Clock speeds/R300 vs. NV30

[Steelwire]

Hello 1 n all...

ATI produced this monster on .15 micron - fair enough - however software that uses all of these features won't be around for quite a while yet.

My suspicion is that ATI are going to go for a clock speed ramping - they may not get a new architecture out for a year or so (bar mods to the current one e.g. new bus technologies, DDRII support perhaps even subtle changes in the chip itself like caching, extra pipelines etc.)

Seeing as most coding is going to move over to HLL which are compiled at run time (well initialisation time) do ATI really need to add new features. Let's say that they are friendly with a company who could (in a years time) crank out 0.09 micron chips cheaply - the R300 would be significantly smaller and clock at maybe 800/900 Mhz - may be faster.... who knows....

If they do this (and of course update the RAMs on their cards) would they really need to change their architecture - I mean they could double the number of pipes by this time next year but what if they could double the card frequency (a process done by the silicon boys) and let the rest of the engineers have 18months or 2 years to generate the next greatest chip on earth.... Hmmmm... just pondering....

 

[Nappe1]

I had a more than a page of my thoughts of future, but this damn XP decided reboot.

anyways, Steel Wire, your topic isn't really interesting because it has nothing new to offer. Everyone knows here these. And one more thing... IMO, you estimation about 0.09µm availbility and and clock rates reachable with that process are a way too optimistic.

But, anyways, welcome to boards.

I'll write my thoughts later... I know that at least nAo has been used to be interested about them.

 

[Steelwire]

your topic isn't really interesting because it has nothing new to offer.

Oh... really... thank u! ;-> actually I been here ages but used to have a different and censorable name b4! I been about a while - even worked for a graphics chip firm.....

I just think we could get into battle similar to AMD vs. Intel... which is essentially the age old optimisation problems of speed vs. size.... BTW What I meant was that Intel and ATI have been very, very cozy recently.

I also don't see 600Mhz+ as a problem - honest - if a CPU can take a 10 times clock increase over 4/5 years then a GPU can take a *2, *3 over a year or two... As these things are now quite programmable a lot of the grunt work of optimisation nowe resides in the compiler.... however.... It didn't much help Itanium.

 

[alexsok]

Again, just for the sake of this thread (u can consider the information whatever u want, a rumour, truth, whatever).

From the information I know, NV30 will most likely be clocked at 400mhz (core) and the memory will be 400/450+ DDRII depending on the model, although 480mhz is likely to be the speed for the top of the line part.

I previously reported this already, but as I said, just for the sake of this thread!

 

[Tahir2]

Just want to add that ATI didn't ramp the Radeon 8500's clockspeed up to compete more favourable with the GF4 Ti but this time I believe tht they will. And I dont believe they need to go to .09nm technology just yet to do it. They will probably get some extra headroom when (if) they switch to .13nm technology.

However again, ATI didnt do this in the past (ramp up a current part on a smaller process) so they might not do it again and wait another 6-9 months or so to release their future card... maybe a VS3.0 and PS3.0 Radeon 9700 with DDR II.

Just adding my thoughts 'for the sake of this thread'.

 

[pascal]

Lets speculate a little. See this .09 micron process from TSMC: http://www.tsmc.com/english/technology/t0113.htm

...With 70-75% linear shrinkage and a two-times performance improvement, compared to TSMC's 0.13-Micron technology...

This .09 micron process is very promissing and will have 300mm wafers.

So the migration from .15 to .09 means:
- .25 (.13 -> .09) x .80 (.15 -> .13) = .2 or 80% of die size reduction
- 2 (.13 -> .09) x 1.3 (.15 -> .13) = 2.6 performance improvment

Now with this data we can imagine a future chip with:
- 850MHz core
- 525Millions transistors
- With the same die size and generating the same heat

My guess this is why JC is forecasting a new quantum leap in performance improvment during the next 2~4 years.

Probably we will see more and more integrated solutions in the mobo for the future. Something like:
- .09 micron chip 150Millions transistors
- multiple clocks speeds inside the chip
-Integrated NB and SB
-Integrated DD sound
-lan, serial ata, peripheral controller
-Integrated 2D/3D DX9 core (800MHz, 8 fp pipelines, 4vs pipelines, large caches)
-UMA controller

 

[RussSchultz]

I don't think your numbers work out quite right:

.15 to .09 produces a linear shrink to 60% of the original in one direction (.09/.15), meaning a area shrink to 36% of the original .15 size(.09/.15)^2.

