AMD’s Upcoming HD 7970 Exposed – a Short-lived Video card?

Changes in 7000 series

AMD’s Radeon HD 7900 series will use a brand new GPU built on the 28nm process, codenamed “Tahiti”.  The top two GPUs – HD 7970 and HD 7950 each have 4.3 billion transistors. This represents a major break from their scalar VLIW-based computing since HD 2900XT was released, to GCN Quad SIMD.

AMD’s new buzzword is “performance per square mm” as it is higher than the old architecture for DX11 and GPU computing.   In this case, the slide below appears to be comparing benchmarks which are much weaker on the HD 6970.  And of course, no parameters for any tests are given.  We look forward to the benchmarks on Thursday.

 AMD  intends to catch up to Fermi in GPU computing with this architecture and we can also expect greatly improved geometry and  Tessellation.  This new Tahiti GPU also supports DX11.1 and PCIe 3.0

And you are going to be hearing a lot of “firsts” – first to DX11.1 and first to PCIe 3.0 even though there are no DX11.1 games yet nor will next year’s upcoming dual-chip HD 7990 saturate even PCIe 2.0

Tahiti features 2,048 stream processors, all of which are enabled on the Radeon HD 7970. These stream processors are arranged along 32 computing units, each with 64 SPs.  It has a 384-bit wide memory interface which provides bandwidth around 260 GB/s.  The stock HD 79×0 comes with 3 GB of GDDR5 memory using twelve 2Mbit memory chips.

In the above picture posted by VR-Zone, the 5+1 phase PWM is visible. We note that the Tahiti die is rotated 45 degrees and partially covered by a heat-spreader although the die itself is exposed to the heatsink.   It’s a big GPU, though a bit smaller than Cayman.  The card features two CrossFire connectors as well as the dual BIOS toggle switch just like the HD 6900 series.

Turkish web site Dominhaber originally posted these specifications with just one significant error regarding the ROPs.

• 4.50 billion transistors, die-area of ~380 mm², built on TSMC 28 nm process
  Advanced GCN 1D architecture
  2048 1D processing cores
  128 TMUs, 48 ROPs
  384-bit wide GDDR5 memory interface, clock slightly below 1 GHz, target bandwidth of 240~264 GB/s
  6pin + 8pin power connector required
  PCI Express Gen 3.0
  DirectX 11.1 support

Another likely AMD PR slide (with misspelled “embargo”) posted by OBR-Hardware confirms that there are 32 ROPs, not 48.   In the new GCN architecture, ROPs are no longer directly coupled to the memory-controllers.

Here is a handy chart which compares the HD 7970 to the HD 6970

Consuming less than 3W at idle is an amazing new feature, considering that the HD 6970 consumes about 20W.  As you can see, AMD is pleased to compare the HD 7970’s idle wattage against their competitors using QuadFire as an example.

Here is another slide from the leaked press deck:

AMD reference board design still retains their signature black and red color scheme. Even the PCB is black. The cooler shroud is curved and a pair of HD 7970s appear to be designed to work better together in CrossFire, both acoustically and with better airflow between the cards than with the relatively noisy HD 6900 series.

Display connectors include DVI, HDMI, and two mini-DisplayPort connectors.  The display ports are all located in one PCIe slot area so designed to exhaust hot air from the second slot and it would be ideal for watercooling.  Board partners are evidently required to include HDMI to DVI and active-DP dongles with the video card.

Even though the HD 7970 is clocked conservatively to increase AMD’s yields, the picture will change dramatically when the HD 7970 is overclocked from its base of 925W with its TDP going well over 300W.  AMD expects this GPU will clock “well in excess of 1GHz”.  This ability to overclock the HD 7970 is its strongest feature and we can expect AMD’s partners to take full advantage of this overclocking headroom to distance the HD 7970 from even the highly overclocked GTX 580s.  ABT will be at CES and we will bring you images and specifications of these new designs.

The HD 7970 is based on GCN architecture. According to the slide, it packs 32 ROPs and 128 texture units.  ROPs are decoupled from the memory controller in this new architecture. Dual geometry engines are featured and an impressively large L2 global cache of 768KB.

With Tahiti’s new GCN architecture, each CU cluster accesses their own TMU/ROPs directly and there is a bigger cache than with VLIW, including a large global L2 cache. In VLIF architecture, the ROPs were shared and an interface regulates the access.

Even so, the HD 7970 still appears to be a bit unbalanced as though some of the function units are disabled, but that is pure speculation.  However, it wouldn’t surprise us to see a “refresh” and a respin of the HD 7970 if/when the GTX 680 turns out to have more performance than AMD is expecting, early next year.

We are getting reports that the HD 7970 is about 20-30% faster than the GTX 580, depending on the game or benchmark used – and faster if it is overclocked.  According to the AMD leaked slide from Dominhaber, tessellation is improved at least 1.5 times over the HD 600o series and we expect the most favorable AMD benches to be DX11 and heavily tessellated.  There will be no emphasis on DX9 gaming as far as we can tell.

These most favorable benches would include Battlefield 3 at very specific benching conditions.  It is very doubtful that the stock HD 7970 will be 1.4 times faster than the GTX 580 across the board.  They will be lucky to get over 25% according to what we are hearing now.

Of course, we are looking forward to AMD’s answer to 3D Vision 2 Surround – Eyefinity 2.0 married to HD3D!  We will have our own comparison of HD3D vs. 3D Vision 2 at ABT later this month.

HD3D now has improved with support for 3GHz HDMI and frame packing support for S3D.

And we see improvements for Eyefinity 2.0

Of course, we look forward to the actual benchmarks on Thursday.  Let’s head to our preview’s conclusion,