Want to know what the best GPU is right now? Check out our guide to the top graphics cards available today.
He was asked directly if the consumer RX version of the Vega GPU would be as fast as the Frontier Edition and responded with, "Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!"
Koduri also answered an earlier question about a 16GB variant of the RX Vega and gave a tantalising "we will definitely look at that..."
Considering they haven't officially announced any final specs for the gaming-focused versions of the new cards that would seem to suggest there will indeed be a 16GB RX Vega card, and if it's going to be quicker than the Frontier Edition we might be looking at a 1,600MHz GPU too.
He also confirmed Vega would be their first Infinity Fabric GPU, which is likely how the two components of the upcoming Raven Ridge APU are likely to talk to each other. With Vega utilising the same Infinity Fabric interconnect which allows the two Zen modules in the Ryzen processors to communicate at high-speed with minimal latency it's not beyond the realms of possibility that we'll see multiple Vega GPUs connected via the Infinity Fabric on a single board.
"Infinity Fabric allows us to join different engines together on a die much easier than before," Koduri explained. "As well it enables some really low latency and high-bandwidth interconnects. This is important to tie together our different IPs (and partner IPs) together efficiently and quickly. It forms the basis of all of our future ASIC designs. We haven't mentioned any multi GPU designs on a single ASIC like Epyc, but the capability is possible with Infinity Fabric."
Original story, March 26, 2017: The AMD Vega GPU architecture is the next generation graphics silicon the Radeon-red team are working on for release early this year, with the flagship Radeon RX Vega promising to deliver AMD a graphics card that can finally compete with the very top-end of rival Nvidia’s GPU stack.
AMD have announced the naming scheme for what will probably be the flagship Vega card - the imaginatively-titled Radeon RX Vega. It's likely to be the direct replacement for the R9 Fury cards from the last generation of high-end AMD graphics cards.
In 2016 AMD promised their Polaris graphics cards would bring their Radeon graphics cards back into the game, but while the RX 480 and RX 470 have shown impressive DirectX 12 performance against the mid-range GeForce-shaped competition AMD have yet to release a high-end card to give them genuine 4K gaming.
This is where the AMD Vega GPU architecture comes in, aiming to jump in at the high-end and providing the Radeon faithful with a serious GTX 1080 Ti contender.
Click on the quick links below to jump to the different sections.
AMD Vega release date
Just one month. That's the most you're going to have to wait because AMD have confirmed that Vega cards are still on track for a release this quarter, with the Radeon Vega Frontier Edition confirmed to be released at the end of June.
AMD previously confirmed they would be releasing Vega graphics cards in the first half of 2017 and, seeing as they’ve already shown a working Vega 10 GPU at the last year's New Horizon event for their Ryzen CPUs, it doesn’t look like it'll be long before the full stack gets released.
AMD also announced the Vega GPU in their professional-class Radeon Instinct deep learning accelerators at the start of December 2016. The top-end Instinct card is the MI25, which is a Vega Frontier Edition by any other name, and likely represents the fastest Vega can go right now.
And what else is happening in the coming months? Firstly the biggest tech show outside of CES, Computex in Taipei, and secondly the gaming extravaganza that is E3. Both of those could be likely launch locations for the sort of high-end GPUs that the RX Vega line is supposed to bring us.
Personally I'm still leaning towards a Computex launch at the very beginning of June. It's a big show for PC technology and could be the perfect place for AMD to announce a genuine return to the fight for the absolute fastest GPU around.
AMD hosted a pre-Computex event in Macau last year to show off the Polaris graphics architecture and the scuttlebutt is that they might be doing the same this Summer to fully unveil the first Vega cards just ahead of their release during the Computex trade show in Taipei.
AMD lifted the lid on their new Vega GPU architecture in a series of short videos on their YouTube channel. If you don't want to listen to a load of AMD marketing folk we've distilled their tech-essence down for you.
AMD's Scott Wasson is calling Vega "the biggest improvement in our graphics IP in the past five years." And given the few details they've given away today that doesn't look like too much marketing hyperbole.
The redesigned geometry engine in the Vega GPUs is promising much higher performance. "It now has the capability to process more than twice as many polygons per clock cycle than we could in our previous generation," explains Wasson.
But it's the High Bandwidth Cache and High Bandwidth Controller silicon which looks the most exciting and that's all related to moving outside of the limits of the graphics card's video memory. In normal GPUs developers have to fit all the data they need to render into the frame buffer, meaning all the polygons, shaders and textures have to squeeze into your card's VRAM.
That can be restrictive and devs have to find clever workarounds for large, open-world games. The revolution with AMD's Vega design is to break free of those limits. The High Bandwidth Cache and High Bandwidth Controller mean the GPU can now stream in rendering data from your PC's system memory or even an SSD, meaning it doesn't have to come via the card's frame buffer.
