Question: Where does Nvidia go from here?

[Natoma]

Lets compare.

ATI R300:

0.15 Micron
325Mhz
310Mhz DDR-I (620Mhz *Effective*) [EDIT]Thanks T2K![/EDIT]

Nvidia NV30:

0.13 Micron
400-500Mhz
500Mhz DDR-II (1Ghz *Effective*)
HUGE PCB and Heat Sink/Fan Combo

Where does Nvidia have room to grow? They're currently using 0.13 Micron process, which means that it will be at least a year, maybe 2, before they can transition to 0.09 Micron. They're already using a huge ass heatsink/fan, and 500Mhz DDR-II, along with a 5V power supply connector. ATI is using a 3pin molex connector, which basically means they're drawing less power. Understandable given the lower clock speeds on the core and the memory.

The only solution I see for them is to widen the bus to 256-bit, which would make an already large chip that much larger, hotter, and power hungry. They could simply increase the speed of the memory, but really, how much can they push through a 128-bit pipe? They can add a second texture unit as well, but again the question remains, what about the bandwidth and the size of this already beastial-sized chip?

The rumors I've read around here have the R-350 still on a 0.15 Micron process, around 400Mhz core, with GDDR-3 (slightly modified DDR-II), and 8x2 (though most likely 8x1).

Even then, there's still the option of moving to 0.13 for ATI when TSMC's process is perfected (no doubt they learned a lot of lessons because of the Nvidia debacle).

So my question is, where does Nvidia have room to grow? 0.09 is not going to be around for a long long time, and with the chip as large as it is right now, it doesn't appear they could easily double the bus bandwidth.

I was talking with a friend of mine and I basically said that I think at this point in time, ATI's experience in the mobile sector is helping them dramatically. They transitioned from building for the low-power mobile sector to the high power desktop sector, while Nvidia transitioned from the high power deskop sector to the low-power mobile sector. And I think it shows in the difference between a 0.15 Micron R300 and a 0.13 Micron NV30.

Thoughts? Because this surely has me thinking that NV30 might be the ceiling for them unless they can make their designs more efficient, and quickly.

 

[T2k]

Just for the records...

ATI R300:

0.15 Micron
325Mhz
310Mhz DDR-I (620Mhz *Effective*)

 

[Natoma]

So uhm, anyone with any thoughts about this? I'd really like to know what people are pondering on this subject, if anything at all....

 

[RussSchultz]

Their first generation product on .13u is most likely not optimal in terms of power/performance, etc. either from architectural issues, or tool and process. All of these can be improved on the next go round to reduce power consumption and/or increase speed.

Some of it is in house skill and experience gained, other parts is TSMCs experience gained.


(Oh, and don't judge how much power is drawn from the power connector by the number of pins. Those powerlines outside your house only have 3 wires, ya know.)

 

[Laa-Yosh]

The only solution I see for them is to widen the bus to 256-bit, which would make an already large chip that much larger, hotter, and power hungry.

What's the relation between external bus width and chip size?
Sure, a new memory controller might add a few million more transistors, but as the process matures, they might be able to lower the voltage and thus maybe even increase the clock frequency as well. I don't see the jump to 256 bit that hard, but then again I might be wrong.

And once they have 256 bit, maybe they don't even have to grow for a while. Remember the GF256? It was indeed helped by replacing the SDR memory with DDR, without changing anything in the card. Now they're already high above the R300 in fill rate, snapping on more bandwith could let this chip wreak havoc... Whereas ATi has to raise the clock by 200 MHz just to catch up with that 4 Gigapixels of speed.

Also, Dave's interview shows that Nvidia is clearly behind schedule with this design. Even the Q4 2002 date was a fallback...
We don't know anything about the individual roles of the NV30 and the NV40, but the later might as well come sooner as expected - perhaps just in a year. In that case Nvidia does not need really need to change anything in the GFX...

 

[Joe DeFuria]

It seems that nVidia, for probably the next year or so, will likely be at the mercy of:

1) Availability of Faster DDR-II
2) Ability to "tweak" the NV30 design to wring more Mhz out of it.

What I mean by number 2, is that, for example, nVidia "tweaked" the N20 core to get the NV25, which allowed for higher clock rates (on the same process) and a refined memory interface offering better efficiency.

