Nvidia's Future GTX 580 Graphics Card Gets Pictured (Rumours)

[exar333]

A purported specifications table for NVIDIA's upcoming GeForce GTX 480 has leaked, courtesy IT168.com. The GeForce GTX 580 features 512 SP, mostly likely arranged in 16 SMs of 32 SP each. This marks a 7% increase in shader count over the GTX 480. The core clock is 772 MHz, with shaders clocked at 1544 MHz, a 10% increase over the GTX 480's clock speed. 1.5GB GDDR5 memory is clocked at 1 GHz (4 GHz effective) over the same 384-bit interface, for a bandwidth of 192.4 GB/s, a 8.5% increase. In essence, GF110 does look like "GF100 done right", with decent boosts in clock speed to boot, and manufacturing optimizations likely. The net result is an average performance jump of 15%-20% over the GTX 480. The GTX 580 is reported to score an immense X12700 in 3D Mark Vantage, which represents a ~30% boost. However, gaming improvements is expected to stay in the 15% to 20% region.

The GeForce GTX 580 is expected to release in November 2010, around the same time as the release of its main competitor, AMD's Radeon HD 6970. A GeForce GTX 570 is also expected, with roughly the same performance as GTX 480, but with much better thermal characteristics. The extent of retail availability is yet unknown, though rumours suggest very limited to no availability on release, much like the GeForce GTX 470/480 release.


http://vr-zone.com/articles/nvidia-geforce-gtx-580-specifications-leaked/10184.html

Nice post. I will have to decide if the GTX580 or the new 6970 will replace my existing 5870. Both look tempting...I can't wait for more news on prices and performance.

 

[SlowSpyder]

If they really are launching right around the time Cayman comes out, then that is great news. Hello competition, hello price wars. :awe:

The leak a few posts above your post says limited to no avaliablity. So, it might be more of a product announcement to try and steal a little of Cayman's thunder. We'll have to wait and see how it plays out, though. It would be nice if there was produtc, a price war right at launch of these parts would be nice.

So, if this turns out to be a revised chip, the GF110, not a GF100, does that mean Charlie was right about his use of 'unmanufacturable'?

 

[Martimus]

Top screenshot is comparison of fan with and without cover. Images have been overlaid and scaled for best fit. Fins are estimated to be double that of visible opening.

Bottom screenshot is comparison of fan between 480 and 580. Scaled according to PCI-E connectors (red lines for verification). Fan opening on 580 is clearly larger, by about 20%.

On far-right, comparison of fans between both cards. Red outline on 480 fan is best-estimate of maximum fin size. Assuming doubling of fin width, 580 is roughly 30% larger based on fin size. Assuming conservative value, where the same amount of fin width is hidden under the cover (blue outline), 580 fan is still roughly 17% larger based on that. Assuming no fins are hidden under cover, 580 is still roughly 6% larger than 480.

If you look at the bottom picture, you can see the entire fan through the opening of the cover, and it does look like the fans are not only the same size, but they look identical. The 580 has a bigger hole though.

 

[n0x1ous]

The leak a few posts above your post says limited to no avaliablity. So, it might be more of a product announcement to try and steal a little of Cayman's thunder. We'll have to wait and see how it plays out, though. It would be nice if there was produtc, a price war right at launch of these parts would be nice.

So, if this turns out to be a revised chip, the GF110, not a GF100, does that mean Charlie was right about his use of 'unmanufacturable'?

It might just be a respin like GF100-A4 or something....

 

[Martimus]

It might just be a respin like GF100-A4 or something....

Enough time has passed that it could even be B0 now.

 

[Idontcare]

The leak a few posts above your post says limited to no avaliablity. So, it might be more of a product announcement to try and steal a little of Cayman's thunder. We'll have to wait and see how it plays out, though. It would be nice if there was produtc, a price war right at launch of these parts would be nice.

So, if this turns out to be a revised chip, the GF110, not a GF100, does that mean Charlie was right about his use of 'unmanufacturable'?

