Business, Graphics, Hardware

AMD Northern Islands Vs. NVIDIA Fermi Die Sizes Compared


AMD’s last week event for invited press and analysts disclosed the details of their upcoming Northern Islands series of GPUs [Radeon HD 6000 Series]. The Northern Island is a refresh of popular Evergreen architecture [Radeon HD 5000 Series], which again is an improved, DirectX 11-supporting version of their RV770 architecture.

Architectures by AMD, Intel and nVidia: We won't see 28nm until late 2011.
Architectures by AMD, Intel and nVidia: We won’t see 28nm until late 2011.

The company worked hard to improve their GPU [Graphics Processing Unit] and decided to play the renaming game. Thus, the first part is known under codename Barts and it is a successor of Juniper – thus a mainstream chip that targets nVidia’s GeForce GTX 460 card belonging to the Fermi architecture. What is interesting is that Barts is named HD 6800, while Juniper belongs to HD 5700 series. AMD shifted the number by one to make room for Fusion. In order to clear things up, you’ll see the comparison in the table below.

War for Performance / Watt and Dollar / mm2 metrics
AMD Northern Islands versus NVIDIA Fermi
AMD Northern Islands versus NVIDIA Fermi

Regardless of the performance war between Barts XT [HD 6870] and GF104 [GTX 460], where on default clocks Barts beats GTX 460 and gets beaten by cards such as Point of View TGT GTX 460 Beast Edition, the biggest battle is the one of die sizes. Just like in CPU world, AMD plays a very cautious game with the ballooning die sizes of the GPUs.

Using a Dollar/mm2 metric from Wedbush Morgan, it is visible that AMD commands higher retail $/mm2 ratio than nVidia. Barts commands $0.98/mm2, while GF104 only brings in $0.46/mm2. Same thing re-appeares with Cayman [HD 6900] which brings $1.05/mm2 to AMD, more than nVidia’s $0.94/mm2 on GF100. With this architectural refresh, AMD jumps ahead of nVidia in terms of potential profitability even with the mainstream part.

On the other hand, the price that nVidia pays for wafers at TSMC is lower than AMD, given the fact that nVidia is ordering more wafers and is a larger customer. In any case, AMD should have an easier time in a price war and be able to earn more money.

Then again, we might say, nVidia is allowed to order more wafers, while AMD is completely tied up with their preparation to launch Southern Islands, 28nm architecture that will come out of GlobalFoundries facilities in Dresden, Germany. This "rumor" got confirmed almost a year ago, when Dirk Meyer, CEO of AMD openly stated that AMD will start producing GPUs in GlobalFoundries as well when they hit the 28nm node.

Die Sizes:  Barts [HD 6800], Cayman [HD 6900] versus Fermi
As we exclusively revealed here, NVIDIA hides their own die sizes for a reason: they’re bigger than the competition. Even though profitability is determined by a combination of $/die and market share [Quadro, Tesla and GeForce all use the same silicon, just like AMD in their FirePro, FireStream and Radeon], comparing both camps turns unfavorable results for the Graphzilla from Santa Clara.

AMD Radeon HD 6870 is based on 40nm Barts GPUAMD’s Cayman [HD 6900] die is exactly 380mm2, i.e. 43mm2 more than previous gen Cypress [HD 5850/5870/5970], while Barts is 230mm2, significantly more than 166mm2 of its predecessor [Juniper].
With these two, AMD will create three product lines: Barts packs in HD 6800, Cayman will come as HD 6900 and two Cayman chips will be turned into Antilles – Radeon HD 6970 X2. Antilles is successor of HD 5970, and the die size comparison there is also quite favorable for AMD: HD 5970 featured 672mm2, while HD 6970 X2 will feature 760mm2.

In comparison, nVidia’s long time rumored dual-GF104 board [GTX 495] would have to beat HD 6970 using 734mm2 of silicon – something that can be called mission impossible, even after you count in the unpleasant fact that nVidia’s SLI technology [currently] scales much better than AMD’s Crossfire.

The problem with that measurement is that Barts is a head to head competitor for GF104, not Cayman [which targets GF100 / GTX 480].

The difference between Barts and GF104 is quite bit – 230mm2 versus 367mm2. Yet, Barts offers higher performance in computer games, which are the key measurement these two are measured.