- May 11, 2008
- 22,218
- 1,410
- 126
An FPGA from Altera will come to the market soon with up to 16GB @ 1TBps HBM memory while making use of Intel memory interconnects.
http://www.eetimes.com/document.asp?doc_id=1329776&
http://www.eetimes.com/document.asp?doc_id=1329776&
Last year, Altera provided a sneak preview of their forthcoming Stratix 10 devices (see Altera's Stratix 10 FPGAs & SoCs -- Breakthroughs in Performance, Integration, Density, and Security). Tempting and tantalizing as this sneak preview was, however, it was a little short on details.
Well, Altera has now thrown the veils open a little wider, providing more concrete details regarding the Stratix 10 MX family of System-in-Package (SiP) devices. First of all we have the core monolithic FPGA fabric boasting up to 2,005,000 logic elements (LEs) and up to 679,680 adaptive logic modules (ALMs) that can run at up to 1 GHz. This core fabric also features a quad-core 64-bit ARM Cortex-A54 subsystem and peripherals running at up to 1.5 GHz.
In addition to hundreds of GPIOs, we also have up to 96 high-speed transceivers. But the real cream on the cake is the (up to) 16 GB of high-bandwidth DRAM memory (HBM). The four blocks of HBM -- along with the transceiver die -- are connected to the main FPGA die using Intel's embedded multi-die interconnect bridge (EMIB) technology, which provides ultra-high-density, ultra-high-speed interconnect between the HBMs and the FPGA fabric. In fact, the four HBM blocks provide an aggregate bandwidth of 1 TBps, which is 10X the bandwidth that can be achieved using discrete, off-package DRAM solutions while consuming much less power (see also this video).
Having up to 16 GB of in-package memory with a bandwidth of 1 TBps is going to be a game-changer for a wide range of applications, including 8K video, high performance computing (HPC), computer vision, big data, radar... the list goes on. This also dramatically reduces the footprint, routing, and power consumption at the board level. Quite apart from anything else, designers are now presented with a completely new memory hierarchy and architecture. Ridiculous as it sounds, we might actually find it difficult to come up with ways to take full advantage of the memory bandwidth that is available to us.