The smartest thing about manufacturing Discipline Programmable Gate Arrays (FPGAs) is you could make the silicon very huge. The nature of the repeatable unit design can soak up points with a course of know-how, and in consequence we regularly see FPGAs be the most important silicon dies that enter the marketplace for a given manufacturing course of. Once you get to the restrict of how huge you can also make a chunk of silicon (generally known as the reticle restrict), the one solution to get larger is to attach that silicon collectively. At the moment Intel is asserting its newest ‘large’ FPGA, and it comes with a reasonably large milestone with its connectivity know-how.

One of many parts driving this business ahead is packaging know-how. We’ve lined intimately parts like TSMC’s 2.5D Chip-On-Wafer-On-Substrate (COWOS) packaging utilized in GPUs, Intel’s embedded multi-die interconnect bridge (EMIB), stacking know-how like FOVEROS, and as we migrate into smaller chiplet primarily based silicon, every will develop into essential to discovering one of the simplest ways to supply the top chip that goes into one million techniques.

Regardless of Intel’s finest diagrams about its EMIB know-how, displaying many die linked collectively from a number of totally different course of nodes, one main barrier has eluded the corporate. Till this level, all we had seen from EMIB was it connecting one high-powered die, like a GPU, to a low-powered die, like HBM. One of many criticisms to having solely identified Intel merchandise join one high-powered and one low-powered die is that if the EMIB connection wasn’t thermally secure to face up to power-cycling between two die.

Intels EMIB Now Between Two High TDP Die: The New Stratix 10 GX 10M FPGA

One in every of Intel’s mock-ups of how future processors would possibly appear like

When connecting two die collectively in a substrate, particularly high-powered die with vias or a BGA design, mechanical stresses must be taken under consideration, particularly if totally different metals are at play. Thermal growth and contraction is a vital level of failure, particularly when coping with embedded and lengthy life-cycle designs. Not solely the growth and contraction of metals, however when coping with natural substrates holding the packaging know-how, making the substrate inextricably skinny additionally severely will increase long-term feasibility issues, particularly if high-powered die are used for connectivity.

With Intel’s new FPGA, the Stratix 10 GX 10M, the priority appears to have disappeared. This new product, designed as an enormous FPGA for the ASIC prototyping and emulation market, combines two giant 5.1M logic aspect FPGAs with three EMIB connections, producing an total chip with a mean TDP from 150W as much as 400W with superior cooling.

Intels EMIB Now Between Two High TDP Die: The New Stratix 10 GX 10M FPGA

A complete of seven EMIB connections, but it surely’s the three within the center which are the milestone

The ASIC prototyping and emulation market, whereas a small market income sensible (Intel said ~$300-$500M/yr), is at all times requesting larger and greater FPGAs so as to have the ability to match increasingly of their ASIC designs onto as few FPGAs as attainable with a view to get essentially the most correct outcomes. These chips in the end find yourself operating at a low frequency for accuracy, anyplace from 50 MHz to 300 MHz, however Intel states that this new Stratix 10 GX 10M design can simply substitute 4 of its previous GX 2800 FPGAs with double the connectivity and even a 40% energy discount for a similar workload.

The design of the FPGA is round these two 5.1M logic aspect dies, linked along with three EMIB connections. These use the AIB protocol operating at over 1 GHz, and type a part of the 25920 connection pins throughout the entire chip which has one other 4 EMIB connections to transceivers as proven within the diagram.

Intels EMIB Now Between Two High TDP Die: The New Stratix 10 GX 10M FPGA

At 10.2 million logic parts, this FPGA eclipses the Xilinx VU19P introduced in August which had 9 million parts (8172okay flip-flops, 4086okay LUTs). The Stratix 10 GX 10M additionally accommodates 6912 DSP parts and 48 transceiver outputs at 17.Four Gbps. Intel states that these are designed primarily as PCIe 3.0/4.Zero help, and that the FPGA helps H-tiles for connectivity for purchasers eager about customized designs.

Intels EMIB Now Between Two High TDP Die: The New Stratix 10 GX 10M FPGA

The launch of this new {hardware} coincides with Intel’s FPGA Tech Occasion in China, which is among the main markets for this product. Intel states the {hardware} has already been with key companions virtually a 12 months (one of many early prospects is China primarily based), however is now in manufacturing for the broader market. On the subject of Intel’s excessive calls for for its 14nm components, the corporate said that the amount isn’t that prime for this form of product, they usually received’t have any points. The firm (at the very least, the FPGA a part of the corporate) did clarify that using EMIB on this style implies that their two-die strategy assists with yield.

Personally, the truth that Intel is strapping two excessive powered die (~75W to ~150W every) collectively utilizing a number of EMIB connections is a key step into driving the EMIB know-how to the broader market. With this as a proof of idea, it paves the way in which for higher multi-die CPU designs in addition to the promise of EMIB (and Foveros) in future discrete GPU merchandise.

Associated Studying



Please enter your comment!
Please enter your name here