[dnavas]

Originally Posted by trinibwoy

If that's the case then it can't be one dot per lane like in the GT200 diagram. Nvidia's diagrams had each SFU unit as being 4 lanes wide. In which case the SP:SFU ratio would have risen to 2:1.

Wouldn't SFUs mostly scale by texture unit, rather than SP?
If textures remain fixed at 8 per TPC, I would think 4 SFUs would suffice (assuming, as you say, that these aren't indicating per-lane). You could even have TUs grow to around 10 and be okay. That's roughly the same SFU/TU ratio (3/8 vs. 4/10) and similar SP/TU ratio (24/8 vs. 32/10). 160TUs total would be ... smooth.

It does make it seem like G[T]200's DP functionality was thrown in (this one seems integrated), and it makes me wonder whether this fellow would choke on inverse/division problems.

But if this thing really "only" has 512 ALUs, what sort of clocks would it need to be competitive with HD 5870, and more importantly the X2?

Well, at 2ops/clock, you'd have 2TF at 2Ghz. Wasn't that the original aim of the G200 chips? [sorry, my mind is fuzzy this early in AM]

Of course, for all we know, the diagram is hugely misleading, the four dots at the top handle branching and instruction re-ordering for multiple WARP/clock issue, and the SFUs are integrated into the data address and setup "bars". Yeah. And the 32 items are really register/cache and TUs, and there's just one big ALU at the bottom which runs at an effective speed of around 100Ghz, but using logic that doesn't require explicit clocking....

Heh.

-Dave

[trinibwoy]

Originally Posted by dnavas

Wouldn't SFUs mostly scale by texture unit, rather than SP?

Why would they? They don't feed the texture units in any way AFAIK.

Well, at 2ops/clock, you'd have 2TF at 2Ghz. Wasn't that the original aim of the G200 chips? [sorry, my mind is fuzzy this early in AM]

Yeah, crazy high clocks are one option but we haven't heard anything to that effect as yet. That fuzzy diagram also doesn't provide any hints of an increase in triangle setup throughput.

[jaredpace]

~545mm2 ?

[Anteru]

Originally Posted by rpg.314

Any body has any idea what prompted the code name change? From GT300->G300->GF100?

Surely as it's the spritiual successor of the VSA-100 chip, and actually, what we see here ladies and gentlemen ... is the Voodoo Rampage! nVidia has been holding it back until a real emergency, and now is the time.

[DegustatoR]

Originally Posted by rpg.314

Ok, if we are in the silly season, I propose that nv release a GG300 after GF100.

Don't know about you but I'd expect them to ride Fermi for some time so we should expect GF11x and GF20x before that GG of yours. Plus it should be GG100 anyway.

[chavvdarrr]

well, "gg" does have special meaning for online gamers

[dnavas]

Originally Posted by trinibwoy

Why would they? They don't feed the texture units in any way AFAIK.

I don't know. It made sense to me an hour ago before the sun rose and after I finished my run. I blame hypoxia [I thought they mostly fed TAs, but that's making a lot less total sense as I start to wake up.]

Were the SFUs rebuilt to be more precise?

http://74.125.155.132/search?q=cache...ient=firefox-a

Higher precision would be useful, and I'm not sure what size LUT is required for interpolation of DP, so maybe they abandon that approach....

Yeah, crazy high clocks are one option but we haven't heard anything to that effect as yet. That fuzzy diagram also doesn't provide any hints of an increase in triangle setup throughput.

Well, I haven't heard jack diddly. :shrug:

-Dave

[nAo]

I wouldn't be so worried about raw flops numbers. HD5890 numbers are huge, but it doesn't seem to be any better than its predecessor when it comes down to use all that raw power (I'd say the only welcome change was moving interpolators workload over the shader cores), it's actually worse given that is more bw constrained (still an impressive part though ).
Modern workloads will be more and more about being fast at less regular computations and data structures than about pushing more MADDs per clock cycle. Although it won't happen overnight

[trinibwoy]

Sure, but what are they going to do about all those not-quite-modern console ports that we'll be playing for the next 3 years?

