Specifications for AMD’s upcoming Polaris 10 lineup of GPU’s has leaked onto the internet thanks to SiSoft Sandra’s benchmarking database.
We’ve heard a lot of rumors over the past few weeks that HBM2 will not debut with the Polaris 10 or Polaris 11 cards released later this year, and instead we’ll see HBM2 appear on Vega in 2017. These recent SiSoft leaks seem to confirm that.
The entry shows the Polaris 10 will feature 8GB GDDR5(x) memory on a 256-bit memory interface and is packed with 2,304 4th generation shaders (which would mean 36 Compute Units, assuming we still see 64 shaders per CU).
At first glance, the most shocking part of the leak is that the GPU core runs at only 800MHz, which assuming SiSoft isn’t reporting an incorrect value (perhaps GPU boosting, or just mis-reading the real Polaris 10 clock value), likely indicates this leak is of an engineering sample, and the full retail silicon will probably be over 1Ghz in speed.
What’s really strange is that the card – even at these conservative clock speeds, beats an R9 290X graphics card. It’s therefore reasonable to guess that at least in the tasks SiSoft ran, Polaris is faster in a clock for clock basis.
The memory configuration is listed as 6Gbps GDDR5, which offers just 192 GB/s bandwidth thanks to a rather narrow 256-bit bus (the same found in the Tonga lineup of cards). This bandwidth is half what you’d find in many enthusiast cards from AMD (such as the R9 290X or the R9 390X) and it’s likely not going to feature in the retail card. Instead AMD will probably shift the production to GDDR5X.
AMD’s Moore’s Law Example shows how mainstream cards provide better performance per dollar than bleeding edge hardware
One of the benefits of GDDR5X is the double bandwidth per pin compared to the older technology, and the fact the memory controllers require very little modification to switch between the two technologies. This, and the lower power requirements and the only marginally higher pricing leads to the GDDR5X being an excellent stand-in until HBM2 memory becomes readily available by the end of the year in sufficient DRAM sizes.
It’s difficult to know which Polaris 10 this GPU actually will be, as its specs could either be the R9 490 or a potential R9 490. Because we don’t have a comparison point, it’s purely guesswork where this card fits in. It is possible AMD may wish to add more of a distinction between their cards than what was found between the R9 390 and R9 390X lineup.
This certainly could be seen in the R9 290 vs R9 290X fiasco too, with many users feeling the price point of the 290 offered the better deal between the two GPU’s. By after Raja Koduri had spoken so much of bringing back the Performance Per Dollar during their GDC conference, it’s anyone’s guess what happens in reality.
Assuming this configuration is similar to the one we saw during AMD’s conference during GDC, it means the card is capable of running something like Hitman at 1440P 60FPS at max settings. Unfortunately it’s difficult to know for certain whether the card at that event was similar to the one we’re seeing leaked here, or a lower end model (or with higher / closer to retail final clocks).
So there you have it, we’ve at least got an idea to the amount of RAM and the number of shaders of one of the cards in the Polaris 10 lineup. Assuming clock speeds are ramped up about 25 percent and memory clocks increase in kind, we should see a damn impressive card when it’s finally available to the public at large.
