After a hell of a lot of waiting, a huge amount of speculation, leaked information and performance numbers - some accurate, some apocryphal - the red team's first high-end GPU in years is here: the AMD Radeon RX Vega 64.
You need a top quality screen to pair with your GPU, so check out our pick of the best gaming monitors around today.
They have also let slip a slightly cut-down version in the Radeon RX Vega 56 which performs a little under this higher-end card, aiming at the GTX 1070 performance wise, but isn't priced aggressively enough to really compete.
This high-end card is exactly why I love AMD. There's something intrinsically loveable about everything they try and do within the realms of PC hardware. It probably stems from my hardwired championing of the underdog that’s endemic in British culture, from a desire to see them, as the Dave of Bible fame, going up against the Goliaths of the tech world. They're always pushing fair, open practises, going above and beyond to find new techniques and technologies to push computing, and computer graphics forward. They're always looking to the horizon.
Unfortunately, such a laudable approach doesn't necessarily translate into hardware with the absolute best real world gaming performance. At least, very rarely right away at launch.
Will things be different with the AMD Radeon RX Vega 64 graphics card? This is the first consumer-facing Vega card, a $499 (£450) GPU designed to go head-to-head with the more expensive Nvidia GTX 1080, and deliver team Radeon their first high-end graphics core since 2015’s R9 Fury X.
So, can the vanguard of the consumer Vega cards deliver where so much AMD tech has previously failed, or are we going to have to wait for the full potential of the new AMD Vega architecture to show itself? I think you can probably guess...
Click on the quick links below to jump to different sections.
The new AMD Vega architecture represents what they’re calling the most sweeping architectural change their engineers have made to the GPU design in five years. That was when the first Graphics Core Next chips hit the market and this fifth generation of the GCN architecture marks the start of a new GPU era for the Radeon team.
Fundamental to the Vega architecture, represented here by the inaugural Vega 10 GPU, is the hunt for higher graphics card clockspeeds. The very building blocks of the Vega 10, the compute units, have been redesigned from the ground up, almost literally. These next-generation compute units (NCU) have had their floorplans completely reworked to optimise and shorten the physical wiring of the connections inside them.
They also include high-speed, mini memory SRAMs, stolen from the Zen CPUs and optimised for use on a GPU. But that’s not the only way the graphics engineers have benefitted from a resurgent CPU design team; they’ve also nabbed the high-performance Infinity Fabric interconnect, which enables the discrete quad-core modules, used in Ryzen and Ryzen Threadripper processors, to talk to each other.
Vega uses the Infinity Fabric to connect the GPU core itself to the rest of the graphics logic in the package. The video acceleration blocks, the PCIe controller and the advanced memory controller, amongst others, are all connected via this high-speed interface. It also has its own clock frequency too, which means it’s not affected by the dynamic scaling and high frequency of the GPU clock itself.
This introduction of Infinity Fabric support for all the different logic blocks makes for a very modular approach to the Vega architecture and that in turn means it will, in theory, be easy for AMD to make a host of different Vega configurations. It also means future GPU and APU designs (think the Ryzen/Vega-powered Raven Ridge) can incorporate pretty much any element of Vega they want to with minimal effort.
The NCUs still contain the same 64 individual GCN cores inside them as the original graphics core next design, with the Vega 10 GPU then capable of housing up to 4,096 of these li’l stream processors. But, with the higher core clockspeeds, and other architectural improvements of Vega, they’re able to offer far greater performance than any previous GCN-based chip.
The new NCUs are also capable of utilising a feature AMD is calling Rapid Packed Math, and which I’m calling Rapid Packed Maths, or RPM to avoid any trouble with our US cousins. RPM essentially allows you to do two mathematical instructions for the price of one, but does sacrifice the accuracy. Given many of today’s calculations, especially in the gaming space, don’t actually need 32-bit floating point precision (FP32), you can get away with using 16-bit data types. Game features, such as lighting and HDR, can use FP16 calculations and with RPM that means Vega can support both FP16 and FP32 calculations as and when they're necessary.
We’ll see the first game supporting RPM, and other Vega-supported features, like asynchronous compute, when Wolfenstein II: The New Colossus launches. The Far Cry 5 developers have also come out in support of RPM, and could make FC5 very Vega-friendly. 3D technical lead, Steve Mcauley, has gone on record stating: “there’s been many occasions recently where I’ve been optimising shaders thinking that I really wish I had rapid packed math available to me right now. [It] means the game will run at a faster, higher frame rate, and a more stable frame rate as well, which will be great for gamers.”
The Vega architecture also incorporates a new geometry engine, capable of supporting both standard DirectX-based rendering as well as the ability to use newer, more efficient rendering pipelines through primitive shader support. The revised pixel engine has been updated to cope with today’s high-resolution, high refresh rate displays, and AMD have doubled the on-die L2 cache available to the GPU. They have also freed the entire cache to be accessible by all the different logic blocks of the Vega 10 chip, and that’s because of the brand new memory setup of Vega.