[apoppin]

i think we can assume (putting it all together and presuming that G80 was GT v1.0):

GT200 = GT architecture, v2.0
GT21x = GT architecture, v2.1
GF100 = GF architecture, v1.0; Fermi 1.00
--- F=Fermi, T=Tesla and G=GPU Core
. . . (GT200 surely means "GPU core Tesla 2.00")

/end speculation .. and i am looking for a new GF

[Ailuros]

Originally Posted by DegustatoR

Why? It does actually.
GT200 = GT architecure, v2.0 (presuming that G80 was GT v1.0)
GT21x = GT architecure, v2.1
GF100 = GF architecure, v1.0
With F=Fermi, T=Tesla and G=... Graphics? GeForce?

I wouldn't be surprised if the GT200 was more of an intermediate addition in an ancient roadmap. G8x->G9x->G1x0.

[nAo]

Originally Posted by trinibwoy

Sure, but what are they going to do about all those not-quite-modern console ports that we'll be playing for the next 3 years?

Nothing, they will run just fine.

[DegustatoR]

Originally Posted by Ailuros

I wouldn't be surprised if the GT200 was more of an intermediate addition in an ancient roadmap. G8x->G9x->G1x0.

Maybe. As G70 was an intermediate addition to the NV line =)
But now it looks more or less solid and will keep everyone guessing what will come next: GF110, GF200 or GG100 -)

[apoppin]

Originally Posted by Ailuros

I wouldn't be surprised if the GT200 was more of an intermediate addition in an ancient roadmap. G8x->G9x->G1x0.

Of course; Nvidia is consistent since their beginning - their cores have always been named for scientists and they have always followed the same numbering convention

i re-did my chart a bit:

GT200 = GT architecture, v2.0
GT21x = GT architecture, v2.1
GF100 = GF architecture, v1.0; Fermi 1.00
--- F=Fermi, T=Tesla and G=GPU Core
. . . (GT200 surely means "GPU core Tesla 2.00")

what is exciting; it is new architecture .. but i doubt much has changed for gaming

[seahawk]

Originally Posted by neliz

Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.

So hardly ahead of 5890.

[DegustatoR]

Originally Posted by apoppin

but i doubt much has changed for gaming

Well, PhysX is pure computation. DXCS is pure computation. And both are "for gaming".
What can be changed for gaming beyond support for DX11 anyway?

[Ailuros]

Originally Posted by seahawk

So hardly ahead of 5890.

[Razor1]

Originally Posted by seahawk

So hardly ahead of 5890.

don't expect a refesh early next year

[trinibwoy]

Originally Posted by nAo

Nothing, they will run just fine.

Of course, I meant vs the competition.

[apoppin]

Originally Posted by DegustatoR

Well, PhysX is pure computation. DXCS is pure computation. And both are "for gaming".
What can be changed for gaming beyond support for DX11 anyway?

that is not what i meant

i believe we will see much more improvement in parallel processing; Nvidia's direction away from pure gaming

[dnavas]

Originally Posted by nAo

Modern workloads will be more and more about being fast at less regular computations

Hmm, are you arguing for moving some of the "SF" instructions into the SPs? I would think log/rcp would be really useful there (hence my previous link). SF could be relegated to sin/cos approximations, or those blue dots might be something else entirely.

and data structures

...caches?

Just trying to keep up.

[nAo]

No, I am not arguing for that.

[w0mbat]

Originally Posted by neliz

Performance figures provided along with the specifications talk about 70% of a 5870CF setup. So 20% over a 5870.
Hence the notion of a X2, it is required for performance leadership.

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

[ShaidarHaran]

Originally Posted by w0mbat

That would be a hughe, negative, surprise. I see GF100 ahead or minimum on par with HD5870 X2.

Seems unlikely to me. Such a chip would need to be well over twice as fast as GT200. On paper, sure, but in real world apps, I don't think so.

[w0mbat]

I think 2.5*GT200 was the main goal.