Last week, Intel announced and demoed a 28-core CPU running at 5GHz. Since then, it’s been discovered that the original part Intel demoed was overclocked to hit these frequencies, that the company was using a water chiller capable of dissipating 1770W to cool the CPU, and the entire demo now looks much more like Intel was attempting to take the wind out of AMD’s 32-core Threadripper announcement than any attempt to launch a series product.
Let’s back up and review. Intel already makes a series of 28-core CPUs, including the 28-core Xeon Platinum 8180. That chip has a 2.5GHz base clock, a 3.8GHz boost clock, and a 205W TDP. Intel TDPs are derived from base clocks only, so we already know that the default version of the core is going to be drawing a significant amount of power. There’s simply no way around it — frequency gains eat power quickly when you’re driving 28 cores in the same socket. WikiChip claims that the full-core boost clock on the Xeon 8180 is 3.2GHz, but take this with a grain of salt since Intel doesn’t formally release these numbers anymore.
The 5GHz CPU that Intel showed at Computex last week already came with some obvious caveats attached. It debuted in a special LGA3647 socket with four 12V CPU power connectors on the motherboard. That’s the first clue that we’re dealing with a truly massive power draw — one of these eight-pin connectors is standard on a high-end board and two of them on the highest-end boards is typically a bit of overkill. The cooler in question — an Hailea water chiller — is another sign that this CPU has truly mammoth power dissipation requirements.
Image by Anandtech
Now it’s possible that Intel was just hedging its bets, but that’s unlikely. Anandtech claims that a Core i9-7980XE at 4.9GHz draws a kilowatt, which means the idea of a 28-core chip at the same clock speed hitting 1.7kW isn’t crazy. The need for four 8-pin power connectors and the use of the enormous cooler further emphasize the issue.
And math isn’t in Intel’s favor, either. If each of the 28 cores had a 99 percent chance to hit 5GHz when manufactured, the chance of getting a CPU with 28 cores that can hit 5GHz is around 75 percent. If each CPU core has a baseline 90 percent chance to hit 5GHz, the chance of an entire 28-core chip hitting that frequency is just 5 percent. Obviously CPU yields are a bit more complicated than this, but napkin-math is good enough for our purposes, particularly since it’s a well-known fact that CPUs with higher core counts struggle more to hit the same frequencies on all core workloads as lower-count chips do. It’s just harder to make certain that every single CPU core can hit the same overclocked frequency.
Put the clues together, along with Intel’s admission that it never told anyone at Computex that the CPU was overclocked, and it’s obvious that standard air-cooled CPUs are never going to be able to hit and maintain these all-core boost frequencies. The power consumption is simply too high. Combined power for CPU and cooling could be over 2kW, and that means working with multiple power supplies in a single chassis — not something most enthusiasts or even most workstations ever offer. This chip, as currently constituted and shown, is built for the niche of a niche of a niche market. And after all of that, you’d better have one hell of a circuit to plug it into — because 2kW is enough to blow your average 15 amp circuit breaker. If you’re going to run it all-out, you’d best be plugged into 20 amp, minimum.
So what would we ultimately conclude, following Intel’s announcement? The company is working on faster Xeon 28-core CPUs. They aren’t going to run at an all-core boost speed of 5GHz, because that clock rate isn’t going to be sustainable unless Intel intends to help create a new ecosystem for a single motherboard with multiple power supplies and water chillers with their own substantial price tag and offer complementary home rewiring. Spoiler: It isn’t. And as a barricade against AMD’s 32-core Threadripper announcement, this move was a bust. If you’re going to block a competitor launch with a product of your own, you’d best be able to deliver. And Intel can’t deliver the part it demonstrated at Computex, because the laws of physics don’t allow it.
We don’t know exactly how Threadripper’s 32 cores will perform against Intel CPUs, and there are genuine questions about how well the CPU will scale given its use of just four memory channels. But Intel ultimately shot itself in the foot with this announcement. By choosing to prioritize a 5GHz overclocked chip it’ll never actually launch, it muddied the very real question of how its 28-core chips will compete against AMD and what kind of features or capabilities Skylake-X users should expect from a potential debut later this year. A more realistic demo could’ve whetted consumer appetites for cores they’ll actually be able to buy a few months from now. The 3.4GHz boost clock AMD is apparently targeting for these chips isn’t very high and Intel has historically hit higher frequencies.
Instead, we’re left with a rather confusing muddle. Intel is apparently building a new 28-core chip. That chip may even be capable of a 5GHz boost clock — on a core or two. But the chances that we’re going to get a 28-core 5GHz CPU, given what Intel went through to get the system working in the first place, is effectively zero.
An Intel spokesperson has contacted ET to state: ““The 28C demo at the keynote is a real product in development targeted at the high end prosumer and enthusiast audience. Intel continues to optimize design and process across its products, and the demo showcased an upcoming Intel product having the capability of 5.0 GHz overclocking across all 28 cores.”
Absent any guarantee of that capability, the chances of materially hitting it are extremely low.