AMD’s Vega architecture uses the second generation of high-bandwidth memory (HBM2) from Hynix. HBM2 has higher data rates, and larger capacities, compared with the first generation used in AMD’s R9 Fury X cards. It can now come in stacks of up to 8GB, with a pair of them sitting directly on the GPU die, making the memory both more efficient and with a smaller footprint compared to standard graphics chip designs. And that could make it a far more tantalising option for notebook GPUs.
Directly connected with the HBM2 is Vega’s new high-bandwidth cache and high-bandwidth cache controller (HBCC). Ostensibly this is likely to be of greater use, at least in the short term, on the professional side of the graphics industry, but the HBCC’s ability to use a portion of the PC’s system memory as video memory should bare gaming fruit in the future. The idea is that games will see the extended pool as one large chunk of video memory, so if tomorrow’s open-world games start to require more than the Vega 64’s 8GB you can chuck it some of your PC’s own memory to compensate for any shortfall.
"You are no longer limited by the amount of graphics memory you have on the chip," AMD’s Scott Wasson explains. "It's only limited by the amount of memory or storage you attach to your system."
The Vega architecture is capable of scaling right up to a maximum of 512TB as the virtual address space available to the graphics silicon. Nobody tell Chris Roberts or we won’t see Star Citizen this side of the 22nd Century.
This version of the RX Vega 64 isn’t quite the top-end iteration of the consumer Vega GPU range; AMD have also created a liquid-chilled version with an all-in-one water-cooling unit attached to its metallic shroud. That card is almost identical in specification to this standard RX Vega 64, except from the wetware version gets a higher base clockspeed and a much higher boost clock too. And, y’know, is a bit less blowy as well.
Other than that, the two full-spec GPUs sing from the same specs sheet. That means they’re both sporting the same Vega 10 silicon, with 12.5bn transistors packed into its pretty massive 486mm2 die, with the smallest working on the 14nm FinFET lithography. It’s worth noting here that part of the reason the chip is so damned big is because it’s got two 4GB stacks of HBM2 directly on-die, rather than arrayed around the outside of the GPU on the circuit board as is more traditional.
There is a limited edition version of the air-cooled RX Vega 64 with a similar metallic shroud to the liquid-chilled card, but that's only available to those taking advantage of the somewhat bizarre Radeon Pack bundles.
The Vega 64 version of the Vega 10 GPU contains 64 of the new next-gen compute units (not just a clever name, eh?) and therefore 4,096 GCN cores. The lower-spec $399 partner card, continuing AMD classic rule of two for GPU releases, is the Vega 56, which has 56 NCUs and 3,584 GCN cores.
Historically the second-tier version of a high-end Radeon card is the card that’s arguably more tempting. It has the same basic GPU, with only a little performance-related hardware stripped out or turned off. That should mean it performs mighty close to the Vega 64, especially if you factor in the overclocking potential of the Vega 56 GPU.
Well, this is where it all gets a bit awkward, isn’t it? We were told the AMD RX Vega 64 was being designed to go head-to-head with Nvidia’s GTX 1080, but, despite that now being the second-tier consumer GeForce card, it’s not a particularly favourable comparison. At least not across the full breadth of our current benchmarking suite.
We still test GPUs against legacy DirectX 11 titles because we still play those games, and the last-gen API still gets used in new games too. And in those legacy titles the Vega architecture is unable to really show its true worth, posting performance numbers which are generally around 20% slower than the reference Founders Edition GTX 1080. The exception is the Shadow of Mordor benchmark where it’s only 10% behind, and that closing of the gap is likely down to the speedy Vega video memory shunting around the high-res texture pack a little easier than the competition's GDDR5X memory.
I’ll admit there was a definite sinking feeling as I started my benchmark testing with the DirectX 11 titles first. Thankfully, things take on a more positive countenance when you feed it the sorts of workloads the advanced architecture has been created for. With the DirectX 12 and Vulkan-based games the RX Vega 64 actually starts to show a performance lead over Nvidia’s GTX 1080. With the DX12 version of Hitman the RX Vega 64 is 16% quicker than a stock GTX 1080, 13% quicker in Rise of the Tomb Raider, and 15% faster in Doom.
That’s a much more reassuring result than it was initially looking like we were going to get. What’s less reassuring, though, is when you start throwing overclocked versions of Nvidia’s finest GPUs into the picture. It’s probably a given that a factory-overclocked GTX 1080 will outperform a stock Vega 64, especially given the ~$100 price premium of those cards. But, when an overclocked GTX 1070 is capable beating AMD’s Vega 64 pride and joy, that’s a lot harder to swallow.
