AMD: R8xx Speculation

[rjc]

The highest-end enthusiast offerings are purportedly called Cypress, with performance offerings dubbed Redwood, mainstream offerings called Juniper and Cedar, and low level entry offerings named after the poisionous shrubbery, Hemlock. [bwing me a shwubbery!]

Er the order of the above in botany seems kindof wrong.

Evergreen trees:

• Redwood(Cupressaceae) > 100m
• Cypress(Cupressaceae) 10-70m(three separate families)
• Juniper(Cupressaceae) 20-40m
• Cedar 30-40m tall (There are 2 types of cedar one in Pinaceae family, the other is a variety of Juniper)
• Hemlock(Pinaceae) 10-60m tall

Seems a bit strange the "Cypress" is ahead of "Redwood" in performance, is a much smaller thing in the plant world.

Interesting Hemlock and maybe Cedar are in different families to the other 3 trees. Note to others if making up further names that all of above trees appear to be of North American origin.

 

[Jawed]

6 codenames needed

http://www.semiaccurate.com/forums/showpost.php?p=1152&postcount=6

There is one other code name of note that hasn't turned up yet. That and some of the linked info is a bit off. I can't say more yet.

What about:

• Juniper - 800 ALUs, 40 TUs, 16 RBEs, 128-bit - 1.2TFLOPs? 750MHz, since RV740 is 750MHz
• Cypress - 1600 ALUs, 80 TUs, 32 RBEs, 256-bit - 2TFLOPs+? 700MHz+?
• 2x Cypress = Redwood - 3200 ALUs, 160 TUs, 64 RBEs, 512-bit - 4TFLOPs+?

for starters?

Could some of these codenames be laptop variants?

Jawed

 

[Scrat]

http://www.semiaccurate.com/forums/showpost.php?p=1152&postcount=6



What about:

• Juniper - 800 ALUs, 40 TUs, 16 RBEs, 128-bit - 1.2TFLOPs? 750MHz, since RV740 is 750MHz
• Cypress - 1600 ALUs, 80 TUs, 32 RBEs, 256-bit - 2TFLOPs+? 700MHz+?
• 2x Cypress = Redwood - 3200 ALUs, 160 TUs, 64 RBEs, 512-bit - 4TFLOPs+?

for starters?

Could some of these codenames be laptop variants?

Jawed

What about this:

• Juniper - 1000 ALUs (still 10 clusters, just longer), 40 TUs, 16 RBEs, 128-bit - 1.5TFLOPs - 750MHz
• Cypress - 2000 ALUs, 80 TUs, 32 RBEs, 256-bit - 3TFLOPs 750MHz
• 2x Cypress = Redwood - 4000 ALUs, 160 TUs, 64 RBEs, 512-bit - 6TFLOPs

Too optimistic? :>

 

[nbohr1more]

What about this:

• Juniper - 1000 ALUs (still 10 clusters, just longer), 40 TUs, 16 RBEs, 128-bit - 1.5TFLOPs - 750MHz
• Cypress - 2000 ALUs, 80 TUs, 32 RBEs, 256-bit - 3TFLOPs 750MHz
• 2x Cypress = Redwood - 4000 ALUs, 160 TUs, 64 RBEs, 512-bit - 6TFLOPs

Too optimistic? :>

I think Jaws is hoping for efficiency gains, so this is either optimistic or pessimistic depending on what you hope for... I'm hoping they take the risky route of beefing up the ratio of Vertex specific logic ... kinda like how Nvidia took the risky route of beefing up the GPGPU hardware... I like my triangles!!! Plus, next-gen HOS support?

 

[LordEC911]

Sooo... something to keep in mind, in the Anandtech.com interview it was stated that the designs they are just starting to work on, due out in 2012, are going to be 4TFlops. I assumed that was the top of the line cards, x2 variants included, but now it sounds like they meant just single GPUs.

The team that I met with in Santa Clara was talking about designs that would be out in 2012, we’re talking 4 TFLOPS of performance here, 4x the speed of RV770. http://www.anandtech.com/video/showdoc.aspx?i=3469&p=3

So we are looking at ~2-2.5TFlops for RV870, ~3TFlops for the next chip(small refresh/tweak?) in late 2010/early 2011 to meet the ~4TFlops for the 2012 release.

 

[Jawed]

What about this:

• Juniper - 1000 ALUs (still 10 clusters, just longer), 40 TUs, 16 RBEs, 128-bit - 1.5TFLOPs - 750MHz
• Cypress - 2000 ALUs, 80 TUs, 32 RBEs, 256-bit - 3TFLOPs 750MHz
• 2x Cypress = Redwood - 4000 ALUs, 160 TUs, 64 RBEs, 512-bit - 6TFLOPs

Too optimistic? :>

Just another variable. Seems unlikely to change. We've gone over the implications of different SIMD widths.

Jawed

 

[Jawed]

I think Jaws is hoping for efficiency gains, so this is either optimistic or pessimistic depending on what you hope for... I'm hoping they take the risky route of beefing up the ratio of Vertex specific logic ... kinda like how Nvidia took the risky route of beefing up the GPGPU hardware... I like my triangles!!! Plus, next-gen HOS support?

