Does AMD have a 2304 stream processor GPU in waiting?

Subject: Graphics Cards | January 2, 2012 - 07:23 PM |
Tagged: amd, southern islands, rumor, leak

The review for the AMD Radeon HD 7970 3GB card based on the Tahiti GPU and the new Southern Islands architecture was released on December 22nd with expected availability on January 9th.  In that review we show a diagram of the Tahiti GPU and its 32 Compute Units (CUs) that combine to form the total 2048 stream processors (SPs).  

We asked and asked, but a die shot was never given to us for our review - a very non-standard practice for new launches.  That started us wondering - was there something AMD was hiding from reviewers about the architecture?  Were there some disabled CUs on the 28nm GPU that they had disabled for business, yield or clock speed reasons?  Think of what Intel has done with Sandy Bridge-E or what NVIDIA did originally with the GTX 480 GPU.  

AMD assured us that was not the case - Tahiti is the full die enabled, 32 CUs and 2048 SPs.  And, based on some of our own internal information, that seems to be 100% the case.

But, an interesting image started floating around last week:

View Full Size

This image from the site ChipHell.com appears to show the development sheets for Sapphire's upcoming Radeon HD 7000 series products and their internal codenames.  There are some really noteworthy things to look at though starting with the Atomic lineup.  

View Full Size

While the Flex 6G is a 6GB card with 6 mini-DP ports on it running at the same clock speeds as our reference designs did initially, the Atomic RX card has a clock speed of 1335 MHz running on 2048 SPs and a pretty good memory overclock as well.  If that is accurate, the performance difference between the Atomic RX and "Da Original" (likely the reference card) would be tremendous!

View Full Size

Here is what is more interesting - another card listed above the HD 7970s that seems to include 2304 SPs, or 36 CUs.  Running at a reference speed of 1000 MHz, this card would have a noticeable advantage over the current HD 7970 cards.  What's more...?

View Full Size

The Toxic ZX, if it exists, would run with 2304 stream processors at 1225 MHz!  The performance of this card could easily beat out the Radeon HD 7970 3GB card by 35-45% with the shader and clock speed differences.

So, what does this all mean?  Probably nothing, but it is fun to speculate on a few things.  It seems possible that AMD either HAD or HAS another GPU waiting in the wings based on Southern Islands to compete with NVIDIA's Kepler when it finally gets released.  Even though these documents seem to indicate that, I kind of find it hard to believe that AMD would have been able to keep this secret from the media and the competition for this long.  It is also equally unlikely that AMD was able to quickly tape out another chip that we are unaware of as even a somewhat moderate change like adding in four very modular CUs takes many months.

And of course, we have to take in the possibility that these are all fake, or a decoy or were written up 18 months ago and plans have changed.  Those are much less fun though. 

Source: Guru3D
January 3, 2012 | 08:56 AM - Posted by nabokovfan87

Short answer is NO. This design has changed drastically from what the leaks had indicated just a month ago. They changed from XDR to GDDR5 memory as well as tweaking some things.

Fact is, things change during the development process due to heat testing and as a means to get better yields. Going from 2048 to 23xx or even 29xx stream processors isn't going to be as powerful as doubling to 4096 w/ 6 GB of memory.

Taking off a few hundred to dramatically increase yields make a heck of a lot more sense to me. It won't piss off the 7970 orders and those who waited 3-6 months for "this" version weren't going to buy one anyway.

January 3, 2012 | 09:15 AM - Posted by Ryan Shrout

In reality though, the kind of changes we are talking about here, can simply NOT HAPPEN in a matter of months. You can't build a new chip with 32 CUs rather than 36 CUs in anything under a year more than likely and you definitely couldn't move from XDR to GDDR5 in that kind of time frame either. The process of building, tweaking and producing a chip is just much too complicated for that.

So either this chip with 2304 SPs NEVER existed or it still exists, it can't have been an on the spot evolution.

January 3, 2012 | 11:32 AM - Posted by nabokovfan87

Completely agree Ryan,

I was saying the rumors a few months ago and released specs were one thing, which indicated when and how this happened.

They probably started with around 2500 stream processors and the XDR memory. The XDR was cut for cost and the stream processors cut due to fabrication, lithography, heat, or yield issues with the new process. What was leaked out was clearly wrong, even though we saw it 3-6 months ago, it was drastically older in my opinion and as you specify as to why.

Has there ever been a history of cards where they add components like this apart from CUDA where cores are "activated" or added but the design had already called for it originally and just wasn't implemented?

January 3, 2012 | 01:25 PM - Posted by guesswho? (not verified)

or that HD 7970 has some CUs disabled due to yield problems like the orginal fermi had

January 3, 2012 | 02:58 PM - Posted by nabokovfan87

No...

If that was the case the 7970 would be the 7950 the 7950 would be the 7930 and the 7970 would be the 23xx stream processor version and be delayed due to yield issues.

With fermi they "refreshed" the lines over time. They had yield issues and stuff at launch which led to retail delays. They launched 104 revision at the time of the 580, which was a different revision, so customer's weren't mad.

It was a while ago, so if I have the details wrong my bad.

January 3, 2012 | 06:06 PM - Posted by Ryan Shrout

That was our first suspicion but I can without a doubt say that is not the case.

January 3, 2012 | 09:41 PM - Posted by Anonymous (not verified)

yeah and you must be Mr know it all, right?

anything is possible, especially since amd releasing a card with massive clock potential but still kept clocks, power, heat & noise low (playing around with nvidia - some fun time).

January 3, 2012 | 11:39 PM - Posted by Ryan Shrout

Your witty comments not withstanding, I am 100% sure the models sent to reviewers in December did not have more than 32 CUs on the die.

January 4, 2012 | 06:58 AM - Posted by Anonymous (not verified)

If you look at the picture on top of BIOS Switch, you can see it says "Susp"...Suspended?

January 4, 2012 | 09:24 AM - Posted by Ryan Shrout

Yeah, I have no idea...would love to know what that word is.

January 4, 2012 | 11:55 AM - Posted by nabokovfan87

Just on verbage and so forth, it would make sense that it says "susepended" or "suspended projects" or something along those lines. Meaning the toxic ZX is on reference in their system for billing and research purposes (test result lookups and lithography billing lookups, part numbers, etc.) but was cancelled and cannot be removed from the system.

On the second section of the sheet instead of -suspended it has -Radeon HD 7970 indicating those parts are to be made or in process/working projects currently in development at the time of the photograph.

Da Original = reference card
Dual-Fan 3G = Low Power/Heat Design (Equivalent to MSI Frozr II/III cards)
VaporX 3G = Unknown (Mild OC?)
Flex 3G = Eyefinity edition (6G of memory and 6xDP)
Toxic 3G = Standard OC Card
Atomic WC = Extreme OC Water Cooled Card ("WC")
Atomic RX = Extreme OC Air Cooled Card

So yeah, that's what it all means in my view.

January 5, 2012 | 06:26 AM - Posted by Anonymous Coward (not verified)

Ah, the mighty 6970. Single link DVI only. You'd think foxconn themselves designed it. Stay classy AMD!

January 5, 2012 | 06:26 AM - Posted by Anonymous Coward (not verified)

Meant 7970 derp

January 5, 2012 | 06:50 AM - Posted by Preetam Nath (not verified)

Well, AMD did say it's a 2048 SP chip only.

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.