Given NV30's current deisgn and apparent heat and power issues, I share the opinion that it would be tough for them to implement a 256 bit bus with that core without a significant re-design to minimize power.

On the other hand, they could possibly reduce the clock rate of the NV30 and DDR-II memory to something like 400 Mhz?, and implement a 256 bit bus with that....

On the other hand ATI's challenge will be:

1) Fully exploit 0.13 microns to ramp up GPU clock speed to at least 450+ Mhz
2) Refine the memory interface to Utilize DDR-II memory efficiently.

By number two: there's a good chance that while R-300 supposedly can work with DDR-II, the memory interface may not be very optimal for it.

I'm not sure which direction will pan out to be better, but I do believe that ATI "controls more of its own destiny" so to speak.

 

[megadrive0088]

where does Nvidia go from here.... NV35 - 256-bit bus - 2 TMUs per pipe
(both maybes) more vertex shader pipelines - higher core and memory frequency. the GeForce FX 2 should be to the GeForce FX what the
GeForce2 GTS was to the first GeForce.


NV40 - Pixel Shader and Vertex Shader 3.0 or somewhat beyond
(like NV30 is somewhat beyond 2.0)

*can't even guess if NV40 will be DX10 or if that will be for NV50.

*start planning XGPU 2 to counter PS3 threat.

 

[megadrive0088]

Also, Dave's interview shows that Nvidia is clearly behind schedule with this design. Even the Q4 2002 date was a fallback...
We don't know anything about the individual roles of the NV30 and the NV40, but the later might as well come sooner as expected - perhaps just in a year. In that case Nvidia does not need really need to change anything in the GFX...

yeah NV30 should have been the spring 2002 product, since NV25 could have been a fall 2001 product. instead, Nvidia did a Ti500 version of the same NV20 core. NV25 moved to spring 2002. NV30 moved to fall 2002 BUT slipped again.

I'm hoping that NV35 and NV40 come alot sooner (.13u process wont be painful anymore) but they would not actually be early and I would not concider Nvidia rushing things, since NV35 should be about done, with NV40 following soon after IF Nvidia is not still behind.

NV35 *should* now be a spring 2003 product (even if Nvidia doesnt release it then) and NV40 the fall 2003 product.... of course, there is no way of knowing at this point.

 

[sc1]

NVIDIA's 0.13 micron process has more upside potential than
ATI's 0.15.

Difficulity of 0.13 micron process has to do with redesigning your
cell libraries so ATI can't avoid same problem. This effected
entire semiconductor industry not just NVDIA.

Basically either you invested for 0.13 processor for future growth
or stuck with fine tuning 0.15 existing processor which will
yield minor improvement. It's massive investment because
your engineers have to learn various effects of your design for 0.13
processor for first time. For example Intel had 2 entire design
team when they did transition to 0.13. While AMD investmed less
and paid for dearly.

So for next year, NVIDIA will refine 0.13 process and will generate
higher frequency, more likely reach 700M-1G Htz and more optimal
design while ATI have to decide whether to take pluge in 0.13 process.

Question is whether they have financial resources or not. My guess
is they will have hard to do it because they have spent money on
0.15 design

 

[Laa-Yosh]

Question is whether they have financial resources or not. My guess is they will have hard to do it because they have spent money on 0.15 design

Or maybe not. Don't forget that ATi has two design teams and the other one that created the Radeon 8500 has been working on the R400 for more than a year now. They could as well be doing their first tests with 0.13 processes at this moment... But they might as well face the same problems that Nvidia had to.

 

[demalion]

Well, I have similar thoughts (Natoma).

For answers, I think nVidia got burned by their traditional push for more expensive/advanced method of progress this cycle, and, if as smart as I think they are, the other "options" that have been alluded to will be their major focus. I think they do have a roadmap for up to nv35 along the "brute force" path, but I too don't see how they can manage it competitvely with the nv30 already taking the measures it does.

I expect these "options" to begin to come to light in nv40...but maybe they are smarter even than that and it will be in nv35 (but my current opinion is that they were not...the 2 slot solution simply doesn't strike me as "sane" in that regard, but maybe the manufacturing issues are behind that more than their design goals).

This all depends on how strong R400 (and the theoretical "R350"... "R" and "RV" ATI notations don't seem to remain distinct in the rumor mill, so it may not exist) showing is...ATI has indicated this part is a major departure (More their "real" nv30 equivalent in terms of the "Mojo" type of thing for ArtX), though that may just be hype (well, they don't do much hype of this nature, and it wasn't repeated often, if at all, so I think it is honest excitement).

