The official GT300/Fermi Thread

[annihilat0r]

Yeah but... why do you expect it to have 2x the performance (I'm suppose you're talking about real world performance) if it's going to have +113% CPs more but only +50% ROP more, +60% mem bandwidth more...

Consider that HD5870 is exactly double the HD4890 (+100% everything at the same clocks) except bandwidth (aprox. +30%) and it's far from double the real world performance (that's one of the most recent proves that doubling everything doesn't mean doubling real world performance), and NVIDIA is not even doubling processing units.

Can they improve the performance per unit and per clock? Sure. Maybe. But how much and why, I think is way soon with the info we have to say it's going to be 2x real world performance of a GTX285. I even would say I hugely doubt it, given that they are more focused in get the new (future?) HPC market before Intel has their Larrabee working (if it happens to be on this century).

You're being almost dogmatic with this post. AMD and Nvidia are completely different brands with completely different chips. Saying "amd doubled everything but didn't double the performance, so nvidia can't double 285's performance" is less worthy than not saying anything, which most people in this thread should do.

[DilTech]

The 5870 is pretty much a 4870 doubled up with a little faster ram. As we all know, architectures do reach their ceilings.

I will remind you, however, that the 4870 had the same amount of rops as the 3870, but they improved how well they handle(ipc if you will). The 4870 wasn't double the 3870 in specifications. 4870 was easily 2x as fast as the 3870 when AA was enabled. The 9800GTX has less rops than the 8800GTX but for the most part was faster(even if only mildly), and also had less memory bandwidth. The only reason the 5870 isn't 2x the 4870 is because AMD are reaching the point of diminishing returns with their architecture. See, they still count on multi-way shaders, which we do know is very hard to get them all working at the same time. NVidia go for simpler shaders which is why, even with less of them, they've had no problem competing. Yes, the shaders run at a faster clock speed, but when you're pitting 240 up against 800, and winning, it's pretty telling about who's more efficient.

This is also why in some titles you'll see the 4890 so close to the 5850, even though the 5850 should destroy that card as they're the same architecture but the 5850 has much better specs. Some titles just don't play nicely with ATi's shader design, but the ones that do FLY on it.

In NVidia's case, they even specifically stated in the article they were disappointed with the shader efficiency with the GTX-280, which is a tell-tale sign that this part is to be a lot more efficient in it's shader use. Now, you increase shader efficiency and OVER double them and tell me what happens, along with more rops, a lot more memory bandwidth... There's a big reason why I say if it's less than double the performance I'll be in shock.

[jfromeo]

I find Farinocco's post pretty logical.

He isn't stating nVidia can't double 285's real performance. He is just saying that by doubling the specs (VS, PS, TMUs, ROPs...) while maintaing clocks (plus, mem clock was increased +225MHz) on a similar architecture doesn't lead to double the real world performance, as we have checked with HD5870 and HD4890.

[DilTech]

jfro, he's forgetting one thing though... The GT300 is a completely new architecture, that's been confirmed. Based loosely on the G80, but it is a new architecture.

4870 to 5870 was just a doubled up chip with DX11 added to it.

[Farinorco]

You're being almost dogmatic with this post. AMD and Nvidia are completely different brands with completely different chips. Saying "amd doubled everything but didn't double the performance, so nvidia can't double 285's performance" is less worthy than not saying anything, which most people in this thread should do.

Who has said what you say that I've said? Read again, please

The 5870 is pretty much a 4870 doubled up with a little faster ram. As we all know, architectures do reach their ceilings.

I will remind you, however, that the 4870 had the same amount of rops as the 3870, but they improved how well they handle(ipc if you will). The 4870 wasn't double the 3870 in specifications. 4870 was easily 2x as fast as the 3870 when AA was enabled. The 9800GTX has less rops than the 8800GTX but for the most part was faster(even if only mildly), and also had less memory bandwidth. The only reason the 5870 isn't 2x the 4870 is because AMD are reaching the point of diminishing returns with their architecture. See, they still count on multi-way shaders, which we do know is very hard to get them all working at the same time. NVidia go for simpler shaders which is why, even with less of them, they've had no problem competing. Yes, the shaders run at a faster clock speed, but when you're pitting 240 up against 800, and winning, it's pretty telling about who's more efficient.

This is also why in some titles you'll see the 4890 so close to the 5850, even though the 5850 should destroy that card as they're the same architecture but the 5850 has much better specs. Some titles just don't play nicely with ATi's shader design, but the ones that do FLY on it.

In NVidia's case, they even specifically stated in the article they were disappointed with the shader efficiency with the GTX-280, which is a tell-tale sign that this part is to be a lot more efficient in it's shader use. Now, you increase shader efficiency and OVER double them and tell me what happens, along with more rops, a lot more memory bandwidth... There's a big reason why I say if it's less than double the performance I'll be in shock.