Technically speaking, since the GF100 was infact manufactured (as evidenced by the fact you can buy them) it was/is very much a manufacturable IC.

Or does he refer to the GF100 as unmanufacturable because they disable some of the cuda cores? IBM does this with their CELL processor, across many nodes, and I've never heard anyone in the professional world refer to IBM's Cell as being an unmanufacturable IC.

When someone tags an IC with the adjective "unmanufacturable" it usually means the device simply fails to work, is broken. Early steppings are usually just that because they are broken in a way that renders them unmanufacturable.

I believe JHH went on record talking about a part of the initial Fermi design that was broken and as such did render Fermi unmanufacturable (something about its IMC iirc)...but they obviously tweaked the design enough to carry it across that threshold of being unmanufacturable to being manufacturable since they do manufacture and sell them now.

I understand it is just semantics, but if one wants to use an adjective like unmanufacturable with the intent that it carries the same weight as it does when used to describe other IC's that were/are unmanufacturable then the context and the semantics do need to be correct. Otherwise it is just misleading, and what is the point of being misleading?

I've worked on some unmanufacturable IC's...people make entire careers out of resolving the bewildering array of issues that can make an IC unmanufacturable. If Fermi is unmanufacturable then it is probably the first mass-produced mass-marketed and mass-purchased unmanufacturable IC in the history of the industry, the first I've encountered anyway.

 

[Scali]

Enough time has passed that it could even be B0 now.

The rumours seem to indicate that the chip was redesigned somewhat. Different balance of texture-units/ROPs... that sort of thing.
We can only hope that this is true. That this is some sort of fine-tuned Fermi++. The big brother of the GF104?
That they did something like they did with G80->G92->GT200:
G80 was the initial DX10 architecture... then G92 trimmed the fat, optimized the lot, and managed to do more with less. Then with GT200 they took the G92 and scaled it up again to G80-like dimensions.
Obvious difference would be that this time they've done GF100->GF104->GF110 without any dieshrinks, so they could only fine-tune the microarchitecture.

 

[n0x1ous]

Enough time has passed that it could even be B0 now.

Sure but I thought they didn't use a B stepping unless it was a die shrink. Like orginal GT200 on 65nm was GT200 -Ax and GT200b on 55nm was GT200-Bx

I could be wrong though.

 

[Soleron]

Technically speaking, since the GF100 was infact manufactured (as evidenced by the fact you can buy them) it was/is very much a manufacturable IC.

He says that other than the risk wafers none were ever produced. That all retail GF100s are from that initial run.

It's unmanufacturable because Nvidia have chosen not to manufacture it beyond what they'd committed to before seeing the outcome. I think that's a valid sense of the word.

 

[SlowSpyder]

He says that other than the risk wafers none were ever produced. That all retail GF100s are from that initial run.

It's unmanufacturable because Nvidia have chosen not to manufacture it beyond what they'd committed to before seeing the outcome. I think that's a valid sense of the word.

I have seen him post in his forums (I read the discussion on his articles, I am not a member over there) when challenged on his 'unmanufacturable' claim. He posts something along the lines of, "Where can I buy a 512SP GF100." So, I don't know if he's changing his story to that particular part is unmanufacturable, or it's a semantics issue (even if 'unmanufacturable' as he is implying it is only that to him and no one else in the industry).

 

[dangerman1337]

Woudln't it be better if the GTX 580 was mostly the GF104 but scaled up 50% than trying to adjust it to GF100 levels, since this is likely based on the GF110?

 

[gorobei]

Sure but I thought they didn't use a B stepping unless it was a die shrink. Like orginal GT200 on 65nm was GT200 -Ax and GT200b on 55nm was GT200-Bx

I could be wrong though.

my understanding was that the letter (sic "B") stood for the base silicon and the number (0) stood for the mask/deposition layers.

so the base silicon design, which is kind of laid in stone so to speak, is very hard to change. The masks that control the other doping layers and copper are easier changed so that's what you revise/respin.

if fermi had fundamental problems based on the A base layer (from either tsmc's end or nv's architecture), no amount of revisions(A0 -> A1, A2) was going to fix it. If they have figured out the design/materials problems, then a new B base layer could get closer to performing as gf100 was originally intended.

