Go Go NV40

[Kruno]

I was reading a magazine and found this. i scanned it and i ll translate it for you.
350 million transistors
0.09 microns
Core @ 800MHz
16MB DRAM
8 Pixel Rendering Engines
16 Vertex Shader Engines
DDR-II
512MB
Mem @ 1.4GHz
44.8GB/s bandwith
DX9.1 or DX10 features
2nd half of 2004 (After the olympic games!!)

http://discuss.futuremark.com/forum/showflat.pl?Board=techdisplayadapters&Number=2204827

http://www.warp2search.net/article.php?sid=12094&mode=thread&order=0

http://www.driverheaven.net/

[PreservedSwine]

Isn't Nvidia on schedule to launch the NV40 at COMDEX this year in Nov? I would have thought they'd stick w/ .13 a little longer than that, but who knows... I doubt we'll be seeing these specs in Nov......

Impressive looking vid-card, though:)

[Lezmaka]

You know this to be complete BS, yet you still post it as if they're real?
lol

[volt]

Wasn't this discussed in another thread? :confused:

[GlowStick]

Yes, yes it was.

April Fools Strikes again!

[mikechai]

I wonder who is the original starter of this BS.
These specs were already posted last year.

[Evildeus]

Let see, february 2003?

http://www.nvnews.net/vbulletin/showthread.php?s=&threadid=7550&highlight=NV40

Or september 2002?

http://forum.pcvsconsole.com/viewthread.php?tid=1783

[retsam]

Let see, february 2003?

http://www.nvnews.net/vbulletin/sho...&highlight=NV40

Or september 2002?

http://forum.pcvsconsole.com/viewthread.php?tid=1783




hahaha somebody please dont tell me that all this speculation is based rumors before the nv30 was even released gawd....

[Jandar]

welcome to the world wide misinformation network.

[Uttar]

Woah, not bad, not a SINGLE thing right.
Okay, maybe one or two, but still, those are only the obvious things ( and even those could be wrong, actually... We know way too little on the NV40 IMO... )


Uttar

[Paul]

Dave had it right over on B3D where he said most of that is more likely to be NV50.

It's certainly not NV40 just from the fact they mention DX10, H2 2004, and 0.09 - None of which are even remotely possible for the NV40, which is a DX9, Q4 2003, 0.13 product.

Oh, and 350million transistors, DDR2 at 700mhz... Can anyone say "Heat"?

[Uttar]

Originally posted by Paul
Oh, and 350million transistors, DDR2 at 700mhz... Can anyone say "Heat"?

Well, it might be possible to have that at a reasonable level of heat IF you had SOI *and* Low K along with a very mature 0.09 process. DDR2 @ 700Mhz seems insane though, more like the job of GDDR-3. And it's hard to speculate on the heat level of a memory type you've never seen :)


Uttar

[Paul]

Heh, sorry. I was still referring to the NV40 there, rather than the NV50. Hence 350million isn't realistic, given that it's a .13 product. Along with the fact that it's not using LowK, AFAIK.

EDIT: But is using SOI, apparently. Thanks to Uttar and GPU:RA for that tidbit :) Probably should have checked that before posting.

EDIT 2: On the same note, if a tapeout of NV40 is expected in July, there's no reason they can't get it to retail in December (although the announcement is in November, so the timeframe is a little short... possibly for the best though). Hmm, kind of makes me think about delaying my purchase of NV35 in September... I'll see how it looks then i guess.

[Uttar]

Originally posted by Paul
Heh, sorry. I was still referring to the NV40 there, rather than the NV50. Hence 350million isn't realistic, given that it's a .13 product. Along with the fact that it's not using LowK, AFAIK.

EDIT: But is using SOI, apparently. Thanks to Uttar and GPU:RA for that tidbit :) Probably should have checked that before posting.

EDIT 2: On the same note, if a tapeout of NV40 is expected in July, there's no reason they can't get it to retail in December (although the announcement is in November, so the timeframe is a little short... possibly for the best though). Hmm, kind of makes me think about delaying my purchase of NV35 in September... I'll see how it looks then i guess.

