Intel sees no point in sticking dies together to increase the number of cores

Intel continues to develop large monolithic processors to increase core count. According to the company, sticking dies together creates latency, with Intel appearing to refer to AMD’s approach.

Increasing the number of cores on a monolithic processor has the advantage of lower latency, said Guy Therien, chief architect at the Intel Client Computing Group for the performance segment. ‘You have heard of that latency in the approach to obtain a large number of cores’, according to the chip architect. With this, Therien seems to refer to AMD, which connects four dies with its Threadripper and Epyc processors to achieve a maximum of 32 cores. In presentations, Intel has previously emphasized the difference in die-to-die latency between its approach to monolithic Xeon processors and AMD’s method.

According to Therien, Intel’s method ensures more consistent performance. He also emphasizes that for most consumers, more cores are not always better because most applications do not use them anyway. “If you don’t need more cores, it can actually be negative because more cores means more heat.” As an example of a niche in client computing that could benefit from it, Therien cites 3D rendering workstations, simulations and 360-degree videos.

Intel slide on the better latency when using less dies

Intel itself also connects several dies on a single processor. The company has high hopes for its emib, or embedded multi-die interconnect bridge. The purpose of this is not to increase the number of cores, but to combine different chips such as fpgas, x86 processors and memory, and different production processes on a single package. Intel’s first Core 2 Quad was also a multi-chip package with two dies, each with two cores.

AMD’s method of linking different dies, via the Infinity Fabric interconnect, allows the company to compete again after a long time in the server market, where Xeon chips have long ruled, and in the market for systems that benefit from many cores. AMD can do this in a scalable way at a relatively low cost and with less risk of poor yields. For example, AMD can relatively easily switch to 7nm for its Epyc processors with possibly more than the current maximum 32 cores.

AMD slide on the advantages of a 4-die multi-chip module versus a single-chip variant