"You are no longer limited by the amount of graphics memory you have on the chip," explains Wasson. "It's only limited by the amount of memory or storage you attach to your sytem."
The Vega architecture is capable of scaling right up to a maximum of 512TB as the virtual address space available to the graphics silicon. AMD are calling Vega 'the world's most scalable GPU memory architecture' and they look on the money with that claim at first glance. But it will still depend on just how many developers will jump onboard with the new programming techniques when Vega launches.
This could potentially mean the recently-announced 4GB versions of AMD's upcoming RX Vega graphics cards won't suffer from having such a relatively small amount of video memory.
It might well get complicated recommending minimum GPU memory capacities in the system specs of new games once the RX Vega cards launch...
The new chip design also allows for more concurrency in processing non-uniform graphics workloads. Previous designs potentially left large chunks of silicon idle when the GPU was processing smaller operations, bottlenecking the graphics system. The new NCU design though is meant to allow parts of the GPU to work on smaller operations when there is spare capacity, meaning that there shouldn’t be as many wasted, idle parts of the chip.
This should mean more work gets done in the same amount of time as previous GPU designs. How much impact this will have on gaming workloads is difficult to say, but it could end up being important for the lower level APIs like DirectX 12 and Vulkan.
The AMD RX Vega cards will appear as standalone named cards sitting outside the rest of the Polaris-based RX 500 series. Those essentially rebadged cards are slightly tweaked Polaris GPUs with a very modest boost to their clockspeeds. The RX Vega cards though are completely different beasts from GPU to memory architecture.
The first Vega GPU, the Vega Frontier Edition, will use Vega 10 silicon and is likely to see AMD coming out with a high-end halo graphics card first, with potentially an RX Vega-branded version coming later in the same way the GTX 1080 Ti followed Nvidia's Titan X.
The Radeon Vega Frontier Edition is going to be AMD's answer to Nvidia's Titan cards, offering pro-level specs in a card that's not really designed for the consumer, but for the creators. That said, I bet there's going to be a fair few well-off AMD fans dropping the cash on a Frontier Edition in the same way the Titan GPUs have always found their way into gaming rigs.
That’s a 16GB card, sporting second-gen high bandwidth memory (HBM2), which will give it an insane level of video memory performance. We’re expecting a 2,048-bit memory bus with bandwidth around the 512GB/s mark.
AMD's Scott Herkelman also announced that AMD's board partners would be able to released different versions of the RX Vega card with different memory capacities. It looks like there will be both 8GB and 4GB versions of the card lower down the stack, both utilising second-gen high-bandwidth memory (HBM2).
The latest rumours show three RX Vega cards designed for the consumer sitting underneath the semi-professional Vega Frontier Edition. At the top is the RX Vega Nova, followed by the RX Vega Eclipse and finally the RX Vega Core.
AMD have said their GCN architecture can be configured to work with both HBM2 and GDDR5, so it's possible AMD will hold the top memory tech for just the top two cards in their RX Vega 10 GPU stack, leaving the cheaper GDDR5 to cover the potential RX Vega Core edition.
There have also been reports of a Vega 11 GPU too, though the rumour mill has been surprisingly quiet on that front recently. It’s possible AMD have decided the Polaris 12 GPUs will shore up the bottom end of the new 500-series range and we won’t see a Vega 11 chip until later on.
There have though been earlier rumours of a Vega 20 GPU that AMD are working on with GlobalFoundaries. The Vega 20 is reportedly going to be a 7nm GPU releasing in 2018 with up to 32GB of HBM2 and memory bandwidth of a ludicrous-speed 1TB/s. There are also rumours of it sporting 64 NCUs, support for PCIe 4.0 and coming with a teeny tiny 150W TDP.
With GlobalFoundaries looking to hit 7nm production in 2018 that part at least looks plausible. The rest? I’m not so sure; it looks like very wishful thinking to me.
Given AMD have shown roadmaps with Vega sticking to 14nm with a 14nm+ refresh to follow, it looks unlikely we'll see 7nm Vega. That's more likely to come with the Navi GPU architecture which is to follow it.
As well as outlining the different levels of AMD RX Vega GPU the recent leaks have defined pricing for them too. The top-end RX Vega Nova is rumoured to be released at $599, around $100 less than the competing Nvidia GTX 1080 Ti. Whether that means it's $100 slower we don't yet know...
The rumours also put a $499 price tag on the middle-order RX Vega Eclipse and the RX Vega Core at $399. That will leave the RX 500-series cards looking after the mainstream, sub-$250 level with Vega shoring up the high-end for the first time since the Fury X in 2015.