HD4870 was 2.5x SPs, 2.5x TUs, near 2x memory bandwidth the HD3870, and even then, it only was about double in MSAA scenaries, with a completely reworked AA logic that was the great fault of HD3870 (AA made deep sadly the performance of HD2000 and 3000 series), so I don't find it to be the best example.

Of course, doubling specs is doubling potential performance, but that doesn't happen everytime. Indeed, it only happens in best case situations. That's why I say that doubling specs is not equal to doubling performance in real world cases (as a whole, there are always concrete cases).

I understand that may be other changes that can be made (mainly architectural changes) that may affect performance. I only say that I wouldn't take for granted that there are going to be so many changes as to make so a drastical architectural improvement, based on what we know right now about the new architecture. I'm not saying it's not going to happen. If it does, great for everybody

jfro, he's forgetting one thing though... The GT300 is a completely new architecture, that's been confirmed. Based loosely on the G80, but it is a new architecture.

4870 to 5870 was just a doubled up chip with DX11 added to it.

I'm not forgetting that. I'm only making my own interpretation of what the words "a completely new architecture based on the previous one" means in the lips of a hardware vendor: an evolution of the previous one, i.e. taking the previous one and applying some architectural changes and improvements. There are going to be architectural changes, because there are several changes announced about the GPGPU matter. Only with this would be enough to justify that "completely new architecture based on the previous one" thing.

Am I saying that there are not going to be per unit and per clock improvements at 3D rendering performance? Nuop. I'm only saying that I wouldn't take it for granted based solely on these words. If I wouldn't take the possibility of that happening, I would say that certainly isn't going to be 2x real world performance, and I'm not saying that, again.

GT200 was said to be "a completely new architecture based on G80 one" also. And it was, if you understand it as an evolution of the previous one, mainly in the GPGPU aspect.

[annihilat0r]

NEWS NEWS NEWS NEWS NEWS

Some guy at the Beyond3D forum got to talk to some Nvidia officials about Fermi, two important points:

- They guys told that what Jensen was holding in his hands at the presentation was a production mockup. However, they do have silicon working
- Ships in '09
- Gaming performance is still mysterious but he said that some guy told him it was about 1.6-1.8x of GTX 285.

[DilTech]

Seeing as how this thread has the most info on the card compared to the rest, it's now the official thread. Please keep all GT300 info in this location.

Thanks in advance guys.

[mibo]

@annihilat0r
I really don't know why you are worrying about performance.
Nvidia will release a card that beats the 5870. This is their goal!! They will push the frequencies and voltages to get to this performance.

[Elfear]

NEWS NEWS NEWS NEWS NEWS

Some guy at the Beyond3D forum got to talk to some Nvidia officials about Fermi, two important points:

- They guys told that what Jensen was holding in his hands at the presentation was a production mockup. However, they do have silicon working
- Ships in '09
- Gaming performance is still mysterious but he said that some guy told him it was about 1.6-1.8x of GTX 285.

So ~SLI GTX 285 performance it sounds like. That sounds about right judging from previous generations. New gen is about equal to old gen multi-gpu. Hopefully the price won't be astronomical.

I'd be very surprised if Fermi ships in '09 unless they mean three cards to Newegg by Christmas time.

[Farinorco]

NEWS NEWS NEWS NEWS NEWS

Some guy at the Beyond3D forum got to talk to some Nvidia officials about Fermi, two important points:

- They guys told that what Jensen was holding in his hands at the presentation was a production mockup. However, they do have silicon working
- Ships in '09
- Gaming performance is still mysterious but he said that some guy told him it was about 1.6-1.8x of GTX 285.

Great about shipping dates news! I was really about they releasing on 2010... let's hope it isn't 20th december

[Bodkin]

Great about shipping dates news! I was really about they releasing on 2010... let's hope it isn't 20th december

Yeah, I am prying I can nab one in time for Christmas

[spajdr]

Die pic for ya

[Kaldor]

Charlie is usually full of , but he makes some very compelling arguments here:
http://www.semiaccurate.com/2009/10/...mi-boards-gtc/

[Bodkin]

Whats with the wired colours?

[DilTech]

Charlie is usually full of , but he makes some very compelling arguments here:
http://www.semiaccurate.com/2009/10/...mi-boards-gtc/

Charlie made one big error, his statement that A1 is the first revision...

Anyone who knows a thing or two about chips knows A0 is the first revision. Anyone remember the Q6600 update, the G0(that's g zero)? If Charlie was correct in that assumption then that famous stepping would have been G1, not G0. These aren't counting the prototype samples which are just to test the features without making the full blown chip.

