AMD is expected to reveal the Radeon R-300 series any day now and all the leaks and rumors point to something pretty significant happening to the high-end cards – the world’s first GPU with stacked-on-die GDDR5 memory, increasing performance dramatically over the traditional memory controllers that we’ve had for years. 3DCenter has collected all the news and leaks so far about the R-300 and the resulting picture looks very interesting for AMD and fairly scary for Nvidia’s Maxwell graphics cards, at least if you don’t care that much about power consumption. Hit the jump for more info.
Looking at the GPU landscape now, things are very different from what they were three months ago. With the GTX 980 effectively replacing the GTX Titan Black, the GTX 970 challenging the R9 290X with only 3.5GB of VRAM and the GTX 960 putting the nails in the coffin for the GTX 760, it’s all a jumble of falling prices and bouts of fisticuffs between the two remaining GPU vendors.
AMD has been making waves of its own, dropping the prices of the Radeon R9 290 and R9 290X in the US and Europe to below GTX 970 levels, challenging the position of the GTX 960 as well. Given how things are at the moment, it looks like AMD was more making space for their new cards rather than just reacting to the GTX 970’s memory woes and the GTX 960 launch.
At the top of 3DCenter’s list is the rumored Radeon R9 395X2, codenamed Bermuda. 3DCenter reckons that the R9 395X2 will just be a doubling of the Radeon R9 390X’s specifications, though the mis-translated disclaimer at the bottom of the cell actually refers to not knowing if AMD will use a liquid cooler again to allow for two R9 390X chips. Two R9 390 chips would be much easier to cool.
Going down, the R9 390X and R9 390 based on the Fiji family have been rumored to be the first graphics cards shipping with High-Bandwidth Memory (HBM), which is soldered on-die to the GPU using a Through-Silic0n-Via (TSV) process. The memory still connects to the traditional memory bus paths, but the chips are stacked on one another to minimise the use of space and simplify construction. Samsung’s 850 Pro and Evo SSDs use similar technology for their flash storage and the use of TSV allows them to maximise the throughput of slower memory dies.
Though the surface area of the memory subsystem is smaller, AMD can design a 1024-bit bus in that space which gives a massive boost to memory bandwidth. If AMD will continue to use the 5GHz SK Hynix memory found on today’s Radeon R9 290 and R9 290X, memory bandwidth stands at…
5000MHz x 1024-bit / 8 = 640GB/s memory bandwidth
Which, given the figure, is astounding. Just double that for the R9 395X2 and AMD would have a monopoly on performance as far as memory bandwidth is concerned until Nvidia debuts their Pascal architecture in 2016, the only one where their future plans incorporate some form of HBM.
The R9 390X and R9 390 seem to be based on a chip family called Grenada, a rebranding of Hawaii that will probably feature some tweaks to the silicon, but will remain largely the same as Hawaii. We’ll probably see these chips being a little more efficient/less leaky as well as being able to boost to higher speeds for longer, and feature some of the colour compression enhancements found in the Tonga GPU family.
Ending things off, Tonga will probably see the R9 285 being rebranded as the R9 370X, while the R9 370 will probably feature a few less shaders and slightly lowered performance to match a lower price point. One change is that AMD will be allowing the manufacture of 4GB Tonga variants and may include 4GB of GDDR5 memory as standard on their reference designs. With AMD billing the R9 285 as their hero card for gaming at 1920 x 1080 or 2560 x 1440, it didn’t make sense to release a card for those resolutions that only shipped with 2GB of VRAM.
For the top of the R7-300 series, AMD could be rebranding their Pitcairn family, which would be super-weird given that Pitcairn is GCN 1.0 silicon (by now about three years old). Its more likely that AMD has transformed Pitcairn into Trinidad, bringing with it new features from all the GCN versions released thus far, including support for colour compression, Adaptive Sync using Displayport 1.2a+ and TrueAudio.
Nvidia still has the rest of the year to release the last of their Maxwell, graphics cards and they still have to fill in the mid-range and budget markets as well as replace the GTX Titan, Titan Black, GTX 780 Ti and GTX Titan-Z in the high-end. The next six months are going to be very interesting.