AMD Launches FireStream Compute Accelerators Featuring Cypress Chip

ATI Cypress Finally Reaches Compute Accelerators

by Anton Shilov
06/25/2010 | 11:37 AM

Advanced Micro Devices this week updated the lineup of AMD FireStream compute accelerators with models based on the Cypress graphics processing units (GPUs). The new cards offer very high levels of performance amid relatively low power consumption. AMD expects the new accelerators to target supercomputer space.


“Heterogeneous systems in which high-performance GPU and x86 CPU technologies work in tandem can deliver enormous computational power. Industry standards like OpenCL are driving rapid adoption of heterogeneous architectures, and commercial customers deploying systems with AMD FireStream accelerators and AMD Opteron processors can immediately experience the benefits of the combined technologies,” said Patricia Harrell, director of Stream computing at AMD.

AMD announced two new FireStream compute solutions targeting different market segments this week. The AMD FireStream 9350 and 9370 GPU compute accelerators are scheduled to be available for purchase from AMD and its technology partners, including One Stop Systems and Supermicro, beginning in Q3 2010.  

It is interesting to note that the FireStream 9370 provides higher double precision performance than Nvidia Tesla M2050/M2070 (515GFLOPS) and over two times higher performance in single-precision operations (1.03TFLOPS).

ATI, graphics business unit of Advanced Micro Devices, released its high-performance code-named Cypress graphics processor on the consumer market back in September ’09, however, AMD itself postponed the release of its FireStream compute accelerators and will only start to ship them in the third quarter of 2010. According to Ms. Harrell, AMD needed time to certify the new AMD FireStream 9300-seris cards with manufacturers of actual servers. But even though the launch of the new compute accelerators took almost a year, the Cypress chip itself does have a number of enhancements for computing.

“We thought about APIs when we designed this chip: DirectX 11 for graphics and DirectCompute and OpenCL 1.0 for compute. Some of the features on the compute side are tailored for compute. For example, small arithmetical operations can be performed locally in memory to be more efficient; local and global data share and global synchronization all help with basic communications between threads; append buffers allows developers to efficiently build in memory database with an application. All of these features are designed specifically for compute,” said Ms. Harrell.

Even though the price difference between the FireStream 9350 and 9370 is gigantic, the more expensive model will still find its clients. For example, applications used in oil and gas industries need a lot of onboard GPU memory, hence, the 9370 with 4GB of GDDR5 is exactly what they need.

“Our customers’ face increasingly demanding enterprise data center requirements for power consumption, cooling, and floor space, and at Supermicro our goal is to deliver solutions that offer maximum performance per watt, per square foot and per dollar. By offering GPU compute accelerators in combination with powerful multi-core CPUs, we are keeping pace with our customers’ demands. The AMD FireStream 9350 and 9370 compute accelerators are a natural fit for our industry-leading server solutions,” said Don Clegg, vice president of marketing at Supermicro.