Jaws is someone else...

You like lots of triangles and HOS? That's what tessellation is for...

Jawed

 

[Jawed]

Sooo... something to keep in mind, in the Anandtech.com interview it was stated that the designs they are just starting to work on, due out in 2012, are going to be 4TFlops. I assumed that was the top of the line cards, x2 variants included, but now it sounds like they meant just single GPUs.

So we are looking at ~2-2.5TFlops for RV870, ~3TFlops for the next chip(small refresh/tweak?) in late 2010/early 2011 to meet the ~4TFlops for the 2012 release.

4TFLOPs sounds really conservative to me. Hmm. That should be the bottom card, bearing in mind the rising tide of IGP. Hell, it could be the only chip, hohoho.

Jawed

 

[CJ]

The original codename on the first video was "EG Broadway", which I interpret to mean "Evergreen Broadway". I've just realised that "Broadway" would be the name of a road (like Avenue, or Boulevard). So Evergreen Broadway would be a road lined with Evergreen trees. Which implies "multiple" evergreens. Which makes me think MCM

Jawed

Check your PM and you know where it came from.

 

[Jawed]

I've noticed, separately, Hudson and Manhattan codenames on mobile roadmaps - something I forgot to mention earlier, so yes, New York New York

So D3D11 laptops are coming soon?...

Jawed

 

[phenix]

I've noticed, separately, Hudson and Manhattan codenames on mobile roadmaps - something I forgot to mention earlier, so yes, New York New York

So D3D11 laptops are coming soon?...

Jawed

XbitLabs quotes Dirk Meyer, predicting that DX11 laptops to show up in the first half of 2010.

http://www.xbitlabs.com/news/video/...rectX_11_Graphics_Lineup_to_Ramp_Up_Fast.html

 

[hoom]

One of the future headaches IHVs might face with advanced tesselation or else a gazillion of miniscule triangles could be multisampling efficiency (compared to supersampling) to suffer quite a bit. One of the recent NV patents showed that they're probably researching in that direction also.

ATI already has its Edge Detect AA mode since R600.

The prospect of an MCM GPU that looks like a single GPU with 32 ATI RBEs at 700+Mhz is very exciting from an AA point of view even with plain old MSAA

 

[CarstenS]

ATI already has its Edge Detect AA mode since R600.

The prospect of an MCM GPU that looks like a single GPU with 32 ATI RBEs at 700+Mhz is very exciting from an AA point of view even with plain old MSAA

Sorry, I don't get it: What's your point?

 

[rjc]

Ha ha didn't notice this before....

.... but the poison code-named R800 (dual RV870) victory sure....

Above orginally from originally from cfcnc at chiphell. There is only one likely to be considered a poison in the lineup, therefore the high end dual card codename is "Hemlock". It's actually the non toxic Tsuga but easily confused with the toxic shrub Conium

That screws my previous theory of the order being based on the tree's size...Looking again at the five available codewords, Hemlock is in a different family:
Cupressaceae: Cypress, Juniper, Cedar, Redwood
Pinaceae: Hemlock

ie The Cupressaceae family is older and spread all over the world, the Pinaceaes relatively young and more concentrated, northern hemisphere, northern climes only.

Have a feeling am putting much more effort into this than whoever invented these codenames...

 

[nbohr1more]

Jaws is someone else...

You like lots of triangles and HOS? That's what tessellation is for...

Jawed

Sorry Jawed ... I think Jaws is a Rage3d regular if I'm not mistaken...?
When I say HOS ... I'm thinking of curved surface support..? Somehing that was hinted to be in the Gamecube GPU... I thought tessellation only amplifies triangle, I though HOS has to do with non-triangle based geometry... I am probably using the wrong term

 

[Jawed]

"D3D11 tessellation" is a kind of generic term for support for HOS and adaptive level of detail:

http://developer.download.nvidia.com/presentations/2009/GDC/GDC09_D3D11Tessellation.pdf

 

[nbohr1more]

I still haven't found the reference but I could have sworn that some developer said that the Gamecube Flipper had something similar to the Nvidia NV1\STG-2000 quadratic texture mapping technology (non-triangle based rendering) but it was specialized just for curved surface effects... it may even have been Factor 5 with RS II : RL

 

[nbohr1more]

I re-read the HOS documentation so I won't be mistaking that for quadratic texture mapping any time soon...

 

[Kaotik]

http://brightsideofnews.com/news/2009/7/22/ati-to-launch-directx-11-gpus-in-seven-weeks.aspx

I know it's theo, one of the most unreliable "sources" for news, but does anyone have any tidbits on if he's in fact correct on this one?

 

[3dilettante]

At best, he may have correctly matched up essentially meaningless code names to the product segments they target, and the September time frame would be intact.

It seems odd that redwood would be so far down, and inserting a picture of a flowering plant for the evergreen line violates the theme.

Not sure what else there is to say.