and yes this is a big if) the same is reflected at consumer level my guess would be.

360 = 448 pipes 1.5GB of RAM
380 GTX = 480 pipes 3GB of RAM

380 ULTRA = 512pipes and 6GB of RAM and an expensive electricity bil and large carbon footprint.

Pure speculation at this stage.. all eyes on the 7th of January I guess eh?
John

The 6GB's or ram costs a lot and i mean a lot, if this is a consumer card nvidia may have to sell it at a loss also i do believe 6 GB's of video ram for games is more than overkill its mega overkill. Hell i prefer 6GB's of ram on a i9 than a fermi.

[Heinz68]

XbitLab claims that the GF100 will have 512 stream processors, I wonder from where they're getting the info.

The flagship Nvidia Fermi “GF100” graphics processor will feature 512 stream processing engines (which are organized as 16 streaming multi-processors with 32 cores in each) that support a type of multi-threading technology to maximize utilization of cores and deliver extreme computing power. Each stream processor has a fully pipelined integer arithmetic logic unit (ALU) and floating point unit (FPU). The top-of-the-range chip contains 3 billion of transistors, features 384-bit memory GDDR5 memory controller with ECC and features rather unprecedented (for GPUs) 768KB unified level-two cache as well as rather complex cache hierarchy in general. Naturally, the Fermi family is compatible with DirectX 11, OpenGL 3.x and OpenCL 1.x application programming interfaces (APIs). The new chips will be made using 40nm process technology at TSMC.

[cowie]

360 = 448 pipes 1.5GB of RAM
380 GTX = 480 pipes 3GB of RAM
380 ULTRA = 512pipes and 6GB of RAM and an expensive electricity bil and large carbon footprint.

Pure speculation at this stage.. all eyes on the 7th of January I guess eh?
John

thats ok its all spec at this point anyways


380 ULTRA = 512pipes and 6GB of RAM and an expensive electricity bil and large carbon footprint.< i want this one

[eric66]

wth consumer card needs 6 gb i doubt that it will have 6. 2 gb max imho

[hennyo]

Well, I really don't understand what this whole ram speculation numbers thing going on right now as I thought we got that information awhile ago. If I remember correctly, 1.5GB was going to the standard amount of ram on the top consumer card with a possibility of a 3GB part in the future kinda how the ATI 5000 series is with 1GB standard and a 2GB possibility part.

[RPGWiZaRD]

1.5 - 2GB for vram is the only capacity that makes sense IMO for non-quadro cards. Personally I see 1.8GB and 1.5GB respectively for GTX 380 and 360 would be the best compromise (cost, size, a good balance between what's actually usable now/a bit of headroom for future games) if it's technically feasible, which I can't be bothered to check. I hope some numbers shows up soon, these speculations and rumors are really boring me to death now lol.

[Mk]

1.5 - 2GB for vram is the only capacity that makes sense IMO for non-quadro cards. Personally I see 1.8GB and 1.5GB respectively for GTX 380 and 360 would be the best compromise (cost, size, a good balance between what's actually usable now/a bit of headroom for future games) if it's technically feasible, which I can't be bothered to check. I hope some numbers shows up soon, these speculations and rumors are really boring me to death now lol.

2 more weeks before any official word assuming Nvidia gonna talk about GF 100 at CES

[ajaidev]

IMO Geforce will have 1536MB "128*12" and for 3072MB "256*12".

If you want 2GB ram you need to use 256*8 so that means 4 slots are empty. Companies generally try and fill as many slots as they can since its cheaper than putting higher capacity memory chips.

[highoctane]

Why would someone want to throw more volts at something already using 225watts at stock?

Just think about the name of the forum you are posting to, look at some people sigs, if anything we need to be buying carbon credits around here...

[Solus Corvus]

What strawmen? You're the one making sweeping assumptions. I just chose not to counter them with my own unfounded predictions. You never explained why you think there are big driver improvements to be had on Cypress even though the architecture barely changed.

32nm is a strawman. I said nothing of the sort. As for my assumptions, I wouldn't say they are unfounded. They are founded on the little bits of information that we have at the moment. Certainly nothing strong like "I have an NDA", but it's still fun to speculate. You don't have to participate if you don't want to.

As for driver improvements, it's just a suspicion of mine. That a chip with the same number of shaders, higher clocks, and higher bandwidth would consistently score lower then a 4870x2 (much less 4890CF) even though it doesn't have to interface through CF doesn't seem right. You keep mentioning a barely changed architecture - then where does the performance discrepancy come from?

That's because I didn't. Mario did and you did as well.

A quick review shows that Marios did mention it, but nobody picked up on it. I certainly didn't mention GTX285 SLI until after you did. Double the speed of GTX285 =/= GTX285 SLI. All the other discussion revolved around GTX295.

Except the GTX280 wasn't 40% faster than the HD4870 and was quite a bit slower than the 4870X2.

Like I said, that's my optimistic (for NV) view. If you don't like it I could present my doom and gloom view.

Personally I think you guys are the ones being more pessimistic (or sandbagging). I think Fermi will be significantly faster then GTX295. If it isn't it'd be a disaster, imo.

[Tim]

XbitLab claims that the GF100 will have 512 stream processors, I wonder from where they're getting the info.

Wow, sounds pretty good.

[Chumbucket843]

As for driver improvements, it's just a suspicion of mine. That a chip with the same number of shaders, higher clocks, and higher bandwidth would consistently score lower then a 4870x2 (much less 4890CF) even though it doesn't have to interface through CF doesn't seem right. You keep mentioning a barely changed architecture - then where does the performance discrepancy come from?

4870x2 can set up 2 triangles per clock. 5870 can set up 1 triangle per clock. you can see in synthetics that geometry shader is the same.

[Solus Corvus]

Does that account for the majority of the performance difference?