My (completely unsupported, but not completely unfounded) guess is that what ArtX primarily brought to the R300 was the critical tweaks that allow it to (over)clock like it does and work on 0.15 micron (the "hand tweaking" referred to earlier), and not so much features and architecture that people assume (I think they give the prior ATI engineers too little credit), and that the R400 may be much "smarter" in the "options" it uses as I think nVidia will have to do, but if it isn't nVidia may have plenty of breathing room if nv40 is even though they seem to be pushing their ability to fit their performance goals into the consumer PC already.

To me, ATI has had expertise in low power consumption relative to performance in both board and chip design, and enhanced it with the ArtX acquisition, and I foresee mobile leadership for them for quite a while. Is there a company or some engineers that nVidia has or could reasonably acquire that would address their apparent deficiency in this regard?

It just boggles my mind they are at a 2 slot configuration without a clearly established performance leadership (the data currently indicating the performance leadership is very incomplete and slanted at this point, but it seems a reasonable assumption that things won't change significantly in the "light of day" and in a wide variety of real world tests...we'll see). It also seems crucial to them that the nv30 configured like that handles ATI's "beyond 9700 Pro" release very well, or that they soon achieve the same performance in a less demanding design/implementation (the nv30 really seems to have little "downward mobility" at this point compared to the R300, but maybe the nv31 addresses this effectively).

 

[demalion]

So for next year, NVIDIA will refine 0.13 process and will generate
higher frequency, more likely reach 700M-1G Htz and more optimal
design while ATI have to decide whether to take pluge in 0.13 process.

Question is whether they have financial resources or not. My guess
is they will have hard to do it because they have spent money on
0.15 design

Well, nVidia was retooling for 0.13 well before releasing the GF4, weren't they? So I don't see how releasing the 0.15 R300 would say anything about their next product.

Also, it is not as if they are NOT going to go to 0.13, so why would there be any further delay? Look at your first statement (I didn't quote it)...I'm pretty sure ATi has known that for a while.

I think all your comments on the challenges they might face doing so make sense, I just don't think your questions about whether they are facing these challenges already do...

EDIT: 2 comments that came to mind that might be pertinent...

I for one thing remember a comment by some ATI person, in conjunction with R400, strongly indicating 0.13 micron process. It may be an association with a conversation I had shortly after reading words from an ATI person concerning the R400, but I don't recall it that way. I think Ichy, if he browses through here, might be able to recall this better.

For another, which process is it that requires similar process engineering investment as another (larger) process such that the larger process could be skipped? To try to unconfuse that convoluted sentence: I read somewhere about some chip make considering (or achieving) skipping a process size and going to a smaller one. Is there a process size between 0.13 and 0.09 that simply got dropped by many? Does anyone recall what might be tickling my memory here...it could be only for one specific application that this was the case...

 

[Bigus Dickus]

My thoughts:

NV30 was probably targeted originally at ~400Mhz, perhaps 450MHz. This is reflected in the amount of bandwidth available - the NV30 is somewhat of an unbalanced design it would appear, similar to the GF2.

The 500MHz "Ultra" version with fusion powered heatsink is probably an attempt at securing the performance crown, which it probably will.

However, the Ultra part is likely what their summer refresh should have been, during the time between now and late next fall when they have an "improved core architecture" version of the NV30 (known as NV35).

I don't see the NV30 being pushed much higher than 500Mhz. It already has external power and one monstrous cooling system.

So, I think they blew their "refresh" wad right at the beginning, and will be stuck there for most of next year, until the NV35 shows up (with perhaps a 256-bit memory controller, improved AA and AF, core tweaks for a bit more clock speed, etc.).

Bottom line: ATi knows what they are facing for most of 2003. If they can lauch an R350 refresh, or even just a clocked bumped R300 that retakes the performance crown (a big if), then they'll likely keep it until late next year.

 

[Geeforcer]

I have doubts whether we will see 2TMU/pipe design from either Nvidia or ATI. The incredibly dubious rumor that R350 (if it exists at all, which still seems unlikely to me) will be 8x2 stemmed from the fact that NV30 was rumored to be 8x2 and "R350 would be the obvious answer, blah-blah...". I think now that NV30 is out, the people who were propagating that rumor will quite down.