Remember the NV30 was supposed to tape-out around March or April, though. So until it tapes-out, might be a good idea to not be *too* sure it'll be available at year's end, even though it's very likely.


Uttar

[marcocom]

Originally posted by Jandar
welcome to the world wide misinformation network.

lol i heard that! good one.

remember that if its going to ship next year it does actually need to be designed and tested and working this year. so its not like were looking into a crystal ball. I think specs are more accurate than release dates. its the dates that are often the BS because nobody but the highest execs decide on that (and they dont even meet the engineers who build it, usually. they also have stocks and vested-interest in the confidentiality more than say...the quality assurance department of testers)

its the 'nv35 will ship next month' BS that not only seeds disinformation , but IMO seeds too much discontent with current cards and keeps people from just enjoying what they have. its vicious and destructive to a company's current product sales to say that crap and i think it should be punishable by law. (especially when 2000 jobs get cut because of it - thats alot of lives affected)

[Dazz]

I very much dobt this is the NV40, for one it won't have 512MB Ram 2nd won't have DX10 as MS have said D9 will be around for a couple of years. 44GB/s Bandwith is way! way! to much bandwith from the NV35 etc. 256bit, 1.4GHz GDDR is very unlikly.
16MB DRAM maybe, if it uses the same memory as the NV35 as DRAM will be used as a large cache. DRAM is VERY VERY expensive so 16MB seems unlikly. I would also like to note no one get hyped up as the same will happen to you again like what happened to the FX.

[Uttar]

Originally posted by marcocom
its the 'nv35 will ship next month' BS that not only seeds disinformation , but IMO seeds too much discontent with current cards and keeps people from just enjoying what they have. its vicious and destructive to a company's current product sales to say that crap and i think it should be punishable by law. (especially when 2000 jobs get cut because of it - thats alot of lives affected)

Alright, alright...
The NV35 will ship this month :p


Uttar

[Lezmaka]

Originally posted by Uttar
Well, it might be possible to have that at a reasonable level of heat IF you had SOI *and* Low K along with a very mature 0.09 process. DDR2 @ 700Mhz seems insane though, more like the job of GDDR-3. And it's hard to speculate on the heat level of a memory type you've never seen :)


Uttar

Samsung is planning at least engineering samples this quarter of ddr2 running that fast.
http://www.samsung.com/Products/Semiconductor/common/product_list.jsp?family_cd=GME0902

[Dazz]

nVidia have learned a valuble lession with just 500MHz DDR-2 i think think they are going to ruin their reputation again by getting 1.4GHz memory.

[MuFu]

I honestly don't think so. IMHO, it'll almost certainly use 2nd-gen GDDR2 (1.8V/1.8V) on a 256-bit bus clocked at ~700MHz for the "Ultra".

MuFu.

[Geforce4ti4200]

You wanna be realistic?


110 million transistors
500MHz ddr2 ram at 256 bit(the nv35 will be 400MHz)
500MHz core at mature .13 microns SOI
128mb and maybe 256mb ram
direct X 9.1

[busyL]

These NV 40 specs would as well have been extrapolated 1)from 2001's statement by Jen-Hsun-Huang in Monaco 2) adding some Moore's law :D
----------------------
"....18/05/2001 at 08:53 GMT
System Builders Summit, Monte Carlo Nvidia's president and CEO has delivered a tantalising glimpse of the kind of performance he expects future generations of Geforce boards to give.

Addressing a group of European System builders, Jen-Hsun Huang's made a presentation which included slides suggesting the GeForce 6 will be out in 2003 and built using 100 million transistors and running at 500MHz. The board will shift between one and five million polygons per frame.

Looking beyond, his spiel suggested that by 2006 there could be a GeForce 10 boasting 300 million transistors and running at 750MHz. Again this would render between one and five million polygons per frame.

Huang was on stage at the IFE/Systems Builders Summit in Monte Carlo. He didn't come out directly and say the GeForce 6 and 10 were on any specific roadmap but coyly put question marks against the board names and the years he suggested they might appear. The future casting seemed to be based on extrapolating the performance of the NV1 back in 1994 and how it handled Sega's Virtua Fighter, and then chuck in a bit of Moore's law.