As such, kind of blows his argument clean out of the water in that regard, doesn't it?

[Roger_D25]

Thanks Kaldor, that was a very interesting article about the Fermi sample shown (I'm not an expert but his points seem accurate)? I assume that many companies do this type of thing when showing off new hardware?

[DilTech]

Thanks Kaldor, that was a very interesting article about the Fermi sample shown (I'm not an expert but his points seem accurate)? I assume that many companies do this type of thing when showing off new hardware?

Mockups are very common place amongst all forms of products.

[Farinorco]

Charlie is usually full of , but he makes some very compelling arguments here:
http://www.semiaccurate.com/2009/10/...mi-boards-gtc/

The first points about the serial numbers and dates on the IHS seem good (excepting the absolutely idiotic one about the "7") about defending their own previous writings.

But even if absolutely pointless (well, they have shown a "fake" card to decorate the presentation, who cares? It's not like this would mean anything) what I have enjoyed more of the article is the part about the "fake" card. Hey, the way it's written it is hilarious... "Those lead to... well, not the power connector", "The 6-pin connector, on the other hand, lines up with, umm, nothing", "Except glue. Notice the connector is black and the hole below it shows white. The only real question now is, Elmers or glue stick"... by here I was

[Roger_D25]

I thought that might be the case DilTech, thanks! Its probably a discussion for another day (or thread) but should people be upset that Nvidia tried to sell the mockup as the actual Fermi product (in this case probably should since it gives the impression their further along than they actually are)?

Farinorco - I also got a good laugh at the way its written!

[mibo]

Charlie made one big error, his statement that A1 is the first revision...

Anyone who knows a thing or two about chips knows A0 is the first revision. Anyone remember the Q6600 update, the G0(that's g zero)? If Charlie was correct in that assumption then that famous stepping would have been G1, not G0. These aren't counting the prototype samples which are just to test the features without making the full blown chip.

As such, kind of blows his argument clean out of the water in that regard, doesn't it?

I know you hate Charlie for being a Nvidia hater, but except for the A0/A1 confusion he is right most of the time.
Didn't he write that a few GT300s might make it in 2009 but real availability will not start before 2010?
And by the huge amount of working GT300s that were shown to the audience, his yield numbers might be not that far off...

[highoctane]

I didn't see the Fermi thread and didn't see this mentioned anywhere yet, looks like Fermi has a chance to gain traction.

http://www.dailytech.com/ORNL+to+Use...ticle16401.htm

[Jamesrt2004]

Charlie made one big error, his statement that A1 is the first revision...

Anyone who knows a thing or two about chips knows A0 is the first revision. Anyone remember the Q6600 update, the G0(that's g zero)? If Charlie was correct in that assumption then that famous stepping would have been G1, not G0. These aren't counting the prototype samples which are just to test the features without making the full blown chip.

As such, kind of blows his argument clean out of the water in that regard, doesn't it?

ahh not to say your wrong or anythign but there was a massive discussion about this some time ago i remeber vividly and apparently Nvidia DO use "A1" as there first revisions


im not in the "know" to be 100% but i remember a argument here a while back

[Farinorco]

ahh not to say your wrong or anythign but there was a massive discussion about this some time ago i remeber vividly and apparently Nvidia DO use "A1" as there first revisions


im not in the "know" to be 100% but i remember a argument here a while back

I didn't wanted to say anything, because I am more or less in the same situation than you, but I have thought exactly the same.

I don't know thought. I don't have any kind of knowledge about this, so I prefer to stay away. I'm only mentioning this because as I have thought the same, and I was (and I'm) unsure, and I was sort of relieved when I read your post ("hey, man, I'm not crazy, that has existed" )...

[DilTech]

ahh not to say your wrong or anythign but there was a massive discussion about this some time ago i remeber vividly and apparently Nvidia DO use "A1" as there first revisions


im not in the "know" to be 100% but i remember a argument here a while back

I can tell you out-right that's not the case... The original NV15 chip(GF2 GTS IIRC) for OEM's was Rev A0. If you don't believe me just run a quick search thru this webpage and you'll find the info listed for you...

http://forums.gentoo.org/viewtopic-t-819-start-125.html

Same here, the gainward driver for that specific chip
http://www.givemefile.net/drivers/video/gainward.html

The reason so many think NV start at A1 is because almost never is the A0 perfect, and as such there's usually 1 to 2 revisions before it's ready for the public. As such most people never know they exist.

[AVB]

pic for ya




http://rs648.rapidshare.com/files/28...Key_Visual.jpg ( res. 6316 x 3240)

Fermi 1.4-1.6 x of GTX 295.

( 1.6-1.8x of GTX 285 is not too much)