Secondly, I don't see ATi doing any major revisions to R300 in the near future, in other words, no R350. On the card level, I think that a 350-400MHz R300 card with 256 of RAM (Be that higher-clocked DDRI, DDR2 or GDDR3) will be releases this spring, possibly under the name of Radeon 9900-10000. Goind into the summer, ATi will bring out RV350 (.13 version of R300 with some tweaks) alongside R400 for the late summer launch.

On the other side, considering that delays effectively made NV30 a 2003 product, I think it will be at the very least 6 month from the day of general availability that we will see any new high-end chips from Nvidia. (and by new I don't mean NV30 with different clockspeed).

 

[megadrive0088]

looking at R300 and now NV30, I am very anxious to see the rumored R350 and eventual NV35. It would be a HUGE mistake IMHO if Nvidia waits until 2004 to get NV35 out.

 

[antlers]

Perhaps the NV30 is just (from nVidia's internal perspective) a dog; rendered ugly and late by an overly-ambitious design and the process troubles incurred in its development (not to suggest the the NV30 is not a very good, competitive part, but I'm sure it's overstayed its welcome in the eyes of nVidia's engineers). Market realities have forced them to put lipstick on it and push it out the door, but the 0.13/DX9+ chip they really want to release--with a much more elegant solution to all the problems they encountered in the NV30--is still in development.

If this were the case, they'd have no problems going forward. nVidia would be heaving a big sigh of release that the NV30 was finally off their backs so they can focus on their more-favored solution, which would be 6 or 9 months off.

 

[BobbleHead]

The only solution I see for them is to widen the bus to 256-bit, which would make an already large chip that much larger, hotter, and power hungry.

What's the relation between external bus width and chip size?
Sure, a new memory controller might add a few million more transistors, but as the process matures, they might be able to lower the voltage and thus maybe even increase the clock frequency as well. I don't see the jump to 256 bit that hard, but then again I might be wrong.

There is quite a relation for both the die and package size. When you design a high speed package for a chip, you generally decide that you want/need a specific ratio between signal and power/ground balls. Right now the biggest interface on the nv30 is the memory interface. 4x32 bit channels, each channel has 32 data bits, 4 mask bits, 8 strobe bits, 14 or 15 address bits, 6 control bits. That adds up to 260 signal balls. Let's say they decide to go to 8x32 bits. That means they now have an additional 260 signal balls to worry about. This is non-trivial. You can't just replace 260 of your power/ground connections with signals, because you'll completely wreck your ratio. Ideally you add 260 new balls, along with X power/ground balls to keep your ratio correct. You've now blown up your package size by a significant amount. You face a similar problem on your die. You need to connect the signals on the package to locations on the back of the die (in flip-chip. In wirebond all connections are around the edge of the die). You do the same sort of rationing, doling out spots for various signals, power and ground connections. Those all have to be certain sizes to make the proper connection between die and package. A good design makes the area of the actual chip and the area of the die-package connections about equal, because if you go too far from that, you're wasting space in one manner or another. Again, toss in an additional 260 signal locations, and more power/ground locations and you need to account for them.

It's well within Nvidia's ability to make the change, but I don't think I'd say that it was "not that hard."

 

[RussSchultz]

Their flip chip already has ~1000 balls, so maybe they're anticipating moving to a wider bus?

 

[sc1]

Well, nVidia was retooling for 0.13 well before releasing the GF4, weren't they? So I don't see how releasing the 0.15 R300 would say anything about their next product.

Also, it is not as if they are NOT going to go to 0.13, so why would there be any further delay? Look at your first statement (I didn't quote it)...I'm pretty sure ATi has known that for a while.

quote]

Hugh? GF4 was comparably small design effort (tweaking existing
GF3 core). Geforce FX is hugh desing challenge. Until you have
taken this route you can't proceed to next generation of chip.

Because ATI have spent bult of their effort on 9700 they have not
enough resources to concentrate 0.13 design. Unless ATI
engineer are 2x fast, it's unlikely come out with 0.13 product
until 2003Q4 or 20041Q. Best they can do is tweak 9700 unlikely
yield same improvement as GeforceFx

 

[Dave Baumann]

Actually, I glibbly asked Alan Tiquet (NV) when 0.09 was on the cards (pardon the pun), and he pointed out that 0.11 was next...