ATI's next gen GPU in next half of 2010

[Super XP]

Originally Posted by Yukikaze

The point was: Never has been =\= Never will be and they have the resources needed to make it happen if they really want to.

You hit the nail on the head, Intel has the money and R&D to do anything they want with a GPU based design. Personally I think they do have a discrete GPU on paper, in design phase and will most likely get released sometime in 2012. It's going to be a competitors nightmare IMO

[wahdangun]

Originally Posted by _33

I've got an article here stating (back in late 2008) that Nvidia has SIMD shaders. Could you point me to something tangible that explains how or when Nvidia got MIMD?
http://perilsofparallel.blogspot.com...d-vs-simd.html

There was a rumour that Nvidia at one point had the GT300 gpu which had MIMD shaders (back in 2009), but it didn't see light of day. That new gpu is still not taped out and is possibly what is called fermi 2. So, ATI and Nvidia are at the same point with MIMD shaders, they still have those technologies for their gpus in the backburner.

http://www.motherboardpoint.com/nvid...r-t190998.html

http://www.nvidia.com/object/GTX_400_architecture.html

it's in GTX 480(just look wizz review), and btw GT300 was meant to be OEM only and it's just a bunch of renamed chip so nothing special about it

[KainXS]

Originally Posted by _33

I've got an article here stating (back in late 2008) that Nvidia has SIMD shaders. Could you point me to something tangible that explains how or when Nvidia got MIMD?
http://perilsofparallel.blogspot.com...d-vs-simd.html

There was a rumour that Nvidia at one point had the GT300 gpu which had MIMD shaders (back in 2009), but it didn't see light of day. That new gpu is still not taped out and is possibly what is called fermi 2. So, ATI and Nvidia are at the same point with MIMD shaders, they still have those technologies for their gpus in the backburner.

http://www.motherboardpoint.com/nvid...r-t190998.html

http://www.nvidia.com/object/GTX_400_architecture.html

???

from what I remember, the GT300 series is what the GF100 was supposed to fall in(or what we originally speculated it would), nvidia changed from GT300 to GT400 series to more than likely prevent confusion with the already rebranded names in the GT300 series(currently all OEM)

but bit-tech also says that the GTX480 is based on a MIMD design(maybe they're wrong???)

http://www.bit-tech.net/hardware/201...1-5gb-review/3

[eidairaman1]

TBH I don't think R6800 is going to be out anytime soon.

[_33]

Originally Posted by wahdangun

it's in GTX 480(just look wizz review), and btw GT300 was meant to be OEM only and it's just a bunch of renamed chip so nothing special about it

Well how come there are no mentions of MIMD in any Nvidia white paper anywhere ? I'd suppose Nvidia would be the first to trumpet about their MIMD shaders? Since Nvidia don't advertise them, then I suppose their shaders are SIMD or some variant (SIMT).

Originally Posted by KainXS

???

from what I remember, the GT300 series is what the GF100 was supposed to fall in(or what we originally speculated it would), nvidia changed from GT300 to GT400 series to more than likely prevent confusion with the already rebranded names in the GT300 series(currently all OEM)

but bit-tech also says that the GTX480 is based on a MIMD design(maybe they're wrong???)

http://www.bit-tech.net/hardware/201...1-5gb-review/3

Proof of what I'm saying is at Nvidia's website. There is no claim of MIMD shaders or MIMD CUDA cores.

[overclocking101]

probably end of this year early next

[erocker]

Yes Fermi is "MIMD" which stands for Multiple Instruction Multiple Data. To put it simply MIMD kind of works in the same way a dual core CPU works meaning it can work on "task A" and "task B" at the same time. It was necessary to implement for Fermi to do Tessellation well. I read all of this when Fermi reviews were released....

Found it (ignore my amateur explanation) : http://www.bit-tech.net/hardware/201...1-5gb-review/3

[_33]

Originally Posted by erocker

Yes Fermi is "MIMD" which stands for Multiple Instruction Multiple Data. To put it simply MIMD kind of works in the same way a dual core CPU works meaning it can work on "task A" and "task B" at the same time. It was necessary to implement for Fermi to do Tessellation well. I read all of this when Fermi reviews were released....

