Disassembly, Cooler & Interposer
Disassembly & Cooler
Some tools are needed to remove the card's cover. Six small screws that hold the shroud in place can be unfastened with a small Phillips-head screwdriver (PH1). This reveals AMD's cooler, along with a frame responsible for adding rigidity and dissipating thermal energy.
AMD leans on a direct heat exhaust solution once again, and that's not a bad thing. We can see the radial fan sitting in its chamber, which brings in air from the case. This flow passes horizontally though the sink’s body and blows out of the card’s slot cover.
The backplate is made from black anodized aluminum. Looking good is its only purpose, which is to say that the plate doesn't help with cooling. AMD secures this piece using six T6 screws.
Our attempt to make the backplate functional with thermal pads didn't get us very far; they didn't draw much waste heat away from the board.
Up top, there's a massive cooling frame that helps keep the card's structure nice and rigid. It also looks like AMD implemented some of the design lessons it learned from past generations. Similar to Gigabyte's Aorus GeForce GTX 1080 Ti Xtreme Edition, the chokes transfer heat to the frame through thermal pads. Indentations in the metal accommodate voltage regulation circuitry as well.
The heat sink’s body is made of thin aluminum cooling fins attached to a large copper vapor chamber. Toward the bottom of the following two pictures, you can see the chamber's outlet, which was soldered shut and should never be snapped off. A large protrusion on the copper plate's surface is situated perfectly to make contact with AMD's GPU/memory package.
The 7cm radial fan is a simple ball bearing-based model from Delta. This is a first for AMD, and it’s certainly nice to see after generations of loud fans on the company's reference cards. Whereas those old fans spun at up to 10,000 RPM, the new BVB1012-family model hits its maximum speed at 5000 RPM. AMD’s target is around 40 to 41% duty cycle, translating to ~2000 RPM.
With the board completely exposed, all eyes turn to AMD's GPU and the HBM2 mounted together on an interposer, which sits on a large package substrate.
The Package: Confusion & Production Problems
Of course, the GPU can't simply be soldered to the PCB. As with AMD's Fiji-based cards, Vega and its complement of memory go through a packaging process that's handled by specialty tools. The resulting module is much easier for board partners to handle later on.
Looking at the Vega Frontier Edition and some Radeon RX Vega cards, pictured on the left, it's clear that AMD now favors the use of molding. The area around the GPU and memory is filled with an epoxy-like material that significantly improves stability. These packages are made in Taiwan and manufactured by ASE. Compare that to the version on the right, which we snapped of our Radeon RX Vega 56 sample. We went into depth on the specific differences in Is AMD Vega's Package Construction A Problem?
That package on the right is actually the third version we've come across; its shim, circuit board, and inscription indicate a different manufacturer and origin.
Thinking back to Fiji, AMD usually shipped its chips first to memory vendor SK hynix, which then assembled its HBM modules, had the packages completed in Korea without molding, and finally sent back, completed, to AMD. Both of the Vega cards we tested previously had molded packages from from ASE in Taiwan, so we now have a new Korean package on the scene.
So, what’s the problem? The HBM2’s positioned 40μm lower on the packages without molding and the underfill’s a bit different too. This causes unexpected production challenges, delaying add-in board partner products. We’ve been told that some of AMD’s partners are having to use six screws to attach their heat sinks instead of just four. Naturally, already-complete thermal solutions and backplates either can't be used or have to be modified to work with molding-less packages.
Is The Memory Different?
It is rumored that HBM2 made by different manufacturers (Samsung and SK hynix) is being used depending on the package version. Although we haven't talked to anyone willing to confirm this, it'd at least explain the different HBM2 module heights and the three packages we've run across. Insufficient packaging process capacity shouldn’t be a problem at this point in time, and should be excluded as the sole reason for all of this.
MORE: Best Graphics Cards
MORE: Desktop GPU Performance Hierarchy Table
MORE: All Graphics Content