Today, Nvidia launched it's hitherto smallest member of the GF10x architecture and also the Fermi family dubbed Geforce GTS 450.
It is based upon a 15x15 mm ASIC called GF106, which boasts 4 Shader Multiprocessors (SM) in one GPC and features the same characteristics as the successful Geforce GTX 460 chip GF104. 48 ALUs - called Cuda-Cores by Nvidia - and eight texture units reside in each SM resulting in 192 ALUs and 32 TMUs across the chip. Also, GF106 has a GDDR5-interface comprised of 3 64 bit channels, each hardwired to an octo-ROP, albeit only two of those are active in currently announced GTS 450 products. Clock rates are 783 MHz for fixed function units and subsequently 1566 MHz for the programmable stuff aka shaders. The memory runs at an underwhelming 900 MHz (3600 GT/s). Double precision (DP) calculations are enabled in consumer parts and run at 1/12th the speed of single precision (SP), where the chip reaches 601,344 GFLOPS.
Gaming performance comes in roughly in the middle between HD 5750 and 5770 with current drivers over a broad set of games run at 1680x1050 and 1920x1200 with 4x FSAA and 16x AF enabled. Power consumption is roughly a tad above HD 5770 levels, except for idle, where the GTS 450 is on par or better than HD 5750. TDP for the card is 106 watts according to Nvidia.
I had the chance to run some benchmarks with TessMark 0.2.2 from my friends at ozone3d.net (check out a full description over there) in my spare time, so here are the (surprising ;-)) results: The Radeon HD 5870 has one mighty tessellation unit, straightly outperforming the Nvidia counterparts - too bad it's only a single one. But that could change with the coming HD6000 series arriving allegedly in october/november 2010 timeframe.
The real surprise though is that something prevents the mighty GTX 480 from achieving the same effectiveness per Tessellator as it's younger, smaller and tweaked siblings. The scaling within the newer Fermi-cores GF104 and GF106 seems fine though.
Update 15. Sept. 2010:
The newly added HD 5770 shows impressively that despite using an acutally useful resolution and adding anti aliasing, in Extreme mode and on Radeon graphic processors at least TessMark 0.2.2 purely depends on tessellation throughput.