Found it: http://www.bit-tech.net/hardware/201...1-5gb-review/3

bit-tech is wrong. Nvidia was supposed to push for MIMD shaders, but they never did so far. This was a feature that was supposed to go with the never released GT300 (the other fermi).
http://www.brightsideofnews.com/news...s-a-cgpu!.aspx

[erocker]

Originally Posted by _33

bit-tech is wrong. Nvidia was supposed to push for MIMD shaders, but they never did so far. This was a feature that was supposed to go with the never released GT300 (the other fermi).
http://www.brightsideofnews.com/news...s-a-cgpu!.aspx

Well, I believe them over a "FUD" site. I'll have to ask someone who really knows.. I don't see any information to the contrary.

[_33]

Originally Posted by erocker

Well, I believe them over a "FUD" site. I'll have to ask someone who really knows..

This is why many believe there will be a fermi 2 (which should be what the GT300 was, with the MIMD CUDA cores).

Check this out, if the GTX480 had MIMD CUDA cores, it would literally crush the HD5870, without a doubt, in shader performance. But, it is not the case. 3D Mark scores would be owned grandly by the GTX480 if that would be the case, and it would be the HD5870 killer that everyone would of hoped for.

Just look at the "shader math test" or the "POM shader test" over here: http://ixbtlabs.com/articles3/video/gf100-2-p9.html

[eidairaman1]

Delusions of Grandeur

[erocker]

Originally Posted by _33

This is why many believe there will be a fermi 2 (which should be what the GT300 was, with the MIMD CUDA cores).

Check this out, if the GTX480 had MIMD CUDA cores, it would literally crush the HD5870, without a doubt, in shader performance. But, it is not the case. 3D Mark scores would be owned grandly by the GTX480 if that would be the case, and it would be the HD5870 killer that everyone would of hoped for.

Just look at the "shader math test" or the "POM shader test" over here: http://ixbtlabs.com/articles3/video/gf100-2-p9.html

Thing is, I read that tessellation would be broken without it in the design. Nvidia markets using marketing speak like "CUDA Cores". They also don't mention if it's SIMD, MIMD or anything else. Link to whitepapers?

[KainXS]

I know they don't the only thing they mention in the whitepapers about MIMD or SIMD is about the G80 being SIMD but not about the GF100 in general, but I have seen nothing that points to it being SIMD, not even rumors,

[phanbuey]

i have always heard that gf100 was mimd... but now that i think about it, those were the pre-rumors.

[mastrdrver]

Originally Posted by Yukikaze

The point was: Never has been =\= Never will be and they have the resources needed to make it happen if they really want to.

You're right on that point. Problem is that this isn't Intel's first try.....or second since there was suppose to be a successor to the i740 but sales were so dismal that it was canned. Infact, how many years have they even had an IGP and even that has suffered terribly from drivers.

I think one of the things going against Larrabee is the transistor density of GPUs. Bloomfield has very similar transistor and die size to the RV770......on an smaller process. Westmere is 240mm2 and 1.17B transistors which ATI does 1.04B in 170mm2 with Juniper.

Personally I think Intel would have a better chance at this if they would just start down at the bottom by making add in OEM cards. It would give them good experience with making GPU type chips and learning to make good drivers.

A lot of money and R&D doesn't do you any good when you don't even know how to make a GPU that even competes on the current OEM market. After reading all the articles on Larrabee everyone is talk about how Intel is trying to make a GPU.....out of a lot of CPUs. I think I'm pretty safe in saying that no one has ever done that with a GPU. Think there might be a reason? How many times has either ATI or nVidia missed the mark when trying to do something new and these companies know how to make GPUs.

[_33]

Originally Posted by erocker

Thing is, I read that tessellation would be broken without it in the design. Nvidia markets using marketing speak like "CUDA Cores". They also don't mention if it's SIMD, MIMD or anything else. Link to whitepapers?

