Nvidia has kicked off the Supercomputing Conference 2013 in Colorado by announcing a new addition to its High-Performance Computing (HPC) GPU lineup, the Tesla K40.
The new card's standout feature is its 12GB of memory, which is over double that of its predecessor, the K20, which has 5GB.
The K40 also sports a few extra CUDA cores (2880 vs 2496), higher memory bandwidth (288GB/sec vs 208GB/sec), and a resulting faster peak performance. The K20 reached 1.17Tflops in double precision mode and 2.52Tflops in single precision mode, while the K40 boosts this to 1.43Tflops and 4.29Tflops respectively.
As with previous Tesla cards, the K40 will be available in both a passively cooled configuration for server environment and an actively cooled configuration for workstations.
With a base clock of 745MHz, the card can also boost to 810MHz and 875Mhz depending on workload. Because of this variable upper limit the card is specified as having a TDP of 235W, though it will run at between 160W-180W in most applications, according to Nvidia.
The news follows Nvidia's pre-emptive announcement last week of a new version of its HPC language, CUDA. The chief addition to CUDA 6 is a unified memory interface that should make it easier for developers to program GPU-accelerated software.
AMD also recently unveiled its latest HPC card, the FirePro S10000 12GB Edition. It also features a hefty 12GB, up from 6GB on the original S10000. However, it is quite a different beast to the K40. Where Nvidia's card uses a single GPU and the card is solely meant as a co-processor, the S10000 uses two GPUs and can actually be used as a graphics card (i.e. video output device) too.
AMD claims performance figures of 1.48Tflops peak double precision and 5.91Tflops single-precision, with each GPU given 240GB/s of memory bandwidth. Its card is much more power-hungry than Nvidia's solution, though, with a TDP of 375W.
The Tesla K40 will be available later this month, some months before AMD's latest cards becomes available. Although normally sold in bulk for undisclosed sums, individual Tesla K40s are already starting to hit retail with prices of around $5,500.
The new card's standout feature is its 12GB of memory, which is over double that of its predecessor, the K20, which has 5GB.
The K40 also sports a few extra CUDA cores (2880 vs 2496), higher memory bandwidth (288GB/sec vs 208GB/sec), and a resulting faster peak performance. The K20 reached 1.17Tflops in double precision mode and 2.52Tflops in single precision mode, while the K40 boosts this to 1.43Tflops and 4.29Tflops respectively.
As with previous Tesla cards, the K40 will be available in both a passively cooled configuration for server environment and an actively cooled configuration for workstations.
With a base clock of 745MHz, the card can also boost to 810MHz and 875Mhz depending on workload. Because of this variable upper limit the card is specified as having a TDP of 235W, though it will run at between 160W-180W in most applications, according to Nvidia.
The news follows Nvidia's pre-emptive announcement last week of a new version of its HPC language, CUDA. The chief addition to CUDA 6 is a unified memory interface that should make it easier for developers to program GPU-accelerated software.
AMD also recently unveiled its latest HPC card, the FirePro S10000 12GB Edition. It also features a hefty 12GB, up from 6GB on the original S10000. However, it is quite a different beast to the K40. Where Nvidia's card uses a single GPU and the card is solely meant as a co-processor, the S10000 uses two GPUs and can actually be used as a graphics card (i.e. video output device) too.
AMD claims performance figures of 1.48Tflops peak double precision and 5.91Tflops single-precision, with each GPU given 240GB/s of memory bandwidth. Its card is much more power-hungry than Nvidia's solution, though, with a TDP of 375W.
The Tesla K40 will be available later this month, some months before AMD's latest cards becomes available. Although normally sold in bulk for undisclosed sums, individual Tesla K40s are already starting to hit retail with prices of around $5,500.
MSI MPG Velox 100R Chassis Review
October 14 2021 | 15:04
Want to comment? Please log in.