Quite how they're going to price the inaugural Radeon Vega Frontier Edition though is not known. Given that it looks like it's specced to directly compete against Nvidia's Titan cards I would expect it to arrive costing anything upwards of $999.
We’ve actually seen quite a lot from AMD’s new Vega GPU in benchmark form so far. The 8GB HBM2 version of the GPU has been shown in public at the recent New Horizon event, where it was running a Ryzen-powered gaming rig with the new Star Wars Battlefront Rogue One DLC. It was playing the game at 4K and was able to keep running consistently at over the 60fps mark.
At the recent AMD Tech Summit in China AMD showed a demo of Deus Ex: Mankind Divided running with the high bandwidth cache controller off and on side-by-side. The GPU-intensive demo showing a 50% improvement in average frame rate and 100% higher minimum frame rates.
AMD have also shown Doom running at 4K using the flagship Vega graphics card. Running at the game's Ultra graphics settings the frame rate is shown at around 70fps with a few dips below 40fps here and there. That's not far off GTX Titan X performance - no wonder Nvidia is waiting for Vega to release before launching the GTX 1080 Ti.
The demo of the unreleased Vega card was running the Vulkan version of Doom live at the Ryzen event and was shown outperforming the GTX 1080 by 10%. That demo also confirmed the 687F:C1 device ID for the Vega GPU. If that sounds familiar it’s because that designation was seen in the Ashes of the Singularity benchmarking database recently, as well as a C3 revision, offering performance around the GTX 1080 mark too.
That device ID has appeared again in a recent 3DMark Timespy benchmark result found online. This appears to be the 8GB version of the RX Vega, but instead of the GTX 1080 performance we've previously seen, the DX12 benchmark seems to show it performing at GTX 1070 speeds.
That’s maybe a little disappointing at first glance, but these are all benchmarks running on unreleased, unoptimised drivers. The clockspeeds for the leaked benchmarks have the RX Vega GPU running with a boost clock of just 1,200MHz which puts it a far cry from the 12.5TFLOPs the peak Vega GPU is capable of.
It’s been reported that the Doom benchmarks were run on a slightly modified driver for the old Fiji GPUs so there is potentially more headroom left to come from the first Vega GPUs when they do finally launch.
The big-boy Vega 10 chip, the one that’s meant to be based on the Radeon Instinct professional MI25 card, could potentially hit 12.5 teraflops of single precision processing. The GTX Titan X runs to a little under 11 teraflops for its part, so even if AMD releases the card with a slightly cut-down GPU compared with the one in the expensive MI25 it may still have a version able to compete with both the Titan X and the GTX 1080 Ti.
The most recent performance indicators though have come from the recent financial analyst day at AMD HQ. Here they showed Vega's 4K capabilities with Sniper Elite 4 running on a single GPU bouncing around the 60 frames per second mark . That's key because Raja Koduri made that a target for Vega from day one.
"One of the goals for Vega is that it needs to cross that 4K/60Hz barrier, a single GPU that can cross the 4K/60Hz barrier," he explained. "That seemed quite daunting two years ago."
Now, we've already seen Vega running Star Wars Battlefront at 4K/60Hz so the fact it can do the same with Sniper Elite 4 shouldn't really come as much of a surprise. What was more revealing from last night's demos, however, was the showcasing of what Vega's high-bandwidth cache controller might be able to offer the games of tomorrow. We have seen HBCC in action with Deus Ex: Mankind Divided at a previous event in China, but the Rise of the Tomb Raider test shows an even greater performance improvement.
The HBCC tech baked into Vega allows the card to stream in larger pools of data from the PC's system memory or even from an attached storage device. Right now that's incredibly useful for professional GPUs using massive data sets, but in-game, not so much.
"Today's games are written for today's memory constraints," explained Koduri, "so we don't see many games cross the requirements for 4GB... We are building this architecture not just for one generation."
In order to show what HBCC can offer games AMD showed a version of Vega limited to just 2GB of video memory where one card had HBCC enabled and one with the memory caching tech turned off. This was to simulate where the new technology can help out when the frame buffer is being maxed out.
The Rise of the Tomb Raider demo showed a massive difference in both the average and minimum frame rates outputted by the same spec of GPU. It's the minimums that are the most interesting part of this though, with the 2GB card without HBCC bottoming out at 13.7fps while the equivalent GPU running the new HBCC tech scores 46.5fps. That's between 2x and 3x the minimum frame rate which will have a huge impact on just how smooth a game feels to play.
"This is going to be a big deal," explained Koduri, "when we put it in the hands of both game developers and gamers."
And he could be right... if those game developers are given enough incentive to code specifically for this AMD-only technology. As Koduri says, current games are not designed to take account for this advance, so AMD are going to have to give them a solid reason to do so.