R4XX to be in the HUNDREDS of millions of transistors

[Geforce4ti4200]

Originally posted by Nebuchadnezzar
And that's for a good reason. :rolleyes: :rolleyes:


Even if he heard it from his friend who works at ati, it still is not finalized. The best chance itll be 12x1. Ill be supprised if its 16x1 or dissapointed if they go backwards with 6x2, why anyone would speculate it to be 6x2 befuddles me as this would be crippling the core. Maybe the mainstream may be 6x2 but the performance cards should be 12x1 or even 16x1 :)

[Ninja Prime]

Geforce4ti4200, I think you are still missing it. It likely won't have normal "pipes" like older cards. It will use it's VS and PS power to do the operations normally done in the "pipelines" on older cards. Thus, it will be able to do a certain amount of pixel drawing operations and texel drawing operations per clock, based on its FP power. That's sorta how it works, but I'm not the best a explaining these things...

[rwolf]

Originally posted by lukar
For long time I'v got full spec for loci, but nobody believes me. The spec of loci will be

0.13 500Mhz design 1TMU 16Pipes, 256bit mem, 128/256 DDR II 500Mhz
~180millions trans, full support for dx 9.1 and open gl 2.0
Smartshader 3.0, AA 8x mode will be introduced
Filtering remains the same 2x 4x 8x 16x performance/quality
Trueform 3. Now fully suppot by hardware and also support for displacement mapping found in Matrox cards.
Videshader II
HyperZIV compression
The card will not need additional power, It will support AGP 4x, 8x. Also the card will support PCI expres, but those cards will not be shipped yet. There will be no adapter for agp 8x/ PCI express.

Actually this sounds reasonable. Think of it this way the 9600 has 75M transistors and the 9800 has 35M more to implement 4 more pipes. If you take 75M transistors (4 pipes) and add 35M * 3 (12 extra pipes 105M transistors) you get 180M transistors. 40M transistors would be used for cache and control logic etc.

On the other hand there may be more cache and more transistors for PS3.0 and the extra pipes so Loci may only be 12x1.

[Dazz]

It a ahigh chance that it will be Xx1 based, why because it's more efficent at doing shaders, unlike nVIDIA 4x1, well they say 8x1 but the performance goes right in line of 4x1.

[Viral]

Originally posted by Geforce4ti4200
Even if he heard it from his friend who works at ati, it still is not finalized. The best chance itll be 12x1. Ill be supprised if its 16x1 or dissapointed if they go backwards with 6x2, why anyone would speculate it to be 6x2 befuddles me as this would be crippling the core. Maybe the mainstream may be 6x2 but the performance cards should be 12x1 or even 16x1 :)

Actually, 6x1 seems like a nice number for new mainstream cards.. perhaps the RV360 replacement will implement such a design.

[Uttar]

Originally posted by Dazz
It a ahigh chance that it will be Xx1 based, why because it's more efficent at doing shaders, unlike nVIDIA 4x1, well they say 8x1 but the performance goes right in line of 4x1.

Hmm, the NV30 is a 4x2... Not a 4x1.


Uttar

[rwolf]

I think the most important thing to do would be to make the solution work with the current drivers. Look how aweful Nvidia's drivers were when the NV30 came out. If you keep the hardware the same with minor differences you can reuse all the work you spend writing, tuning, and testing the drivers.

[eesa]

Originally posted by lukar
For long time I'v got full spec for loci, but nobody believes me. The spec of loci will be

0.13 500Mhz design 1TMU 16Pipes, 256bit mem, 128/256 DDR II 500Mhz
~180millions trans, full support for dx 9.1 and open gl 2.0
Smartshader 3.0, AA 8x mode will be introduced
Filtering remains the same 2x 4x 8x 16x performance/quality
Trueform 3. Now fully suppot by hardware and also support for displacement mapping found in Matrox cards.
Videshader II
HyperZIV compression
The card will not need additional power, It will support AGP 4x, 8x. Also the card will support PCI expres, but those cards will not be shipped yet. There will be no adapter for agp 8x/ PCI express.

interesting... will you just settle for "this card will be amazing? :)"
I'm actually pretty excited about pci express - It's no big secret that pci express vid cards are planned for next year but what about all my other pci cards? Are they at the same time release pci express versions of sound cards... etc? Or will the mobos have both like pci/isa combo of yesteryear? Can pci and pciexpress exist on the same bus?

