The 9800 GTX+ is NVIDIA’s first foray into the world of 55 nm chips. While it does not add anything new to the old 9800 GTX, it does run faster and slightly cooler/more power efficient.
While NVIDIA obviously was planning on utilizing these smaller process nodes, it is likely that their schedule for transitioning to 55 nm was brought forward after they saw the success that AMD had with taking the R600 architecture, optimizing it, and transitioning it to 55 nm as quickly as they did. The latest round of graphics cards almost seems a verification that NVIDIA has been a little too conservative about new process nodes when we see that the 575 mm square + GTX 200 series of cards barely surpass the performance of the 276 mm square RV770. While NVIDIA has relegated a larger transistor percentage into making the GTX 200 a more effective general processing unit, the HD 4800 series do not give up much in terms of overall 3D performance as well as some impressive numbers when it comes to single and double precision floating point operations. This summer NVIDIA did release their 9800 GTX+ cards based on their first 55 nm part, the G92b. NVIDIA had promised that it would be the first part of July when we would see these parts, but it was actually in August when we first started to see them hit the streets. A few weeks is not a huge delay, but it certainly did not do NVIDIA any favors when considering that AMD had their Radeon 4850 out in force.
At the beginning of this summer, when the GTX 200 series of cards were released, we started to hear that NVIDIA would be quickly transitioning their GTX 200 production to 55 nm. Some people refrained from buying cards because they expected these parts to start leaking out around early August if some rumor sites were to be believed. Well, it is now in the first week of September and I personally have not heard a single thing about a 55 nm GTX 200 derivative. And I am not entirely sure that NVIDIA will have one out anytime soon.
While not exactly another 9700 Pro, the HD 4800 series have been giving the folks over at NV a headache. It also forced NV to drastically cut the prices on the new GTX 200 series of cards, which became a bit of an unpopular move with their partners and forced the companies to give early adopters a partial refund.
The original G80 chip was big and was designed from the ground up to be a 90 nm part. The G92 was a partially redesigned G80 that was adapted for use on TSMC’s 65 nm process. The GTX 200 was also designed from the ground up as a 65 nm part, but its complexity is far greater than that of the G80 or G92. NVIDIA almost doubled the amount of custom designed portions in the GTX 200 (going from 128 SPs to 240 SPs) as well as a much more complex cache heirarchy. My gut feeling here is that NVIDIA is having a harder time porting the GTX 200 design to 55 nm than originally thought. TSMC’s 65 and 55 nm processes are quite similar, but from my understanding TSMC did change a few things around at the 55 nm node. So this is not exactly a case of an optical shrink, but there are also some physical material differences that must be accounted for in the design of a new GPU. It looks like the original G92b transition went “fairly” smoothly, but it still was not exactly on time. Now we are dealing with a part with revised SPs as well as the massive 1.4 billion transistor count, and I truly believe that NVIDIA is having a hard time getting the 55 nm versions out the door.
If there was ever a chip that needed shrunk, it is the GTX 200. By going to 55 nm they should achieve around a 18% decrease in die size. So instead of the 576 mm square the new chip should be around 488 mm square, which is right around what the old G80 was in terms of die size. The second issue we may be seeing is that yields on their 55 nm process with this new, still large part, could be holding NVIDIA back from releasing these cards in a timely manner. This may not have been a big problem if the prices on the GTX 280 and 260 parts had stayed where they were initially, but because of the price drops NVIDIA cannot afford to release a new part at those prices if there is not enough product to fulfill demand.
The stock 65 nm GTX 280 gets awfully close to the theoretical 225 watt power limit that can be supported by PCI-E and the 6 pin/8 pin power combination. Even a half node shrink could do wonders for this product.
I could certainly be wrong here, and that 55 nm GTX 200s are rolling off the line as we speak and are being implemented into current products without us knowing it. That is likely not the case though, as we saw with the introduction of the G92b parts with the GTX+ moniker.
The second part of this post deals with the latest leak about dual core Phenom parts being released by AMD. It appears that the first parts coming to the market will be clocked at 2.3 GHz at the highest, yet AMD is positioning this part as a faster unit than the previous 3.2 GHz Athlon X2. While the Phenom design is faster per clock than the older Athlon, it is not that much faster. We also know that the dual core Phenom, while based on the Phenom architecture, was worked on by a different design team than the quad core Phenoms. This part was always intended to be introduced later than the quad cores, with the older dual core Athlon 64s handling that particular market until the dual core Phenoms were released. Putting 2 + 2 together here, we can likely assume that there are some signficant design changes with the dual core Phenoms as compared to the B3 revision quad core Phenoms. I am guessing that many of the architectural enhancements that will make their way into the upcoming 45 nm parts were first implemented into the dual core Phenom design at 65 nm. So while in fact these parts are running at 2.3 GHz, their performance will exceed that of a dual core Athlon 64 part running 900 MHz faster. We also must figure that the dual core Phenom will be faster per clock than the current quad core Phenoms. So if a user only is focusing on single and dual threaded applications, then the dual core Phenom may be a better buy for the money. This is being positioned as a budget enthusiast/gamer’s chip for the time being.
Does the 65 nm dual core Phenom have more in common with the upcoming (though still distant) 45 nm Phenoms? I’m guessing that it does.
So while some folks may be running benchmarks with B3 Phenom chips with two cores disabled, those scores will likely not accurately simulate the performance of these upcoming dual core Phenom parts from AMD. We also have no idea how much overclocking headroom we have with these chips, but from everything that I have heard these chips will not exactly be powder puffs when it comes to performance.