NVIDIA Fermi: Architecture discussion

[Groo The Wanderer]

I'm sure Fellix will want to have a go at this: I have just scaled the die shot of GT215 so that the PCI Express blocks match in size with the GF100 die shot. This leads me to estimate 480mm² die size for GF100.

Strangely the GDDR5 interfaces on the two chips are scaling very differently, to the extent that the address/command section and the data section are scaling differently from each other. And also don't agree with the PCI Express based scaling.

Jawed

The chip is 23.x * 23.x, the first eyeball estimate I got was 23 * 23, but that is a tad low.
http://www.semiaccurate.com/2009/08/18/nvidia-takes-huge-risk/

There is a reason for the GDDR5 scaling.

-Charlie

 

[Blazkowicz]

7 years ago NVIDIA tested packed GPUs from Fab like this:
...
We know Fermi is a working silicon:

...

So I wonder how does today's Fermi test board look like? Does it really have a lot of wires sticking out, as well as test modules, and looks like a character from Terminator like Fudo describes or is it something similar to year 2002's test board

I guess it looks like this

 

[XMAN26]

The chip is 23.x * 23.x, the first eyeball estimate I got was 23 * 23, but that is a tad low.
http://www.semiaccurate.com/2009/08/18/nvidia-takes-huge-risk/

There is a reason for the GDDR5 scaling.

-Charlie

When will it end with the predictions on performance:

"Cypress will slaughter Nvidia if it takes this route."

From the link from his own article.

Charlie, you were wrong with R600, you were Wrong with R770. You really need to stop while your ahead.

 

[eastmen]

I don't see any volumes of ATIs DX11-Chips. All they need is a (Paper)-launch to make people wait a little bit further for Fermi.

Only Problem for nvidia is the performance. Fermi must beat RV870 by 30-40%. Then everything is fine. Even when they deliver 3-4 months later than ATI.

I'm just wondering

1) Isn't the Fermi going to be almost 1B tranistors bigger than the RV870 ? 3B vs 2.15B tranistors ? So shouldn't the Fermi be much bigger on 40nm than the RV870 meaning it will cost more also having lower yields ?

2) Do we really know if this is all ati has ? We haven't seen the x2 part performances and if its a paper launch this year with us waiting until febuary or later for a real launch on the fermi ati can release a 5890 that would close the gap on the 30-40%

I think nvidia can come out on top , I just don't see them really responding this year and I think ati can counter punch.

 

[seahawk]

That scenario is very likely. And do not forget that the high-ed Fermi is worth nothing against Juniper.

 

[rpg.314]

May be the shortage/marketshare discussion belongs in the nv strain thread? *shrugs*

 

[DegustatoR]

Haven't you been paying attention to Degustator? he said GF100 taped-out at the beginning of this year,

Never said that.

MP would've started before May ..

Never said that either.

so secretly all graphics chips made by nv are GT300's which will be unlocked in a "Can of Whoop-ass/Power Of 3" driver that will be released soon! honestly!

You're funny as Charlie.
And I'm still waiting on this "9 months later than Cypress" thing you were speaking some weeks ago, yeah.
How's that GT212 coming BTW?

And when does it come out? November (nobody will give that credit)?

November-December clearly.

And where are the smaller parts? November too?

Will they need a smaller parts before 2010?

 

[AlexV]

Aaaaand back to Fermi now, thanks.

 

[trinibwoy]

Can one of the mods please move the off-topic posts to the Nvidia strain thread? I think the alleged pulling out of the high-end market by Nvidia is discussion worthy.

 

[AlexV]

Since you asked so nicely, I did just that...and I see some 1 liner post pruning is still needed. Please keep this thread on topic gentlemen!

 

[trinibwoy]

Thank you

 

[Jawed]

Mostly i agree. It's quite hard to filter out the different limitations, but i think there's a lot improved internally in Fermi, whereas AMD chose more or less the brute force approach,

Like RV770?

scaling mainly the number of execution units while leaving internal bandwith from and to caches etc. untouched an thus on the level of the identically clocked HD 4890.

L2->L1 is untouched, and GDS->anything could be untouched (though GDS is 4x larger - and it may have additional features aimed specifically at compute), but everything else looks like it's scaled.

L2->L1 not scaling looks like a severe problem - and theoretically indicates that the next architecture has to be quite different in this respect. I'm interested to see how well prunedtree's matrix multiplication works on HD5870, to see if L2->L1 is causing a problem.

Nvidia otoh did chose to overhaul much of their architecture to the point that we still don't even know the #s and caps of their TMUs and ROPs.

Nor how useful the prior architecture was for D3D11. e.g. VS-HS-TS-DS-GS-PS may not work so well on the prior architecture.

Still have to wait to see how well it works on ATI.

Jawed

 

[seahawk]

The most interesting question about Fermi is, how are the little GPUs are doing? This is more important then the final release date of D12U.

 

[Ailuros]

The most interesting question about Fermi is, how are the little GPUs are doing? This is more important then the final release date of D12U.

