AMD today is putting out a server oriented GPU called the S9170 with 32GB of GDDR5. This level of memory would be slightly overkill for 8K gaming but it is only closer to adequate for some HPC tasks.
On the spec side we can’t say much because the press release we have on it was basically info free. We can say the S9170 has a 512b GDDR5 interface and sports 320GBps of bandwidth, both strongly suggesting it is based on Hawaii or its Caribbean Islands equivalent called Granada. If this is true the claimed 5.24TFLOPS of SP compute suggests a clock hovering around the 930MHz mark, down from 1050MHz in Grenada. DP is half rate on this one or a claimed 2.62 TFLOPS, a number that easily obliterates anything else on the market from AMD or Nvidia.
One reason for the clock drop is power, 32GB of GDDR5 is going to suck a lot of energy. Since the S9170 has a 275W TDP, the same as a 1050MHz consumer Granada, that gives you an idea of the tradeoffs on clock vs GPU energy vs GDDR5 draw. In short the S9170 is nothing radical, just about what you would expect it to be. On the record front AMD now holds the crown for the largest GPU memory space on a single or dual card, the fastest DP performance on a single or dual card, and the only GPU named after the island of Granada. That last one isn’t all that important to the HPC set but we would be remiss if we didn’t mention it.
We can’t say how much it costs, how the memory is arranged, or other technical details yet, but we hope to bring those, and a picture, to you soon. More when we get it.S|A
Latest posts by Charlie Demerjian (see all)
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018
- Intel can’t supply 14nm Xeons, HPE directly recommends AMD Epyc - Sep 7, 2018
- AMD reintroduces the Athlon name with two CPUs - Sep 6, 2018