One of the topics we’ve been covering in the past month is the behavior of the Core i9-7900X when overclocked, as well as the performance characteristics of the X299 chipset. In both cases, Intel has been criticized. First, for a CPU that traps a great deal of heat below the heatspreader, thanks to the use of thermal paste instead of solder, and second, for X299 board specs that apparently allow motherboard manufacturers to use poor VRM cooling.
It’s not clear who’s actually responsible for the latter issue. It could be that Intel didn’t provide the necessary platform documents within a reasonable manufacturing window, or it could be that motherboard manufacturers made their own assumptions about product specs and wound up with VRMs that run too hot for their own on-board cooling solutions. Either way, it’s not a great look.
Now, veteran high-end overclocker der8auer has moved over to the Threadripper side of the equation. The first thing to know is that AMD, unlike Intel, is specifically using solder in Threadripper, whereas Intel’s 10-core Core i9-7900X doesn’t use it. The second is that Threadripper has a different CPU configuration under that solder than we would’ve expected.
According to der8auer, who spoke with AMD, Threadripper is an Epyc CPU configuration with four dies, each packing eight cores. At first, we thought this meant AMD had gone for a similar arrangement to what it used for Ryzen 3: four cores active on each die (with SMT enabled in this case) and 16 cores/32 threads on the entire chip. But that’s not the case if der8auer’s information is accurate. According to him, Threadripper only uses two of the four Ryzen dies mounted on the PCB. The other two aren’t functional. His video on the topic is embedded below.
This seems rather odd for a host of reasons. Generally speaking, the entire point of using an MCM (Multi-Chip Module), as opposed to a unified die, is so you can leave chips off the product. If you think about it, this makes perfect sense: You don’t want to build a 32-core part only to find out that your yield at that enormous die size and power consumption is too low to make the chips profitable. It’s economically much safer to build and mount 4×8-core CPUs than to build one unified 32-core CPU. If AMD were offering Threadripper with just two cores active in every CCX, it would probably help with hot spot formation and allow AMD to selectively choose the very highest clocking cores. Instead, we’ve got a situation where it apparently made sense for AMD to mount four cores, but only activate two of them.
The only explanation I’ve come up with is that this is related to AMD’s economies of scale and manufacturing costs. AMD doesn’t own its packaging and test facility anymore, and its Epyc design costs aren’t going to be immediately replaced by Epyc revenue. AMD has stated that while it’s targeting a 10-percent server market share, it doesn’t expect to hit that this year. By tossing expected Threadripper sales in alongside Epyc sales, AMD may have cut its overall costs by improving its economy of scale.
Besides the use of four dies instead of two, we got confirmation AMD is using indium solder for its dies, while gold-plating the “back” of the heatspreader to ensure proper thermal cycling. This kind of approach is standard for higher core-count chips, but Intel has been reserving it for Xeon these past few years. Whether Threadripper’s 16-cores can handle their top clocks as well as Xeon is something we’ll have to wait and test, but AMD definitely opted for the top-drawer solution on these parts. Der8auer also explains why a properly soldered chip should look the way this one does, for those of you who are concerned the solder TIM distribution after de-soldering reflected poor mounting on AMD’s part.
Watch der8auer’s video for more info on how he managed to delid the chip. The use of solder means this isn’t a CPU you’ll want or need to pop the heatspreader off of, and der8auer’s efforts killed his own Threadripper, despite being careful with the core.