AMD: R8xx Speculation

[Wirmish]

THERE'S A RUBY ACTION FIGURE?!?!

Personally I prefer the new black Ruby....


EDIT:

Red Ruby : YouTube
Black Ruby : YouTube (watch in HD)

 

[digitalwanderer]

Oh man, now THIS is the kind of silly season excitement I LIVE FOR!!!

 

[turtle]

THERE'S A RUBY ACTION FIGURE?!?!

Ha...I was just looking at that. That's pretty bad ass.

I wish AMD would sell that shit...I'd buy it...something to appease the fanboi in me jealous of those with their bad-fill keychains. My ATi mini fridge is so lonely...jk.

Any advertising is good advertising...Did AMD actually learn this?

Granted, that's probably the only one, as it's all their PR/Advertising could afford.

Yeah, I said it...you know you were thinking it.

 

[BRiT]

THERE'S A RUBY ACTION FIGURE?!?!

And a blow up doll...

 

[Squilliam]

And a blow up doll...

I want... where can I has both?

 

[rjc]

Nordic Hardware, March 2008 version:

RV770 is in production, 480sps and 32TMUs indeed
The old HD 4000 specifications are a bit off, but not completely. RV770 builds on the R600 framework and it seems that RV770 will have 480 unified shaders. RV670 has clusters of 16 shading units, and each units have five sub-units that do the actual work. Although, it's more of a 4+1 design, where the fifth can do more things that first four. This will remain, but there will be more 10 more clusters (160 shaders) and twice as many TMUs, up to 32 units.

Nordic Hardware, August 2009 version:

Evergreen at Quakecon, odd memory configs, samples shipping
"You might have heard rumors of some odd memory configurations going around and a source close to AMD has confirmed this. Numbers are pointing to something like 384-bit and GDDR5 with the high-end X2 part, 192-bit and GDDR5 with the performance part, and then scaled down to 96-bit and perhaps even 48-bit with lower-end parts. The numbers are still pending confirmation, but we are quite confident AMD will bring some odd memory configurations to the market with Evergreen."

Who said people don't learn from their mistakes?

 

[NH|Delph1]

Nordic Hardware, March 2008 version:
Nordic Hardware, August 2009 version:
Who said people don't learn from their mistakes?

The misinformation campaign from the pre-RV770 era mislead several of my sources. And they actually confirmed 480sps, but were wrong (very annoying experience...). AMD managed to screw with practically every news site in the Western media.

You should perhaps disregard the ROPs calculation in my older news post. I must have been tired. Whether its 192 or 384-bit for RV870 is hard to say, but I'd prefer to stay conservative. The memory bandwidth of 384-bit + DDR5 seems a bit unnecessary, and it would lead to a more expensive PCB, higher costs, etc. etc. not something you identify AMD/ATI with today.

//Andreas

 

[rjc]

The misinformation campaign from the pre-RV770 era mislead several of my sources. And they actually confirmed 480sps, but were wrong (very annoying experience...). AMD managed to screw with practically every news site in the Western media.

Yes must have been very frustrating, i (half) remember something about during the design process for the RV770 the spec was upgraded (pad limiting maybe?), so maybe the 480sps was just out of date information rather than per se wrong information?

You should perhaps disregard the ROPs calculation in my older news post. I must have been tired. Whether its 192 or 384-bit for RV870 is hard to say, but I'd prefer to stay conservative. The memory bandwidth of 384-bit + DDR5 seems a bit unnecessary, and it would lead to a more expensive PCB, higher costs, etc. etc. not something you identify AMD/ATI with today.

I looked through all the previous ATI cards i could find and i couldn't find any the have any of the memory interfaces you mentioned. Is always 64bit, 128bit, 256bit. I did find the 3200 IGP did use 32bit though at least according to wikipedia.

Forgive my ignorance but how does a 48bit memory interface work? I was maybe under the false assumption that GDDR3 chips only work in mutiples of 32 bits. I think there is a x16 mode for DDR but would you design a chip around it?

 

[NH|Delph1]

Yes must have been very frustrating, i (half) remember something about during the design process for the RV770 the spec was upgraded (pad limiting maybe?), so maybe the 480sps was just out of date information rather than per se wrong information?

They haven't officially admitted to it, but grins, smiles and the overall smug sense of selfsatisfaction along with some [add Nelson voice]"Ha-Ha"[remove Nelson-voice], and advice of maybe not trusting industry sources in the future (and a drunken admittance) have more or less confirmed the misinformation campaign

If I remember correctly, Theo, Fuad, and many more was fed the exact same info, on purpose ;-)

I looked through all the previous ATI cards i could find and i couldn't find any the have any of the memory interfaces you mentioned. Is always 64bit, 128bit, 256bit. I did find the 3200 IGP did use 32bit though at least according to wikipedia.

Forgive my ignorance but how does a 48bit memory interface work? I was maybe under the false assumption that GDDR3 chips only work in mutiples of 32 bits. I think there is a x16 mode for DDR but would you design a chip around it?

All signs point to other memory configurations than the regular 64, 128, 256, but there are many possible multiples of 32 to go with. 192 and 384 are the two that have been mentioned to me though.

As in the lower range chips, it's much harder to find information. All the cards shown so far has been mid-range. The card at Quakecon is for the $100 segment. And you're right about the 48-bit part, I guess they won't scale below 96-bit. I'm still digging and for some reason there have been extremely few leaks.

