Last week, we covered the successful delidding of AMD's Threadripper CPU. It's not for the faint of heart, given that it requires multiple hot air soldering stations and a variety of razor blades, and the expert overclocker who managed to remove the heat spreader still killed the chip.
Even so, der8auer's findings left us a touch confused. What the delid showed was that AMD had put four dies on the package, but only two of them were apparently active, as opposed to using four cores across each die for a total of 16. We've since spoken to AMD about this, as has Overclock3D, and we can explain what's actually going on.
The explanation is straightforward: The reason Threadripper has four dies under the hood with only two in use is because leaving them off would create mechanical instability when the heatsink is mounted. AMD apparently needs four dies in that arrangement to stabilize the packaging. This does make sense, given some of the esoteric cooling solutions with high pressure attachments that overclockers use.Why Not Use All Four Die?
There are several reasons AMD isn't using four 2+2 configurations. First, 2+2 would increase latency and hit gaming performance, as data has to be passed across the Infinity Fabric. So far, tests have shown that configuring Ryzen 7 1800X in a 2+2 or 4+0 configuration has a fairly limited impact on performance, but that could change when core counts rise to the level Threadripper offers.
The other important thing to know about Threadripper is that the other two die apparently aren't actually chips at all. They're apparently structural inserts required for support, but do not represent "bad" Epyc cores or any kind of core recycling. OC3D also states(Opens in a new window) that the two active die in Threadripper are on the diagonal from each other.
Image by der8auer
The die arrangement makes sense -- using the two die with the most space between them ensures that hot spot formation will be minimized. And there's still likely some truth to our speculation as to why Threadripper and Epyc share so much infrastructure (though it's not clear if Epyc CPUs can be used in X399 motherboards, or if Threadripper would function in an Epyc board).
Either way, we can see a common pattern to what AMD is doing. Ryzen, Epyc, Threadripper -- these chips aren't just built on a common CPU core, they're built according to a fairly rigid set of design specifications. Intel tends to offer a more varied range of products, with various features fused off depending on which market segment each belongs to. This can make it difficult to buy a chip that supports all of the features you want.
AMD went a different, simpler route. The same CCX configurations that power Epyc also power Ryzen 3, though these chips are tested to different specifications and intended for vastly different markets. As such, ramping Threadripper and Epyc at the same time, on a common platform, is probably a major cost-saver for the company. When you're trying to break into major markets you haven't had a share in for 6-7 years, it's extremely important to leverage economies of scale, even if the ideal solution would be a separate CPU socket or set of products.
Feature image by der8auer
