AMD might just have released their latest Radeon branded graphics cards, but that isn’t the only upgrade they have in stock for us. Today they are releasing the new FirePro S9170 server GPU with an industry-leading 32GB memory for high-performance compute.
The server card is designed for DGEMM heavy double-precision workloads and with support for OpenCL 2.0. It is based on the second-generation AMD Graphics Core Next (GCN) GPU architecture and this new monster is able to deliver up to 5.24 TFLOPS of peak single precision compute performance while enabling full throughput double precision performance, up to 2.62 TFLOPS of peak double precision performance. That is some serious performance.
FirePro graphics cards aren’t something for gamers and if you came here to find an amazing game card, then you’ll be disappointed. GPUs are great with parallel workloads and can be used to offload those tasks from the CPU and achieve a great performance per watt. In fact, AMD is the market leader in that field and are ranked number 1 in Green when counting performance per watt.
The new AMD FirePro S9170 isn’t a replacement for the S9150 but rather thought as a complimentary card for those who need that extra memory to solve those extra big problems. It is the new flagship of server graphics cards and comes with ECC support.
We don’t find the new HBM memory here, which is a given due to the amount and HBM’s current limitation. Instead, it is the good old and well-performing GDDR5 that is used. The card is passive cooled and requires an airflow of 20 cubic meters per minute – which any current server setup should deliver. Users shouldn’t experience any thermal restrictions or throttling if cooled according to specifications.
The GPU clock of the S9170 got a slight boost over the S9150 and it now runs at 930MHz vs the old 900MHz. Shipments are scheduled for Q3 2015 and the street price will be slightly higher than the current S9150.
- Passively cooled solution for server environments
- AMD Graphics Core Next architecture
- 2,816 stream processors (44 compute units)
- 5.24 TFLOPS peak single-precision floating point
- 2.62 TFLOPS peak double-precision floating point
- Full throughput double-precision
- Error-correcting code memory support (external only)
- 32GB ultrafast GDDR5 memory
- 512-bit memory interface
- Up to 320GB/s memory bandwidth
- 275W maximum power consumption
- Support for SMBus temperature reporting at boot-up
- AMD PowerTune technology3
- AMD STREAM technology4
- OpenCL™, OpenGL support
- PCIe® x16 bus interface, PCIe 3.0 compliant
- Full-height/full-length dual-slot form factor
- Headless display support
- Linux OS support (64- and 32-bit)
- FCC, CE, C-Tick, BSMI, KCC, UL, VCCI, RoHS, and WEEE compliance
- Designed, built, and tested by AMD
- Planned minimum three-year life cycle
- Limited three-year warranty