//Andreas

 

[rpg.314]

There werent many real leaks leading up to the launch of rv770 either. So AMD has pretty tght control over them it seems. Personally, I think rv870 will have <1600 alu's, even if it is ~300mm2.

 

[Scrat]

What about this:

Juniper - 128bit memory interface
Cypress - 2xJuniper, MCM, 256bit memory interface
Trillian - 3xJuniper, MCM, 384bit memory interface
Hemlock - 2xTrillian | 2xCypress, CF on a board, 2x384bit memory interface.

This would confirm the 384bit rumor...

 

[rjc]

They haven't officially admitted to it, but grins, smiles and the overall smug sense of selfsatisfaction along with some [add Nelson voice]"Ha-Ha"[remove Nelson-voice], and advice of maybe not trusting industry sources in the future (and a drunken admittance) have more or less confirmed the misinformation campaign

I found this in the RV770 article at Anandtech:

Initially, RV770 was targeted at 1.5x the performance of R600, which looking back would not have been enough. During the next 1.5 years that 1.5x turned into 2x R600 and finally settled at 2.5x the speed of R600, at a price in the $200 - $300 range

So if you interpret the above as performance = no of shaders. It would have started as 1.5x320=480, then 2x320=640 and finally 2.5x320=800. Maybe they just leaked the first design document?

All signs point to other memory configurations than the regular 64, 128, 256, but there are many possible multiples of 32 to go with. 192 and 384 are the two that have been mentioned to me though.

Speaking of the original specifications for chips, did you see this from chiphell admin(post #15 in thread)

其实RV870最早的规格很黄很暴力.....1920SP...384bit....可惜TSMC不争气....
(Actually RV870 earliest specifications very violent...1920SP...384bit....it's a pity TSMC fail to live up to expectations....)

As in the lower range chips, it's much harder to find information. All the cards shown so far has been mid-range. The card at Quakecon is for the $100 segment. And you're right about the 48-bit part, I guess they won't scale below 96-bit. I'm still digging and for some reason there have been extremely few leaks.

Did you see the mobile chips leak a few pages back in this thread? (codenames, memory interfaces and power). I did a post just after with a quick comparison to the RV7X0 mobile parts based on the power used. As lower chips dont scale nearly as well, probably if you were brave, could possibly estimate unit counts and die sizes from that.

 

[mczak]

I looked through all the previous ATI cards i could find and i couldn't find any the have any of the memory interfaces you mentioned. Is always 64bit, 128bit, 256bit. I did find the 3200 IGP did use 32bit though at least according to wikipedia.

Pretty sure this is wrong (though I don't know if that was also true for the older rs690 IGP). Sideport memory is only 16bit. At least I haven't seen any board which uses 2 memory chips...

Forgive my ignorance but how does a 48bit memory interface work? I was maybe under the false assumption that GDDR3 chips only work in mutiples of 32 bits.

Doesn't seem to make much sense indeed, though if you use DDR3 memory instead they only come in 8-bit and 16-bit flavors, not 32-bit, so there wouldn't be a problem.

 

[no-X]

I found this in the RV770 article at Anandtech:
So if you interpret the above as performance = no of shaders. It would have started as 1.5x320=480, then 2x320=640 and finally 2.5x320=800. Maybe they just leaked the first design document?

Orton mentioned 96 ALUs (480 SPs) GPU even pre-R600 launch (he likely talked about "R650", which changed to RV770 during time), so some of the rumour could be based upon this info.

 

[NH|Delph1]

I found this in the RV770 article at Anandtech:
So if you interpret the above as performance = no of shaders. It would have started as 1.5x320=480, then 2x320=640 and finally 2.5x320=800. Maybe they just leaked the first design document?

Not at all impossible, but when several independent sources say the same thing it becomes harder to explain.

Speaking of the original specifications for chips, did you see this from chiphell admin(post #15 in thread)
Did you see the mobile chips leak a few pages back in this thread? (codenames, memory interfaces and power). I did a post just after with a quick comparison to the RV7X0 mobile parts based on the power used. As lower chips dont scale nearly as well, probably if you were brave, could possibly estimate unit counts and die sizes from that.

I noticed, but since I don't know for sure yet I can't really elaborate further. He has a good striking average though.

//Andreas

 

[Vincent]

Anyone notice the last post by Chiphell's admin-

It is rumored that RV8XX series will be launched during the mid Sep.

-------------------------------------------------------------------------------------

September is the greatest month among the rest in 2009

1. 12, Septermber---> Porsche Panamera

2. 16, September-----> AMD's first DX11 product line-up

 

[Kaotik]

Anyone notice the last post by Chiphell's admin-

It is rumored that RV8XX series will be launched during the mid Sep.

-------------------------------------------------------------------------------------

September is the greatest month among the rest in 2009

1. 12, Septermber---> Porsche Panamera

2. 16, September-----> AMD's first DX11 product line-up

Well we already got the AMD slide/pic whch says September 10

 

[w0mbat]

Maybe its press only...

 

[Kaotik]

Maybe its press only...

That's a possibility of course, then again, HD4850 part deux is also possible

 

[Humus]

THERE'S A RUBY ACTION FIGURE?!?!

Yup, I have one.

Personally I prefer the new black Ruby....

Rubies are red dammit!!!111one