90nm what it means to next years gpu's

[borntosoul]

what can we expect when ati and nvidia reach 90nm process?
i think it will be a turning point and make one of the biggest jumps in performance since the 9700. also do you think that we have hit an iq wall and the next gen cards will be focusing more on pure speed?

[MUYA]

erm maybe towards the mid of next year or q1 2006 when both IHVs willl launch new GPUS , NV50 and r500 maybe?

These new GPU may use 90nm which allows for lower voltages and lower power consumption, which means possibly higher clocks and of course more complex GPUS due to more trannies being able to crammed in per area. There could also be a problem with yields as process get smaller and comlexity rises.

IQ wall? nah they are just beggining to fall down...more AA, new AA techniques? more AF? Definately more Pixel shading capabilties in a more efficient manner. As well as Vertex shading....maybe PS and VS will integrate..i think according to some as DirextX next?

performance jump? I'd say that the next new gen GPUS will also bring a significant jump in performance as well. By then gaming with 4X AA and 8XAF will be standard at 1600 by 1200. and those GPUs will bring Doom III, Far Cry and the unrealsed yet Half Life2 to its knees.

[coldpower27]

Hopefully we can reach 90nm for GPU soon, though yeah late 2005 timeframe to early 2006 timeframe might be a good time for ATI and NV to start experimenting with this new technology. Hopefully by then 110nm will be a very mature technology.

It will be interesting what NV50 and R520 could bring but that is very very far off from now.

[borntosoul]

there will be a big push for both ati and nvidia to release chips on 90nm before christmass 2005 - thats my prediction! wont be too difficult (thats a full year and bit after AMD and over a year and a half after intel).

[MUYA]

GPU complexity is much more than of CPUs . The GPUs will be designed with whicever fab's 90nm process in mind but...they will be either employing TSMC or IBM (for NV) still solving yeild problems when adopting 90nm process will for most of the time not that easy..

[Riptide]

I bet we don't see another major refresh until 2006.

[borntosoul]

i heard its more difficult to design cpu's than gpu's.

[retsam]

i heard its more difficult to design cpu's than gpu's. thats what i heard as well... cpu's have to worry about backwords compatibility. gpu's can just send those calls that arent supported in hardware onto the cpu.

[jolle]

GPUs are atleast alot more complex, pack a whole lot more transistors and pack a whole lot more power..

NV40 = 220 million transistors, 100 gigaflops (I think)
P4 3.2Ghz = 55 million transistors, 59 Gigaflops (what i got when googling, not sure, think that includes SSE and all the SIMD stuff)


I think MUYA is on the right track, eventually (seemingly with DX Next") the VS and PS units will prolly merge into 1 powerful unit that can do both, which is prolly a good idea cause then the VS units cant bottleneck the PS units (or any other part of the card) and viceversa, it can just work fullspeed with any combo of "VS" and "PS" load put on it..
But more raw power, perhaps refined shaderlanguages or something, evolved AA and filtering techniques sounds reasonable..

[Viral]

NV50/RV500

And somehow, we may see cooling solutions get even larger ;\

[borntosoul]

GPUs are atleast alot more complex, pack a whole lot more transistors and pack a whole lot more power..

NV40 = 220 million transistors, 100 gigaflops (I think)
P4 3.2Ghz = 55 million transistors, 59 Gigaflops (what i got when googling, not sure, think that includes SSE and all the SIMD stuff)


I think MUYA is on the right track, eventually (seemingly with DX Next") the VS and PS units will prolly merge into 1 powerful unit that can do both, which is prolly a good idea cause then the VS units cant bottleneck the PS units (or any other part of the card) and viceversa, it can just work fullspeed with any combo of "VS" and "PS" load put on it..
But more raw power, perhaps refined shaderlanguages or something, evolved AA and filtering techniques sounds reasonable..

dont get sucked in to the marketing hype, if cpu's were so easy to make compared to gpu's than nvidia and ati would be in the cpu business!

[jolle]

I didnt say CPUs were easy to make, and I didnt say they were easier then GPUs to make either..
I said they are more complex and pack more raw power.

and how is any of that marketing hype?
had to google like a madman to find it all..

[MUYA]

CPUs and GPUs are apples to oranges to compare complexity as each do very very different tasks!!! There was once an article somewhere pr sp,ething that precisily described this. Arstechnica maybe not sure.

[jolle]

ya, its a bit like RISC vs CISC isnt it?
GPUs resemble RISC abit, limited but dedicated, while CPUs has to be more versitile, and as mentioned compatible.

[Bill The Cat]

I think MUYA is right - Ars had an article some time back on the topic.

As far as 90 nm goes, I won't be holding my breath. Everyone seems to have trouble fabricating silicon with features that close, and the heat from gate leakage is enormous. 110 is probably the farthest they dare venture for some time, I would guess.

[bkswaney]

I think MUYA is right - Ars had an article some time back on the topic.

As far as 90 nm goes, I won't be holding my breath. Everyone seems to have trouble fabricating silicon with features that close, and the heat from gate leakage is enormous. 110 is probably the farthest they dare venture for some time, I would guess.

I agree.... 110nm will be the build size for 05 and maybe halfway
through 06.

The first 9's will be low end parts. Maybe by christmas of 06
the 9nm will be ready for prime time. ;)
We are almost into 05 now and 13 is still the king of
the high end parts.

[borntosoul]

AMD dont seem to be having much trouble with 90nm so we'll see how ati and nv do by end of next year, i guess time will tell.

[bkswaney]

AMD dont seem to be having much trouble with 90nm so we'll see how ati and nv do by end of next year, i guess time will tell.

But intel and amd have been gearing up on it for some time now.
Plus they are not packing 222mil in trans into them.
Less than half that.

