AMD is on a tear now

Today at Computex, AMD has unveiled its first DirectX 11 compliant part, but the announcement is more about the overall features that DX11 introduces rather than the specific part that AMD has already fabricated in conjunction with TSMC on their fledgling 40nm process. Read on for the full scoop.
Today at Computex, AMD has unveiled its first DirectX 11 compliant part, but the announcement is more about the overall features that DX11 introduces rather than the specific part that AMD has already fabricated in conjunction with TSMC on their fledgling 40 nm process.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 22

Windows 7 will be the enabler of DX11 and Stream Compute for the desktop, and AMD is hoping (and is likely to be) the only launch partner with DX11 products for sale when Windows 7 launches.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 23

ATI/AMD has not been shy from integrating the latest and greatest into their products.  While they were slow on the 130 nm uptake with their R300, they quickly gained a reputation for leveraging the latest process nodes and memory technologies.  While each product may not have been entirely successful (Radeon X1800, HD 2900 XT), the company continues to push the envelope with their products.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 24

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 25

The proof that AMD does have 40 nm DX11 silicon (of course… it could be RV740 chips, but I somehow doubt it).  Doing some napkin type math, and assuming this is a 300 mm wafer, then the individual dies shown here are around 180 mm square.  This puts it right in the sweet spot in terms of die size.  For comparison, the RV770 is approximately 260 mm square on the 55 nm process.  The 40 nm process is one full process node down, so die savings are around 45% to 50%, depending on factors such as SRAM cells, custom logic, etc.  So, it appears as though this product is going to be around 1.2 billion transistors in size (again, napkin math).

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 26

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 27

These are the big changes in DX 11 vs. DX 10/10.1.  We have seen most of them before, but not all in one technology specification.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 28

TruForm was introduced with the Radeon 8500 way back in 2002, which was based on the R200 chip from ATI.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 29

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 30

Yes, the Ruby demo was about the only real showcase for tesselation at this time.  Interesting, but not supported by PC games.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 31

The HD 4000 series still supports ATI’s tesselation unit, but again it is not used at this time.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 32

Finally with DX11 we will see tessellation used in actual games.  I remember graphics folks talking about tessellation and gaming since GDC 2000.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 33

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 34

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 35

In this example compute shaders will control the behavior of these creatures, which is reactive to the physical situation and surroundings.  These creatures appear to be choosing the path of least resistance while hiking through rugged terrain.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 36

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 37

Currently MediaShow Espresso can utilize ATI Stream, and the results in this particular test show a signficant advantage in performance with the HD 4670 against the GeForce 9500 GT.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 38

Drag and Drop refers to a realtime transcode of a video going from the Windows 7 desktop to a portable media device.  So instead of doing the transcode separately, then copying the file to the device… this allows users to just drop the media file onto the device without having the first transcode it.

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 39

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 40

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 41

Computex 2009: AMD Demonstrates working 40nm DirectX 11 Silicon - Graphics Cards 42

It is actually quite a surprise that AMD has actual working DX11 silicon, and from the picture we can gather that the first product might just be the high end unit from AMD.  Consider that the HD 3780 was based on the RV670 chip, which was also right around 180 mm square in size on 55 nm, and it was the high end product for AMD until the release of the larger and more powerful HD 4870.

Another thing to take away from this presentation is that currently TSMC is having a hard time producing 40 nm parts in good quantities.  This is exhibited by the dearth of available Radeon HD 4770 which are based on the RV740, a 40 nm chip which is around 137 mm square in size.  At the very least, we can expect the upcoming chip to be slightly faster than the current HD 4890, but with the added bonus of full DX11 support.  I am unsure if the jump from DX10.1 to DX11 will be as transistor intensive as going from DX9 to DX10, but I do not think that it is going to be as drastic a change.

AMD certainly has set a aggressive pace, and so far we have not seen any information from NVIDIA that they have a working DX11 part in house (all rumors point to them NOT having any working silicon at this time).  If TSMC can fix its 40 nm issues, and AMD can ramp these parts in a timely manner, then it is not unreasonable to expect actual product to be available for the October launch of Windows 7.