[B&R]

200 mil wow!!

OT:

Today in my univ i was jus learning abt the first microprocessor with 1000 transistor in 1970's and to know we are at 200 mil is jus amazing. :afro2:

Jus wanted to share.:afro:

[GlowStick]

Originally posted by B&R
200 mil wow!!

OT:

Today in my univ i was jus learning abt the first microprocessor with 1000 transistor in 1970's and to know we are at 200 mil is jus amazing. :afro2:

Jus wanted to share.:afro: Well, CPU's are actually alot less transistors, but GPU's just havealot.

[nutball]

Originally posted by GlowStick
Well, CPU's are actually alot less transistors, but GPU's just havealot.

Depends if you're counting the cache in the statistics doesn't it?

[GlowStick]

Originally posted by nutball
Depends if you're counting the cache in the statistics doesn't it? Im not sure, nvidia allways did boast that the nv30 dose have more transistors than a p4 i belive.

[nutball]

Originally posted by GlowStick
Im not sure, nvidia allways did boast that the nv30 dose have more transistors than a p4 i belive.

I wouldn't be surprised if NV20 had more transistors than the P4 core.

Itanium2 is something like 221M transistors I think, but the vast majority (two-thirds) are cache, not core.

[Uttar]

Talking of cache, the NV30 got a lot more cache than the R300.


Uttar

[Ninja Prime]

Talking of cache, the NV30 got a lot more cache than the R300.

Too bad it doesn't help it any lol. :D

[eesa]

CPUs are _way_ more complex than gpu's. You can't even compare the two. Only reason gpus have so many transistors is because there's so much repetition of the various units in the pipeline.

[GlowStick]

Originally posted by eesa
CPUs are _way_ more complex than gpu's. You can't even compare the two. Only reason gpus have so many transistors is because there's so much repetition of the various units in the pipeline. I disagree, current GPU espically the new ones comming out are very very complex. Ask your self this, if the gpu's are less complex, and they are built on a smaller process than the Orginal P4, why dont we see them clocked at speeds like 1.8ghz?

[aaronspink]

Originally posted by GlowStick
I disagree, current GPU espically the new ones comming out are very very complex. Ask your self this, if the gpu's are less complex, and they are built on a smaller process than the Orginal P4, why dont we see them clocked at speeds like 1.8ghz?

Because their design methodology isn't even in the same performance league as most microprocessor designs. I'd be very supprised it something like the NV3x or R3x0 had any full custom design outside of of ram structures. In addition, the NV3x and R3x0 are roughly 50-60 FO4( fanout of 4 inverter) designs where as most cpus are in the range of 8-20 FO4. Additionally most microprocessors are using advanced sequential designs ( high performance flipflops and latches, aka pulse latches and sense amp latches ) where most synthesis based designs like NV3x and R3x0 are using bog standard master-slave flipflops.

And yes, the GPUs are simpler than CPUs. They are pretty much just datapaths. They don't have to worry about a LOT of compexity that are required to have high performance general purpose cpus ( cache coherence, multiple memory models, fast branching, dynamic resource allocation, OoO logic, advanced scheduling, etc). It is simple to design a simple CPU. Most EE/CE students will design and debug a simple 5 stage risc in one semester. A high performance cpu on the other hand will take 4-5 years to design and test with teams in range of 400-1000 people.

To give you an idea of the compexity and scale of a modern high performance cpu, both amd and intel have teams working on single designs that are as large if not larger than the total number of people in all of ATI's or Nvidia's design teams.

If they were simple, they wouldn't need that many people.


Aaron Spink
speaking for myself inc

[rwolf]

Originally posted by Uttar
Talking of cache, the NV30 got a lot more cache than the R300.


Uttar

How do you know this to be true?

I think the secret is to design for latency while keeping the chip busy. Cache has its uses, but has limited usefulness on a video card.

[Carbon Unit]

You guys are all behind the times, the final spec on the NV 40 is 175M transistors built buy IBM (flip chip) 600/1500 43gb of bandwith:eek:

[rwolf]

Yeah. Like you know what the final clock speed is going to be when the chip hasn't even taped out yet.