I’ve received word from a reader that some Germans wrote a story containing details about RV870, e.g. Radeon “5870”. Neoseeker brought the translation forward , and while some parts make a lot of sense, some really don’t.
First of all, the RV870 is supposed to be a 40nm part, but that’s not something that we didn’t know already. Both Nvidia and AMD are going to bring 40nm half-node parts first, followed by 32 and 28nm full-nodes. According to the story, the GPU is supposed to contain 25% more shaders than Radeon 4800 series, bringing the theoretical computational power to 1.5 TFLOPS.
Well, you don’t need 25% more shaders to get 50% performance increase. Radeon 4800 showed the path that the company is going to take, and the name of the game is how to increase the performance of those 10 shaders that now sit in one “pipeline” (or shader cluster), and increase the capacity of scratch cache to enable faster GPGPU computation.
The alleged die size is 205mm2, and that would go in-line with die-shrink of 4800, which would be roughly 170mm2 if it was manufactured in 40nm (instead of actual 256mm2). 30-35mm2 should be enough to slap around 1000-1200 shaders, if those rumors are true.
However, there is just one thing that does not hold ground in the story – and that is that RV870 should use 512-bit memory interface and GDDR5 memory. I may be forced to eat my own words, but no, ATI RV870 will not bring 512-bit memory controller. RV870 will feature much improved 256-bit memory controller, and it will offer bandwidth of some 150-200 GB/s per GPU. When you combine the two GPUs, possibly on the same substrate, you will get 512-bit memory controller… in a way. 512-bit memory controller with current GDDR5 memory (900 MHz QDR, e.g. 3.6 “GHz”) yields 230 GB/s. And that is the amount of bandwidth GTX280 would have if nV went for GDDR5 instead of older GDDR3 memory.
Nvidia’s next-gen part will however, bring 512-bit memory controller coupled with GDDR5 memory, offering insane amount of bandwidth – 200-250 GB/s, to be more precise.
Dual-die GPU is a good idea, but can TSMC pack the chips like Intel can?
It will be interesting to see can TSMC pack two RV870s on the same substrate, or that idea will go the way of do-do birds. We’ll see.
Oh yeah, cooling will be vapor chamber, and we should see some really interesting cooling designs. AMD already got its feet wet with vapor-chamber technology (both 4870X2 and Sapphire’s Atomic 3870 and 4870 come with vapor chamber cooling).
According to the story, the codename for the 5800 board is Lil’ Dragon. However, claim about DX11 being ready in summer of 2009 is something that I would take with a big grain of salt. During PDC’08, held last week in LA, there were talk that Microsoft will even send Windows 7 to manufacturing without DirectX 11, putting 10.1 until DirectX 11 shows up at later date. My sources compared the situation for 2009 equal to the one in 2002, when ATI shipped DirectX 9 part five months before DX9 came out.
As it stands right now, both Nvidia and ATI will have their DX11 parts ahead of actual API, giving developers enough time to optimize their respective drivers. Let’s hope for the best.
