New architecture, better memory design
If you didn’t already know, the Vega 64 comes in two reference cooler designs: one with a limited edition silver shroud, and the other being the signature matte black and anodized aluminum design that AMD has fitted on its cards in years (since the Radeon R9 Fury cards, if we’re not mistaken). With that said, you’ll also see that the same reference cooler shroud decorates the Vega 56.
On an aesthetic level, the ratio of anodized black aluminum is higher than the amount of plastic that was used to build the cooler shroud, and for obvious (engineering) reasons. Compared to plastic, metal conducts heat better. That benefit is also something of a drawback: you’re not going to want to be putting your hands on this card when it’s working in full swing, not unless you’re intentionally looking to give your fingers something of a burn. Like all reference coolers, the card employs a blower-style fan, which sucks in air from the top of the card and exhausts through the back of the card, where you’ll find all of the card’s output ports. On that note, the card offers three DisplayPort 1.4 outputs and a single HDMI 2.0b output. Much like NVIDIA’s high-end GeForce GTX 1080 Ti, the RX Vega 56 doesn’t come with a dual-link DVI port. Then again, we are fairly certain that the gaming monitor that one would use with a card of this caliber won’t have a DVI input.
Moving forward, let’s talk about the driving force and the beating heart of the Vega 64: its Vega 10 GPU, and the new High Bandwidth Memory 2 (HBM2) format. Compared to Fiji and the first iteration of HBM, the Vega 10 is built in and around a new 14nm lithography, versus the now archaic 28nm die architecture that was used in Fiji.
The GPU die is also smaller, packing 8GB of HBM2, 4,096 stream processors, and 64 compute units (hence its namesake), all spread across a 2,048-bit memory bus. Its memory clock speed is rated at 945MHz, but don’t let that lower-than-average memory clock speed fool you. Thanks to some improvements implemented on HBM2, the GPU’s base clock speed is set to run at 1,630MHz out of the box, which is actually a little higher than what AMD rated during the company’s pre-SIGGRAPH presentation.
Alas, all that graphical prowess comes with a pretty hefty price: power consumption. To power it properly, the card requires two 8-pin PCIe connectors to be, well, connected to it. Oh, and the GPUTach meter makes a return to the Vega 64, essentially providing users with a visual sense of the card’s workload in real time.
On the other end of the card, the RX Vega 56 also comes with a dual switch that allows you to switch to the secondary BIOS, which makes the card consumes less power (at the cost of performance, obviously). On its default BIOS configuration, the card will pull a maximum TDP of 210W, while running on the power-saving configuration will enable the card to pull between 135W (in power saving mode) and 173W (in turbo mode).
Putting it through its paces
So then, the million dollar question still remains: how does the Vega 64 perform? To answer the question, we tested the card using our test bed, which comprises the following components:
• AMD Ryzen 7 1800X • Gigabyte AORUS AX370-Gaming 5 • 16GB (2x 8GB) Geil EVO X DDR4-3200 RAM
(auto-timings: 16-16-16-36) • Noctua NH-U12S SE-UM4 Cooler • Kingston HyperX Predator
PCIe SSD (480GB) • WD Caviar Black (6TB) • Corsair RM1000 PSU Our benchmarking programs also included the following:
• Futuremark 3DMark 2013 • Unigine Superposition
• Deus Ex: Mankind Divided • DOOM • Hitman
That brings us to our next point: overclocking. Now, we should make it clear that while we did manage to overclock the card, the amount of overclocking headroom that the card gave us wasn’t what we’d call substantial. At best, we managed to bring up the GPU’s core clock by an additional five percent (1.72GHz, more or less), and the memory clock from 945MHz to 970MHz. By comparison, the Vega 56 that we tested shows nearly double the overclocking headroom, but not without some form of diminishing returns.
With the Vega 64, the card just wouldn’t budge any further than five percent. That brings us into another related segue, we also compared the Vega 64 with the Vega 56 and NVIDIA’s GeForce GTX 1080 FE.
During 3DMark’s Fire Strike and Time Spy tests, it was quite obvious that, despite the limited overclocking headroom, the Vega 64 was clearly pulling ahead of the other two cards. Likewise, the card was also a pretty solid performer in Unigine’s Superposition benchmark on all three resolutions, with the shaders set at 4K optimization.
Interestingly, the game portion of our Lab Exam confirms that NVIDIA’s year-old GeForce GTX 1080 actually putting the Vega 64 through its paces (albeit not by a lot), with the card going almost toe-to-toe with AMD’s top-dollar card in certain titles, and in different resolutions too. Nevertheless, it does show that the Vega 64 is capable of holding its own in the triple-A titles we used.
The only issue we found with the card was its power consumption. Out of the box, we noted that the card was drawing an approximate 450W off the wall, and this was before we overclocked it. After overclocking (and cranking up the fan’s speed to run at nearly 5,000 rpm), that number spiked to nearly 500W, making it one of the thirstiest reference-cooled cards that we’ve tested to date.