The Galax GTX 1070 EXOC SNPR is around the same price as the Vega 64’s suggested retail price and, in terms of DirectX 11 performance, it’s often faster than the new Radeon card. The modern APIs, however, do allow the Vega 10 GPU to show it’s genuine potential. You wouldn’t necessarily recommend someone picks up a suped-up GTX 1070 instead of the Vega 64, but equally GTX 1070 owners shouldn’t feel too despondent about their card’s relative performance just yet.
The benchmark results of the DX12 and Vulkan titles give me a lot of hope for the future of the Vega architecture and this RX Vega 64 card too. It looks like the new architecture is going to be a classic case of relying on AMD’s ‘fine wine’ attitude, where they’ve traditionally seen performance improve, sometimes quite dramatically, over a component’s lifetime. You may well get great future performance, even if Vega’s legacy frame rates remain off the pace.
What’s got me a little less enthused, however, is the fact that AMD have needed to create a GPU with 5.3 billion more transistors, and a total platform power consumption that’s some 43% greater, in order to compete with the GP104 GPU. And don't forget it’s been released more than a year later and still finds itself behind the older silicon in some instances.
That power draw is a little concerning - though might keep the miners at bay - and we had a minor issue with our sampling shutting down under load. The reference RX Vega 64 requires a pair of 8pin PCIe power connectors, and our 1,200W Corsair PSU has individual PCIe power cables which split out into two of these 8pin plugs. Match made in heaven, right? Nope. If you try and draw all the juice down one of these cables our Vega card freaks out and reboots the system. It's absolutely rock solid when using separate 8pin cables, but it's something worth considering if you're in the market for a Vega GPU. Your PSU has to be up to scratch.
The AMD Radeon RX Vega 64 is a tough card to recommend you buy... at least it is right now. And probably will continue to be so in this AMD-designed air-cooled guise too. It's a card that would’ve been great if it was released a year ago and may still be in twelve months time. That probably sounds pretty damning, but I actually feel rather positive about the new Vega architecture and its future gaming potential.
The classic AMD black blower cooler gets rather loud when it's running at full chat, though is more than capable of keeping the chip chilled even when you're overclocking the nuts off it. Which is lucky because it doesn't seem like there are going to be a lot of AMD board partners willing to take a punt on the Vega architecture just yet.
Pricing is a real issue for the new AMD Vega 64, however, with it seemingly being revealed, by Overclockers that the $499 launch price was just that, the card's price at launch. The first batch of Vega 64 GPUs was apparently subsidised by AMD to allow retailers to sell at the $499 (£450) SEP, but as that stock vanished in a matter of hours - minutes in some places - the cheapest Vega was suddenly at least $100 (£100) more expensive.
At that level the Vega is moving into overclocked GTX 1080 territory, and that's a place where the new AMD GPU struggles to compete, even in DirectX 12 and Vulkan tests. Right now the frame rate performance isn't quite there for the vast majority of games, at least not in terms of trying to compete against a reference-clocked GTX 1080. But that's based on games that are currently installed on our rigs, based on legacy technologies and optimised for traditional GPU designs. The titles that really take advantage of the modern DirectX 12 and Vulkan APIs, however, show much more promise in terms of the relative Vega gaming performance. Just not against an equivalently priced GTX 1080...
Still, that extra performance gives me hope the classic AMD ‘fine wine’ approach might just bare even more fruity things if/when more developers jump on the RPM bandwagon, or start to take advantage of the open world benefits HBCC might be capable of delivering. I'll admit, that's a relatively big ‘if’ considering the amount of pull Nvidia has in the gaming market as a whole, and the number of devs who may not be willing to knowingly ring-fence the best gaming performance just for AMD’s quarter of PC gaming’s installed user base.
But, with the latest console custom ASICs starting to use similar AMD technology to that which has now appeared in Vega, there's a possibility more cross-platform developers will start to take advantage of these new features. If that's the case then, even if AMD gamers only make up a quarter of PC gamers, the PS4 Pro and Xbox One X install base will add in a huge potential pool of users for devs to target.
To be fair, it's also worth remembering AMD is still the underdog. They may have a similar level of R&D investment as Nvidia, but that's based on their fighting on the twin CPU and GPU battlefronts. That means AMD have to pick their fights, and if it means trading legacy performance for long-term benefit, that could be the right choice... but, honestly, only time will tell.
I kinda want to give two scores for the Radeon RX Vega 64... so I will. On the one hand it's well off the pace in most legacy games, but on the other it manages to push ahead of the competition in the modern APIs it's been designed for. If I'm just scoring it for its DX11 performance then I'm looking at a lowly 6. With it being $50 cheaper than a GTX 1080, but 20% slower, well over a year behind its GeForce rival, it looks pretty poor value. That flips on its head when you're purely looking at its performance in modern APIs - $50 less for faster gaming performance. That's a definite win for the Radeon, I give it an 8.
Overall then, you're looking at a score somewhere in the middle. But AMD are in the high-end conversation again, and their silicon future looks pretty rosy too.