The reason why tessellation is fast on the GTX480 is not because they implemented MIMD CUDA cores, but because Nvidia rewrote the way they handle geometry. Here's a quote from the Nvidia GF100 white paper (page 13):

Parallel Geometry Processing
Previous GPU designs have used a single monolithic front-end to fetch, assemble, and rasterize triangles. This fixed pipeline provided a fixed amount of performance to an arbitrary number of parallel execution cores. As applications differ in their workload, this pipeline was often bottlenecked or underutilized. The difficulty of parallelizing rasterization while maintaining API order also discouraged major innovations in this area. While the single front-end design has worked well in past GPU designs, it became a major roadblock as the need for geometric complexity increased.
The use of tessellation fundamentally changes the GPU’s graphics workload balance. With tessellation, the triangle density of a given frame can increase by multiple orders of magnitude, putting enormous strain on serial resources such as the setup and rasterization units. To sustain high tessellation performance, it is necessary to rebalance the graphics pipeline.
To facilitate high triangle rates, we designed a scalable geometry engine called the PolyMorph Engine. Each of the 16 PolyMorph engines has its own dedicated vertex fetch unit and tessellator, greatly expanding geometry performance. In conjunction, we also designed four parallel Raster Engines, allowing up to four triangles to be setup per clock. Together, they enable breakthrough triangle fetch, tessellation, and rasterization performance.

http://www.nvidia.com/object/IO_86775.html

Oh, and on P.16 they talk about the CUDA cores.

[nt300]

Hope this Southern Island performs well enough to strip gtx 480 bare. Hav you seen this and is truth or not? He speek like he works for ATI

ATI Hybrid Code Name: Southern Islands — ATI Boy!!! 2010-04-05 18:39
Fermi 1 & 2 Killer!!!
Despite the 40nm process, three key features that will be based on Evergreen have been put on steroids. Please observe bellow:

- 40nm Process
- 2400 Stream Processing Units (1)
- 256 Texture Address Units (2)
- 128 ROPs (Rasterization Operator Units) (3)

AMD plans to differentiate this re-fresh line from the past HD 4890 naming sceme due to its architectural changes. So don't expect the name HD 5890.

Though we will see an improvement in Tessellation, it?s not highly emphasised in Southern Islands (SI). You can look for this feature being greatly improved in code name Northern Islands.

ATI Code Name called ?Hecatoncheires? is something that will feature a series of technological advancements never heard of in the GPU industry. You will never find any information about this special architecture due to is being extremely top secret.
http://benchmarkreviews.com/index.ph...=480&Itemid=72

[yogurt_21]

Originally Posted by nt300

Hope this Southern Island performs well enough to strip gtx 480 bare. Hav you seen this and is truth or not? He speek like he works for ATI

those specs look entirely made up, 128 rop's? why?

[TVman]

Originally Posted by yogurt_21

those specs look entirely made up, 128 rop's? why?

128rop`s...... why not? more rops = better performance

[wahdangun]

ups it's really still SMID look at tomshardware :

"With GT200, a Streaming Multiprocessor was made up of one 8-way SIMD units, two special function units, and one double-precision unit. With GF100, they now have two 16-way SIMD units and four special function units."

source: http://www.tomshardware.com/reviews/...0,2585-18.html


so it's just a very powerful SMD

[wiak]

Originally Posted by yogurt_21

those specs look entirely made up, 128 rop's? why?

well HD 4870 had 800 shaders, HD 5870 has 1600 shaders and so on :P

[erocker]

Originally Posted by nt300

Hope this Southern Island performs well enough to strip gtx 480 bare. Hav you seen this and is truth or not? He speek like he works for ATI

I'm sorry to say, no. I can confidently tell you that no one who works for ATi would make any post like that on any internet forum. Absolutely ridiculous.

[nt300]

Originally Posted by yogurt_21

those specs look entirely made up, 128 rop's? why?

Question the ATI Boy that said it. I just post what I found. I don't know, for better FPS in higher res?

[erocker]

Originally Posted by nt300

Question the ATI Boy that said it. I just post what I found. I don't know, for better FPS in higher res?

Can you please not bother posting other people's posts from other forums? It's unnecessary and spreads rumor which is against this forums guidelines. Thanks.

[nt300]

Originally Posted by erocker

Can you please not bother posting other people's posts from other forums? It's unnecessary and spreads rumor which is against this forums guidelines. Thanks.

Woops didn't now that