Today Intel announced its the Intel Stratix 10 MX FPGA which has a few interesting features. As part of the Stratix 10 line, it has a high-performance (Altera sourced) FPGA onboard. What makes it interesting is that it has High Bandwidth Memory DRAM (HBM2) integrated. Integrating HBM2 with the FPGA addresses one of the primary bottlenecks many companies see with FPGAs, memory bandwidth. The Intel Stratix 10 MX has up to 512GB/s of memory bandwidth using HBM2 which is awesome.
Intel Stratix 10 MX FPGA with HBM2
In terms of market segmentation, the Intel Stratix 10 MX FPGA with HBM2 is targeted at HPC and big data analytics workloads (such as Apache Spark Streaming) where one can compress, encrypt, decrypt and accelerate data sets faster with FPGAs, leaving x86 CPU resources for compute.
Integrating the FPGA silicon with HBM2 uses Intel’s Embedded Multi-Die Interconnect Bridge (EMIB) technology. That means that Intel is able to use package micro bumps to provide high-speed links through the heterogeneous architecture. Intel’s Stratix 10 line is built upon its 14nm process and Intel is able to integrate non-Intel fabbed components (e.g. HBM2) onto a high speed interconnect using EMIB.
We see EMIB as a major enabling technology that will allow Intel (and others with variants of the technology) to make amazing compute platforms in the future. The FPGA space is specialized enough where this makes sense today. In the future, we fully expect to see architectures such as Intel’s x86 line with EMIB technology to integrate exotic memory technology or other accelerators. Beyond this, it could potentially be used to stitch together many x86 compute resources onto a single package while maintaining higher yields.
The FPGA industry is extremely specialized, but the EMIB technology finding its way into another application is a step in a direction we want to see. We did think it was interesting that Microsoft was not a major launch partner for the HBM2 enabled Stratix 10 MX since it relies upon high-speed memory in Intel FPGAs for its BrainWave architecture. See Microsoft Shows off Project BrainWave Persistent Inferencing from FPGA Cache.
Of course there’s no price. That’s not how Altera and Xilinx work. This isn’t a gaming GPU