[borntosoul]

thats true, but we are talking over a year later, i know im being a little optimistic:)

[MUYA]

You have remember that the die sizes for GPUs almost doubles that of CPUs. SO the sheer complexity of the dies will make yeilds a lil less favourable than GPUs.

[Karma]

I agree.... 110nm will be the build size for 05 and maybe halfway
through 06.

The first 9's will be low end parts. Maybe by christmas of 06
the 9nm will be ready for prime time. ;)
We are almost into 05 now and 13 is still the king of
the high end parts.

Remember that ATi has to move to a 90nm process faster than nVidia does, not for desktop video cards but for the Xbox2 R500. So ATi will have a 90nm chip in 2005, although maybe not for desktops. Maybe R520.They both have certain hardware similarities but from what I've known the R520'll be 110nm, right?

[Bill The Cat]

AMD dont seem to be having much trouble with 90nm so we'll see how ati and nv do by end of next year, i guess time will tell.

AMD are having a huge problem with 90 nm. Their high end FX is built on 130nm, not 90nm as it would be if all were well. The new S939 chips are only sporting 90 nm on the low end, no high end 90 nms are in sight. AMD isn't running a magic show here at all, they're having the same trouble the entire industry is having.

[borntosoul]

it makes more sence to use 90nm on their mid/lower range cpu's anyway, gives them time to iron out any glitches and the ruduced die size will mean more profit for them on the chips they sell most of. i thinks its more sensible for them to do it that way- doesnt mean they are having huge problems with it.
also why move the fx 55 to 90nm when they can make it on 130nm?

[coldpower27]

From what we can see AMD's 90nm transition isn't any more smooth then Intel's was. They are releasing the lower clocked versions first 1.8GHz - 2.2GHZ on 90nm process and using the mature 130nm process for the higher clocked parts.

Intel did this with Prescott, yields were and still are excellent on their 2.8 - 3.2GHZ parts, they released a Northwood 3.4C, on a more mature process first then replace it with 3.4E and 3.6 as the 90nm process matures.

The only really smooth transition to 90nm process to me was the Pentium M Dothan which was released with 1.8 and 2.0GHZ models first, which surpassed the previous process clock rates, and later the other parts, 1.5, 1.6, 1.7GHZ were transitioned to the process as well.

We see ATI and Nvidia do the exact same thing, ATI tries out 110nm on the X300 Series, then they move it up a notch to the X700 Series a mainstream part, and finally to complete the puzzle they should move it to the enthusiast part the R480 hopefully.

Nvidia is doing it a bit different they are trying it out on the 6600 Mainstream series first, but it is unknown which core will get it next, will it be the value core or the ethusiast core, there are reason for either transistion.

[JoshMST]

Wow, I think we are actually on the "apples and oranges and plums" now with this discussion!

I personally think that we will see limited 90 nm production in the first half of 2005 from TSMC, but it may not be terribly viable as a mainstream source of production. I also do not think that TSMC's 90 nm process will be anywhere near the performance of Intel or AMD's processes. However, it will be a bump in overall transistor performance over 130 nm Low-K and 110 nm (and if there is ever a low-k varient of that process). Now, in GPU's running at 500 to 600 MHz, transistor leakage may not be as huge a problem as it is with Intel's Prescott. Even though these companies are using standard cells based on whatever process is being used, there will be mechanisms built into the process and the cell design that will help to combat leakage.

So, I figure we will hear about initial production of 90 nm R500's early in 2005, and I wouldn't be surprised if the NV50 and R520 are both 90 nm products. But, TSMC has yet to produce a complex ASIC on their 90 nm process (which I think is in testing and actually throwing out test wafers).

Now, AMD is having fewer problems with their initial 90 nm production than Intel did. AMD was much more conservative with how they achieved their results. First off they optimized the process for strong yields and low power. Since the 130 nm process still has another speed grade in it, they felt that they can hone their process by achieving strong yields with good power charactersitics. Now, if AMD had their back to the wall, and absolutely had to go with a 90 nm A64 at high speeds, they could have probably achieved it... but at a severe cost in both yields and power consumption. Process technology is a lot like alchemy, they are always fooling with the mixture to try to achieve better results. In this case AMD thought to achieve solid yeilds for a new process, as well as lower the power, so they tailored their 90 nm process at this time to achieve those goals. Now, as they continue to produce 90 nm A64's, they will tweak the process to achieve better yields AND speed bins. This is why the next speed grade for the A64's on 90 nm is scheduled for January. This gives AMD a lot of time to really work on their process, make it fast, yet keep it economical.

Intel was forced from the get go to get the Prescott up to speed with the Northwood core, and to achieve that they had to really push the design and the process. We all know the results. Intel's 90 nm process is very strong, but the weakness is the design of the Prescott. Intel really pushed the clockspeed, and their overall transistor layout and design reflects that. It also means that the faster one wants a transistor to go, the more power they have to apply to it. To get a design like the Prescott to go 3.6 GHz, a lot of power has to be applied! Dothan is a great example of a more balanced design that is tuned for high IPC, yet low power consumption. Dothan is a great processor, and it shows off the strengths of Intel's 90 nm process.

As for CPU's and GPU's, this was explained to me very quickly and easily. CPU's are single pipeline designs optimized for single thread performance and speed, and they are backed up by large on chip caches and a slow external memory. GPU's are multi-pipeline designs that can do a lot of parallel work each clock, smaller ondie caches are utilized, while a fast external memory subsystem is used. So basically GPU's are already dual core (multi-core), and have been since the TnT days. Until recently those pipelines were very specialized, while a CPU is much more general purpose and "programmable".

Clear as mud, eh?