Huang got quite hung up on what Nvidia boards, and the 3D graphics they generate, will mean to the world. He believe's the company's technology is riding a youth cult wave but is also part of something far more culturally significant. Being the subject of fevered roadmap speculation and hardcore gaming reviews pleased him greatly. "We're nearly the heroes of every 18 year old," he said. "3D graphics - it's cool," said his Powerpoint slide.

But later in his address he offered up quite lofty ambitions for his company's products. "3D graphics is going to be one of the most important art forms in the 21st century," he said. "The Geforce 3 provides an infinite palette for artists."

He then went beyond this. "In 10 years time we're going to transform your computer into a lifeform." It's possible he meant this literally, but he was probably looking towards more and more realistic Avatars - those computer generated folk who wish you a 'nice day' when you turn on your PC...."

link (http://www.theregister.co.uk/content/2/19036.html)

Besides :one of the most important thing - imho - is not mentioned >> which is the memory bus width 128 or 256 bits?


NEway I've got some newer info from a pal who's under NDA :angel: .....please don't spread the news yet :cool3:

Here goes...
------------------------------
>>NVIDIAZILLA's SHADOW RESEARCH CENTER - somewhere in ALASKA - May 7th 2003 -

After numerous failed press attempts to contact graphics chip manufacturing firm Nvidiazilla, J.H.H.(John - High - Himself) President and CEO of Nvidiazilla,(BTW I still don't get the point about his nickname "Jen-Hsun-Huang"?) held an exclusive press conference to announce a "killer" line of 62 new graphics cards to be released within the next three... or maybe thirty... or maybe a hundred months.

"I am excited to announce the 'XXXNVFXtreme' line of consumer-level graphics cards, each of which has at least 500 times the power of all the governmental supercomputers combined. These cards will take advantage of our newly acquired 'QRAM', as we've discovered that our current line of eDRAM is too slow and inferior, often only running games like 'Unreal Tournament' in 1600 x 1200 resolution at a paltry 429 frames per second. Then we moved to RRAM, URAM, and LARAM, but they weren't fast enough." When asked for a demonstration of the current line of cards, J.H.H. responded by saying, "I would, but I don't have one handy. I think I left the card in my hotel room. No wait, actually I accidentally left it behind at the airport's bar. I'll go back and get it later or something." He then proceeded to show off some very impressive graphs and diagrams, most of which look like they cost a lot to print up at Kinko's Copies.

"As you can see by this first chart, our Glaze Superion NFiniteFx Pro Quad Delta Mark VII is more powerful than Jesus Christ. It is capable of performing more than 50 billion calculations per nanosecond, most of which involve some kind of really hard arithmetic like long division. WITH a reminder"... J.H.H. then showed off a diagram which compared their upcoming video cards to all major political leaders, both alive and dead. Their card won every time.

The upcoming Nvidiazilla card will be 4.02x10^506 times more powerful than the nearest competition.
"The Glaze Superion NFiniteFx Pro Quad Delta Mark VII will also take advantage of our really cool 0.00000002 micron manufacturing technology which we discovered from a crashed UFO a few weeks ago. We had to delay the manufacturing process by a few days, because Uncle Dave came over and accidentally inhaled one. He's been coughing up blood ever since, but you can tell his pulse rate has increased by over 5,000 times. That's the power of the Glaze Superion NFiniteFx Pro Quad Delta Mark VII. It also doesn't require any power whatever to operate, and can function by solely running on the user's psychic energy."

Gary Talbeese, a columnist for British online magazine "Overclockin' Gits", asked J.H.H whatever happened to the dazzling 3D line of cards we've been promised for the past two years? Are we ever going to see those instead of noisy hair-driers?"

"Ah yes, Gary Talbeese, my old nemesis. Don't you remember the product demonstration I personally gave you yesterday? Of the Glaze 3D video card? The one you said was excellent?" J.H.H. then began to dangle a pendulum before Talbeese's face, and he began to speak in a slow, monotone voice. "Remember? You... thought... the Glaze3D... was... excellent..."

