Intel & AMD Partnership For Single Package Chips – The Next Wave Of All-In-One Systems

A rumor about an Intel/AMD partnership has circulated for decades, with copious amounts of ifs and buts thrown in. Today we have confirmation of a partnership that may both surprise and delight, as Intel lays the groundwork for a new generation of on-package processors, that of third-party GPU integration with its CPUs.

It’s no secret that Intel’s graphical capabilities have been somewhat lackluster (understatement of the year), when it comes to modern computing and rendering capabilities. Intel does have a very strong hold on CPUs, but it’s always been third place when it comes to graphics. Laptop manufacturers have typically resorted to discrete graphics in their higher end systems, which adds complexity, power, and cost to the overall system. Wouldn’t it be nice to integrate an AMD or NVIDIA GPU on the same package as where Intel’s own GPUs have occupied?

Allowing third-party GPUs on the same package as Intel’s CPUs has been a pipe-dream for a long time; no two companies could come to an agreement as to how this could be done. It appears that Intel is relenting on this issue and now allowing other GPUs to share the same package space as Intel’s CPUs.

Thing is, this goes beyond just the GPU die, as we are looking at probably the first generation of single package processors, full system on chip deployments, or what Intel calls Embedded Multi-Die Interconnect Bridge (EMIB) technology. This is a single package of CPU, GPU and now, memory, with the inclusion of HBM2, so the GPU doesn’t share system memory.

Now, this is where things are a little fuzzy, and we’re waiting on confirmation, but the HBM2 memory may be just for the GPU only, and the CPU is likely to use normal DDR4, rather than use the HBM2 on-package – unless Intel has come up with a CPU that has an HBM2 memory controller. The use of HBM2 and AMD GPU likely means a Vega derivative is going to be used It would seem that Polaris architecture is to be used, rather than Vega, likely due to the better power efficiency, while AMD’s own APUs which will be coming out soon, will use Vega.

Conceivably, it’s possible that a full system integration, including HBM2 as a replacement for standard DDR4. It would take a lot less power, require less cooling, smaller footprint, thinner devices, and so on. Latency is comparable as well. So why isn’t HBM2 used for system memory? Plain and simple – cost. HBM2 is far more expensive than any other memory type on the market, be it DDR4, GDDR5 and GDDR5X.

If the memory were on package with the CPU and GPU, then there would be no way to upgrade without a full system replacement. For the most part, this actually isn’t a problem, since a lot of tablets and laptops solder the memory to the motherboard, so upgrades are impossible anyway. But the possibility of a full single package chip has been in the eyes of a number of enthusiasts for a while, and I would have guessed AMD could have pulled it off first (think of an AMD NUC, but even smaller, and graphical capabilities that are worthwhile).

There is some confusion though from this, as the partnership comes at a very odd time. AMD’s Ryzen launch put its CPUs back into a competitive space; it doesn’t really need Intel to create these chips. However, Intel still has a superior CPU in regards to IPC and power efficiency, which is something that AMD may be lacking, something that’s critical to power-sensitive devices like laptops. Whatever the case though, this partnership would have been in the works before the Ryzen launch anyway, and both companies have something to gain.

Of course, will we see NVIDIA GPUs on the same package as Intel CPUs? Well, reading the press release, the key line is “custom-to-Intel third-party discrete graphics chip from AMD’s Radeon Technologies Group” which might mean that AMD built the GPU specifically to Intel’s specifications, rather than just bolting on a standard Vega GPU die. AMD has a long history of producing application specific devices, as the last couple of generations of consoles will attest, while NVIDIA hasn’t been very forthcoming in producing such custom chips. There is still a possibility though, and I’m sure there would be a market for such integration, though it will likely be up to NVIDIA, rather than Intel.

No specific devices or SKUs were listed with this press release, as this is more of a technology announcement/partnership. The CPU aspect will of course be Intel’s latest generation Core architecture, Coffee Lake, but future CPUs will also be available as time goes on, so it’s best to see this as a framework or test, rather an actual product at this time.

Intel says that a closer look at the chips and systems from OEMs will be announced first quarter next year, so we may not see products until Q2 2018. Still, this has been a long time coming, and is genuinely exciting. I would hazard a guess that perhaps some of the first products out the gate would in fact be something like a new NUC from Intel.