 

[Skurge]

That they did something like they did with G80->G92->GT200:

This sounds just like what AMD are doing.

Cypress->Barts->Cayman

Cypress was the base DX11 arch, Barts trimmed the fat (was also slower than Cypress, just like G92 was slower than G80) and Cayman is Scaled up again.

 

[n0x1ous]

This sounds just like what AMD are doing.

Cypress->Barts->Cayman

Cypress was the base DX11 arch, Barts trimmed the fat (was also slower than Cypress, just like G92 was slower than G80) and Cayman is Scaled up again.

agreed. Remember how FUDO said that Cayman was going to be the biggest chip AMD ever made? (Not that FUDO is a beacon of accuracy, but I think the odds are that Cayman is bigger than Cyrpess) did he say AMD or ATI? Because if its just the biggest chip AMD ever made it might only be 350-360mm right?

 

[Skurge]

agreed. Remember how FUDO said that Cayman was going to be the biggest chip AMD ever made? (Not that FUDO is a beacon of accuracy, but I think the odds are that Cayman is bigger than Cyrpess) did he say AMD or ATI? Because if its just the biggest chip AMD ever made it might only be 350-360mm right?

Well, he could have said AMD's graphics division.

 

[Idontcare]

my understanding was that the letter (sic "B") stood for the base silicon and the number (0) stood for the mask/deposition layers.

so the base silicon design, which is kind of laid in stone so to speak, is very hard to change. The masks that control the other doping layers and copper are easier changed so that's what you revise/respin.

if fermi had fundamental problems based on the A base layer (from either tsmc's end or nv's architecture), no amount of revisions(A0 -> A1, A2) was going to fix it. If they have figured out the design/materials problems, then a new B base layer could get closer to performing as gf100 was originally intended.

There's no "set in stone" definition for stepping nomenclature but generally the iterations are intended to communicate what is being iterated within the die itself.

A change in the stepping letter A->B->C is usually done to communicate that changes have occurred at the transistor level (which inherently requires changes at the wiring level, aka the BEOL). Circuits are moved around, if not redesigned or re-layed out to mitigate some undesirable attribute of the original circuit layout.

A change in the stepping number 1->2->3 is usually done to communicate that changes have primarily occurred in the wiring levels of the layout. The transistors pretty much stay where they were but the wires may get routed a little differently.

The way masksets are generated, a change in the upper levels that preserves the masksets already made for the lower-levels is less costly to implement and validate (hence it is dubbed "less complex"). A stepping change that involves transistors will likely invalidate the upper level masks as well, increasing cost and validation (hence it is dubbed "more complex").

Thus in laymens terms we are told to view a number-changing stepping revision A1->A2 or B0->B3 as a "minor stepping revision" while viewing a letter-changing stepping revision B3->G0 as being a "major stepping revision".

 

[nitromullet]

It's unmanufacturable because Nvidia have chosen not to manufacture it beyond what they'd committed to before seeing the outcome. I think that's a valid sense of the word.

IDC already explained what the term 'unmanufacturable' means in the industry (which he works in), and Fermi clearly doesn't fit that definition.

Fermi may be unfeasible or unprofitable, but the two Fermi chips in my rig refute the notion that it's unmanufacturable. You obviously can have your own opinion on the matter, but i fail to see the benefit of incorrectly redefining words to win a semantics argument.

 

[Kuzi]

I remember asking about this when the GTX460 first came out because I use CUDA and was interested in knowing if the 460 was neutered when it came to GPGPU.