The speed follows the same pattern, but in the opposite sense. (I think your 2.6x is about right)

The power will go down much more than you're suggesting, since the Vdd is (rule of thumb) the micron size * 10, meaning .9V will be the operating voltage, rather than 1.2 or 1.5. So, the power required will be closer to 36%-50% of the original.

Yes, smaller is always better. I doubt, however, we'll see the NV35 or R350 on .09. TSMC is just now sampling and its going to be an even more heinous uphill battle than .13 is as more of deep submicron effects come into play in layout as we go smaller. Tools will not be cheap, and neither will masks.

 

[Dio]

Smaller minimum feature size does not necessarily mean smaller overall transistor size, smaller gate size, smaller sram cell size etc. - or at least, the changes are not necessarily as linear as is presented here.

Faster and smaller designs require more space to be added for noise insulation etc. in some circumstances. This is why chips usually require at least some minor layout tweaks for smaller processes.

 

[pascal]

I don't think your numbers work out quite right:

.15 to .09 produces a linear shrink to 60% of the original in one direction (.09/.15), meaning a area shrink to 36% of the original .15 size(.09/.15)^2.

I understand and agree with what you are doing, but lets recalculate using the TSMC data:
- The new technology (.13 micron) will reduce die size by more than 20 percent compared to .15 micron.
- 70-75% linear shrinkage from .13 micron to .09 micron
- Then .7 (.13 -> .09) x .8 (.15 -> .13) = .56 or 44% die size reduction

It was my fault in my first calculation

The speed follows the same pattern, but in the opposite sense. (I think your 2.6x is about right) .

The speed improvment is exciting, inst?

The power will go down much more than you're suggesting, since the Vdd is (rule of thumb) the micron size * 10, meaning .9V will be the operating voltage, rather than 1.2 or 1.5. So, the power required will be closer to 36%-50% of the original.

The TSMC vcore are 1v and 1.2v then I expect maybe 36% when comparing 1.5v chips with 1.2v chips.

The R300 with the .09 micron could have:
-800Mhz core
-44% smaller die
-36% less heat
Looks like a mainstream chip

Yes, smaller is always better. I doubt, however, we'll see the NV35 or R350 on .09. TSMC is just now sampling and its going to be an even more heinous uphill battle than .13 is as more of deep submicron effects come into play in layout as we go smaller. Tools will not be cheap, and neither will masks.

I dont expect to see any new chip using the .09 micron during 2003.

 

[Fuz]

Who wants to bet that one of the big CPU manufacturing companies will buy one the big boys of the graphics industry, call it a merger if you like, within 5 years?

 

[Dave Baumann]

Well. NVIDIA buy AMD, then Intel buys ATI to compete. If anything that radical happened in the marketplace then thats what I'd see happening.

 

[Fuz]

Well, if you look at whats happening in the industry, it points very strongly towards a merger with a CPU company and graphics company.

Correct me if I am wrong, but the graphics processors are fast becoming more and more like CPU's. Eventually they will become the same thing.

 

[pascal]

At the same time it is hard to justify a +1GHz CPU only for internet browsing and simple office applications (the majority of the target market).

 

[Gunhead]

A CPU + GPU maker? I thought VIA was that already (C3 + ex-S3 stuff)...

 

[KimB]

VIA sucks, and Cyrix always sucked. I don't see any possible way for them to compete.

 

[Johnny Rotten]

Well. NVIDIA buy AMD, then Intel buys ATI to compete. If anything that radical happened in the marketplace then thats what I'd see happening.

Oh dear, what would DoomTrooper do then??

 

[Doomtrooper]

 

[Mummy]

At the same time it is hard to justify a +1GHz CPU only for internet browsing and simple office applications (the majority of the target market).

Dont worry, M$ will have in a "distant" future Longhorn + they are pushing .NET crap now, they will add thousands of features we dont need at all, your 1GHZ cpu will start to be slow even for GUI and simple applications, im sure

 

[pascal]

Now I have to worry
I imagine layer above layer of useless software doing nothing and adding no real value to word processing, e-mail and browsing.

 

[Nite_Hawk]

pascal:

Ah, but don't fret! Think of all the "security" enhancing features to protect you from rogue hackers and viruses! It will also conveniently make sure that you don't "inadvertently" use your computer to do anything the mpaa or riaa wouldn't like. Nice to see our processing power put to good use.



Nite_Hawk