They're doing fine and are waving hello....LOL

Seriously now I don't think anything regarding possible release dates is known yet.

 

[mapel110]

I'm just wondering

1) Isn't the Fermi going to be almost 1B tranistors bigger than the RV870 ? 3B vs 2.15B tranistors ? So shouldn't the Fermi be much bigger on 40nm than the RV870 meaning it will cost more also having lower yields ?

2) Do we really know if this is all ati has ? We haven't seen the x2 part performances and if its a paper launch this year with us waiting until febuary or later for a real launch on the fermi ati can release a 5890 that would close the gap on the 30-40%

I think nvidia can come out on top , I just don't see them really responding this year and I think ati can counter punch.

1. thats why it should be 30-40% faster, so nvidia can get more money for it.
2. X2 is multiGPU with all its benefis and limitations. Single GPU High End is more important to most High End Users I know.
A possible 5890 will only raise the frequencies a bit. I doubt this can create 30-40 % more performance.

ATI has launched mid range Chips. nvidia will hopefully come with real high end.

 

[leoneazzurro]

November-December clearly.

Link to the official announcement? No, Fudzilla isn't official.
ATI publicy demonstrasted working cards in June before launching in September. Nvidia did not demonstrate even a non-working card. I really hope for competition that they are not really in troubles, but until now the situation hints in that direction.

Will they need a smaller parts before 2010?

Yes, they need. Juniper, if really a 1120 So part, will perform in the 4870-4890 range and go in direct competition with GTX260-275. With a 128 bit bus and much smaller die. 5850 practically EOLed GTX285 thanks to lower price and higher performance and DX11 compatibility.
So Nvidia needs smaller parts competing with Juniper, so a GT200 shrink to 40 nm or Fermi derivatives that at the moment are not yet even given a name.

 

[rpg.314]

And by the looks of it, 57xx will EOL G92 and it's various off-springs/namesakes...

 

[DegustatoR]

Link to the official announcement? No, Fudzilla isn't official.
ATI publicy demonstrasted working cards in June before launching in September. Nvidia did not demonstrate even a non-working card. I really hope for competition that they are not really in troubles, but until now the situation hints in that direction.

Intel demostrated something LRB-like a long time ago. That's all dust in your eyes and nothing more. Only real launches and sales matter. NV is clearly betting on Nov-Dec launch right now. I (and you) have no idea if they'll be able to do it.

Yes, they need. Juniper, if really a 1120 So part, will perform in the 4870-4890 range and go in direct competition with GTX260-275. With a 128 bit bus and much smaller die. 5850 practically EOLed GTX285 thanks to lower price and higher performance and DX11 compatibility.
So Nvidia needs smaller parts competing with Juniper, so a GT200 shrink to 40 nm or Fermi derivatives that at the moment are not yet even given a name.

Juniper is not a 1120 SP part, it's a 800 SP part with 128-bit bus. It will end up close to RV770 (not RV790) in lower resolutions and will compete with GTS250 and GTX260 as of now. GTX275 will be somewhere between 5770 and 5850.
That's clearly a bad situation for NV but it's far from catastrophy of galactic proportions. And 55GP wafers are still cost substantially less than 40G wafers thus Juniper being smaller doesn't really mean anything at the moment.

 

[leoneazzurro]

Intel demostrated something LRB-like a long time ago. That's all dust in your eyes and nothing more. Only real launches and sales matter. NV is clearly betting on Nov-Dec launch right now. I (and you) have no idea if they'll be able to do it.

Betting is not sure win, especially when you take high risks

Juniper is not a 1120 SP part, it's a 800 SP part with 128-bit bus. It will end up close to RV770 (not RV790) in lower resolutions and will compete with GTS250 and GTX260 as of now. GTX275 will be somewhere between 5770 and 5850.
That's clearly a bad situation for NV but it's far from catastrophy of galactic proportions. And 55GP wafers are still cost substantially less than 40G wafers thus Juniper being smaller doesn't really mean anything at the moment.

All stores out there are quoting 1120 SP parts. Of course they can be all wrong, but there is also this:

http://translate.google.com/transla...d.php?t=6947&prev=hp&rurl=translate.google.it

if you look a the feature tests, especially FT1, FT3 and FT6, where the others are clearly BW limited... Color fill hints also to a 8 ROP configuration, but this could be also a driver glitch and/or BW limitation.

 

[Kaotik]

A possible 5890 will only raise the frequencies a bit. I doubt this can create 30-40 % more performance.

ATI has launched mid range Chips. nvidia will hopefully come with real high end.

Will it? Based on all the current info there is there's no way Cypress is as big as it is unless there's something "hidden" (like those possible extra SIMDs which could be enabled for HD5890), when Juniper apparently has 1120 SPs aka 14 SIMDs

Degustator, how come you're so sure that Juniper would be 800SP part, when practicly every possible bit of info excluding ATI slides which we haven't seen yet say it's 1120SP?