The responses at the time were that it was not architecturally neutered any more so than the rest of the architecture was scaled back from GF100 (fewer cuda cores means fewer cuda cores but not less capable cuda cores, etc)...and I vaguely recall being referred to a Tomshardware link where they benched the 460 in a bevy of GPGPU apps and its performance fell in-line with what you'd expect for a Fermi chip that was simply cut-down (i.e. fps fell just as much as GPGPU performance compared to a 470 and 465).

Now this is my fuzzy memory, was I under the wrong impression? Is there a review you could link me to that would help dispel my ignorance?

Honestly I haven't seen a review directly comparing the GPGPU performance of both cards, with their gaming fps results. I'll try to look into that IDC

I've rechecked Anand's review though and found a few interesting things:

* ECC is gone from GF104
* GF104 has 25% less CUDA cores (384 vs. 512) but has the same number of texture units (64). Though in reality GF104 has 7% less "enabled" texture units (56 vs. 60).
* FP64 was removed from 2 of the 3 blocks of CUDA cores, so FP64 instructions execute at 1/4 FP32 performance (handicapped from a native 1/2).

From those points alone you can see that NV wanted a smaller, more gamer oriented card that would compete better against ATI's offerings, and I believe they did succeed with that.

 

[Zstream]

I'll take a wager that they take the cuda cores off and sacrifice tessellation performance.

 

[nitromullet]

From what I hear...... You will miss it by 12 days. ,
but I've been wrong before.

Honestly I'm leaning towards the 6970 right now anyway, but a step-up option would probably make the decision a little bit more difficult.

This is a slighter tougher new generation to justify than most previous new generations. There are no new features and no die shrink, so it's all about who can offer the best evolutionary performance gains. Hearing numbers like 10-15% don't exactly make me to enthusiastic about eating the loss on selling two slightly used GTX 470s, and turning around and dumping a few $100 on the new cards. The main driver would be that the GTX 470s will probably completely worthless if I hang on to them too long.

 

[AtenRa]

GF100 vs GF104

No ECC
Half the Polymorph Engines (16 vs 8)
Half the Raster Engines (4 vs 2)
FP64 ¼ vs ½
256-bit vs 384-bit Memory controllers
Because we have 2 less 64-bit Memory Controllers we have 256kb less L2 cache and less 16 ROPs (48 vs 32)

Texture Units analogy goes up to 1 Texture Unit for every 6 Cuda Cores vs 1 Texture Unit for every 8 Cuda Cores in GF100.

More compact SM design with 48 Cuda Cores and 2 Schedulers vs 32 Cuda Cores with 2 Schedulers. (Superscalar design)



Most of those changes don’t effect GPGPU that much (except ECC and FP64). I believe most of those changes effect Tessellation performance and Power Usage.

 

[tviceman]

I'll take a wager that they take the cuda cores off and sacrifice tessellation performance.

Errrhhmmmm ummmmm.... I don't think you can take away "cuda" from Nvidia's cores. That's just what they are now, or rather, what Nvidia calls them.

 

[AtenRa]

http://www.chiphell.com/forum.php?mod=viewthread&tid=133004&extra=page/=1&page=4

GTX 580 @ 772/1544/2004MHz ? 64TMU

GTX 480 @ 925/1850/1950 MHz with 260.99 WHQL

http://bbs.expreview.com/viewthread.php?tid=37391&rpid=274471&ordertype=0&page=1#pid274471

 

[Spike]

http://www.chiphell.com/forum.php?mod=viewthread&tid=133004&extra=page/=1&page=4

GTX 580 @ 772/1544/2004MHz ? 64TMU

GTX 480 @ 925/1850/1950 MHz with 260.99 WHQL

http://bbs.expreview.com/viewthread.php?tid=37391&rpid=274471&ordertype=0&page=1#pid274471

That looks pretty cool... except for the scary panda. I don't want that looking out at me from my gpu...

 

[tviceman]

gtx580 @ 772mhz is beating a gtx480 @ 925mhz? If true, that is more than a 20% increase in performance. If all the leaked benchmarks regarding Cayman have truth to them, then the recent price war just started may end up lasting all the way to 28nm!