AMD Quietly Reveals Next Radeon Series, Launches Next Week

84 Comments

View All Comments

• OneArmedScissorB - Friday, October 15, 2010 - link

  Radeon 3870 -55nm
  
  Radeon 4870 -55nm
  
  You fail.

• animekenji - Sunday, October 17, 2010 - link

  4870 was a huge jump in performance over 3870, ran cooler and was more power efficient. I don't see what point you're trying to make here?

• Voldenuit - Saturday, October 16, 2010 - link

  If architectural advances on the same node don't yield any performance advantage, why has intel been successfully executing their 'tick-tock' strategy for the past 4 years?
  
  Conroe was on the same node as Presler (65 nm).
  
  Bloomfield and Lynnfield were on the same node as Wolfdale (45 nm).
  
  Sandy Bridge will be on the same node as Gulftown (32 nm).
  
  You can't change the laws of physics, but I guess you also can't cut down on the number of idiots who continue to misquote and misattribute them either. -_-

• GeorgeH - Saturday, October 16, 2010 - link

  Presler was essentially a 65nm die shrink of a 90nm design.
  Conroe was designed for 65nm.
  
  Wolfdale was essentially a die shrink of Conroe to 45nm.
  Nehalem/Bloomfield/Lynnfield were designed for 45nm.
  
  Gulftown is essentially a die shrink of Nehalem to 32nm.
  Sandy Bridge was designed for 32nm.
  
  Sandy Bridge wouldn't work nearly as well at 45nm as 32nm.
  Nehalem wouldn't work nearly as well at 65nm as 45nm.
  Conroe wouldn't work nearly as well at 90nm as at 65nm.
  
  Similarly, Northern Islands won't work nearly as well as AMD intended at 40nm.

• Targon - Sunday, October 17, 2010 - link

  That is the very definition of the Intel tick-tock strategy. New design on old process, then new process with the same design, then new design on the last process, back and forth.
  
  Of course, there is a limit to how far this can go since there is an increasing degree of difficulty in getting a new fab process to work.
  
  Northern Islands/Southern Islands was initially set as a 32nm part, but TSMC issues changed that into a two phase design rollout.
  
  Now, due to the use of a 40nm process, AMD has had to cut out a bit of the design for the 6800 series due to thermal constraints. That doesn't mean that the new design is bad, but it does limit how fast it will be on this "next generation" part.
  
  We will see in a week how well the new design works.

• coldpower27 - Monday, October 18, 2010 - link

  Given it's been couple of months I hope the 6800's are good chunk faster then the 5800 and still within a good thermal envelope.
  
  The 5800 are already at 334mm2, on 40nm, and that already pretty efficient compared to nVidia's GPU. Though nVidia GPU does have alot of extra "stuff" that isn't needed in the consumer market.
  
  I just hope ATi can make things interesting again.

• Fergy - Friday, October 15, 2010 - link

  You might want to look up the Radeon 3800 and 4800 series. Both 55nm.

• GeorgeH - Friday, October 15, 2010 - link

  R600 (2/3xxx) was designed for 80/65nm and got shrunk to 55nm (the 80nm 2900XT was about the same fps-performance wise as a 55nm 3870.)
  R700 (4xxx) was designed for 55nm.
  
  Evergreen (5xxx) was designed for 40nm.
  Northern Islands (6xxx) was also forced to be designed for 40nm.
  
  Kind of a big difference. You can't make two good designs at the same node and expect revolutionary performance differences (as with the 4 and 5 series) unless you add heat, power, and cost for significantly more transistors in one design. 40nm is pretty mature now, so AMD will be able to push the boundaries more with the 6-series - just not to the extent we've grown accustomed to.

• JimmiG - Friday, October 15, 2010 - link

  It's definitely possible sometimes to significantly improve the performance without a die shrink. Take NV20 vs NV25 vs R300 for example. All built on the same 150nm process, but with a significant performance improvement each time. G80 to G92 was far less exciting in spite of a die shrink.
  
  According to Anand's RV870 article, the chip was originally meant to be much bigger, but it was changed in 2008 due to concerns about cost, yields etc. This last-minute change must mean that some parts of the chip are not as optimized as they could be since they had to throw out stuff in panic to stop the ship from sinking. It also means the design for a much larger, higher performing chip is already there. Now, with a more mature 40nm process, I don't see why they wouldn't resurrect that.

• GeorgeH - Friday, October 15, 2010 - link

  RV300 had almost 2x as many transistors as NV25 (~110M vs ~63M.)
  
  NV20 had ~57m, but IIRC at the end of the day the GeForce3 Ti500 wasn't really much slower than a GeForce4 Ti4600 in most games - after a quick search I found ~139FPS in Quake III vs ~159FPS @1280x1024. That's ~114% of the FPS for ~111% of the transistors.
  
  If you need X bits flipped and have Y transistors, a new architecture isn't going to make your task much faster unless your current architecture isn't well suited for flipping.
  
  If the 5xxx parts are much better at adding bits than flipping them, the 6xxx parts might be significantly faster - but I really don't think that's the case. That means any dramatic FPS-performance increase from NI will have to come with a similarly dramatic heat, power, and/or cost increase, minus the small benefits AMD can derive from a more mature 40nm manufacturing process.