"YES MASTER, I WAS IMPRESSED BY THE RAW POWER OF THE GLAZE 3D. NVFXtreme TECHNOLOGY IS TRULY THE WAY OF THE FUTURE. ALL HAIL GLAZE 3D," Gary replied while lumbering out of his seat and slowly leaving the conference room. We were later told the Police found him crouched in a gutter of an alleyway, mumbling something about mummies and tanks.

This diagram is proof that there is actually a card. I mean, if there wasn't a video card, how could we have this diagram? Huh?
"Our upcoming line of 62 new Glaze Superion NFiniteFx Pro Quad Delta Mark VII cards will appease every single consumer out there, from the businessman to the hardcore gamer. They will also come in different colors to match your computer case. They will be able to detect what mood you're in, and change accordingly. Some of the cards will have 'Smart Technology', allowing them to determine if you're about to lose in a particular game, and then adjusting the difficulty accordingly. Our 'Performance Line' of cards will feature incredible 'Next Gen' special effects. Hate-tee-Hi may be promoting their inferior 'Full Screen Anti-aliasing" technology, but our cards will be the first to ever have '5-Mile Radius Anti-aliasing' technology. Not only will it smooth out the jaggies on your screen, but it will alias everybody else's computer games within the surrounding five miles. It will also support 'Force Feedback Viewing' and 'The Atomic Discombobulator'."

J.H.H. finished up the press conference by presenting the audience with a myriad of new charts, most of which had a few words we could understand. One of the charts was very glossy and had a lot of red in it. When a reporter attempted to ask J.H.H. when the projected release date for the Glaze3D cards was, he responded by vanishing in a cloud of smoke, cackling evilly..."
---------------------- :cool3:

[MuFu]

Here's my guess:

0.13u SOI @ ~700MHz
~150-200mill transistors
8 full FP pipelines, 1 TMU per pipe
8 VS units
Same ISA/constructs for PS/VS
DX9 + PS/VS 3.0 compliancy
Programmable tesselation unit
256-bit memory interface
2nd-gen GDDR2 @ ~700MHz.

That is not unrealistic at all. In fact, I doubt such a configuration would consume much more power than the 5900 Ultra.

MuFu.

Edit - yes, sorry... meant to say "150-200mill".

[Uttar]

Pretty sure you couldn't fit that much in 150M transistors, though.
The Inquirer hinted at 2FPU/pipeline, with 8 pipelines.

I think that might actually make sense.
The NV30 can do 1FP op or 2 TEX ops/clock/pipe.
So what if the NV40 could do 2FP ops or 2 TEX ops/clock/pipe ( or 1 FP & 1 TEX, of course ) - but on eight pipes?

That'd give us something that can do 16 PS operations in ALL situations. nVidia already managed to do COS/SIN at full speed, so all they'd need now is some dedicated hardware for LRP and they've got some truly always great performance.

Where to get the transistors for that, though?
My guess is that they're going to have a TBDR architecture and cheat in the VS. Let me explain.

They'd actually shift all non-transform related instructions to the PS automatically. So they can support all non-transform related instructions with little speed in the VS, and support transform-related instructions at very high speed - so you get the equivalent of 8 VS or 4 VS or something in most situations, but you'd get also need more PS power.

The point, however, is that using a TBDR, this would actually make sure you NEVER do any useless non-transform related instructions. So in practice, this might be a lot better than it sounds.

I don't know if we're gonna get that type of stuff in NV40. Might be more of a NV50 target. Don't know. But I believe it's certainly a way the industry could go for.
ATI tried to do a hybrid PS/VS architecture ( truly shared resources ) - it might have been a good thing. But with a TBDR, I believe you can do BETTER.

So, IMO, even if TBDRs don't have huge advantages theorically, you can do loads of very nifty things with them to save a lot of transistors in other places.
A good question, thus, is whether the NV40 is a TBDR or not. It's practically all we need to know, IMO.


Uttar

[MuFu]

I'm almost certain it isn't now due to the likely memory interface tech/config.

MuFu.