The upcoming AMD Navi graphics cards are going to have to work some to challenge the latest Nvidia cards, but they could be specced out to dominate the mid-range market rather than try and go toe-to-toe at the high end. Whatever the specs, we do at least know that they’re going to be hitting the shelves sometime in the first half of next year.
From all the industry folk we’ve spoken to, it’s not too far-fetched to think that the red team will manage to get its long-awaited graphics architecture onto the shelves in time for an early 2019 launch, but why should you wait for the next Radeon GPU?
The new AMD Navi GPU architecture will most likely be the direct successor to the current mid-range Polaris 500-series graphics cards, and not the more-recent, more high-end RX Vega architecture. That seems to have become even more likely with Nvidia releasing its top-spec RTX 2080 Ti cards this year without waiting to see what AMD can do in response to the brand new Nvidia Turing cards.
We don’t 100% know for sure that AMD will be replacing the Polaris architecture that has been powering gaming rigs for a few years now, as it might still decide to re-target against Nvidia’s top graphics cards. We know it wants to compete with the top GeForce GPUs, but whether it will manage that feat is still up in the air. But we’re betting on a mainstream graphics card range from Navi… at least at first.
AMD Navi release date
AMD’s roadmaps have indicated a launch sometime in 2019. However, the latest rumours hint this could be as early as the end of Q1 2019. We’ve also heard from more pessimistic industry sources stating June is a more realistic launch window.
AMD Navi specs
The 14nm Vega 10 and Polaris 10 GPUs, used in the RX Vega and RX 500-series cards respectively, hold a total of 4,096 Stream Processors for Vega and 2,304 inside the Polaris chip. Thanks to the 7nm process, AMD could fit roughly 1.6x more logic into the same die space with Navi… if TSMC’s numbers are to be believed.
AMD Navi pricing
Pricing all depends on whether AMD target the high-end or mid-range markets with Navi. This will likely also affect whether AMD utilise pricey HBM2 memory, or GDDR6. A midrange RX 680 could be somewhere in the realm of $330 to $400 at the most, while the top high-end card closer to $600.
AMD Navi performance
It’s still far too early to guess at performance figures. However, Navi will likely be pretty adept with the Vulkan API, and should be a little more efficient than Vega thanks to the 7nm process.
The RX 580 has been a great choice for mid-range gaming, and before the whole mining boom rendered them nigh-impossible to actually buy anywhere, it was our top pick for best graphics card. However, will Navi be able to offer something just as viable for gamers, and more importantly, will it be able to finally break Nvidia’s stranglehold on the high-performance GPU market?
It’s not yet confirmed whether team red will opt for the RX 600-series nomenclature, or name the cards RX Navi, like it did with the AMD Vega architecture. It all really depends on where the architecture ends up in the market.
Current rumours suggest that the next graphics card we see from the AMD built specifically for gamers will be the RX 680 – following on from the Polaris architecture – and laser-focused on mainstream performance ready for Sony’s PS5.
The first graphics cards featuring AMD’s Navi architecture are rumoured to launch by the end of Q1 2019. Computex saw AMD launch its first 7nm graphics card, 7nm Vega for Radeon Instinct, but rumours whizzing around the back alleys of the conference hall indicated that Navi wouldn’t be far behind.
Current roadmaps from team red indicate a 7nm Navi is “on track” for launch sometime in 2019. However, these public roadmaps aren’t all that specific and AMD have a huge amount of leeway as to when Navi’s release window is set. It does have a laser focus on nailing a consistent roadmap, so we can be pretty sure Navi will be launched next year. A couple of industry sources, however, have indicated to us AMD is aiming for a Q1 release, but have said a more realistic launch window would be sometime around June, 2019.
AMD already confirmed that Navi would be ready for both GDDR6 and HBM2 memory. The speedier memory tech has already started to go into mass production from the big three memory manufacturers – Samsung, SK Hynix, and Micron – and will be well into volume production before the end of the year.
“I remember the ATI / Nvidia days, this every year,” says David Wang, senior VP of engineering, alternating one hand ahead of the other. “I think that’s how you make this business so excitement, so interesting. That’s how you make gamers so excited about new hardware every year.”
Launching Navi cards early in Q1 2019 would put AMD face-to-face with Nvidia’s next-generation of graphics cards, such as the RTX 2080. Although, if Navi is a mid-range graphics card architecture as rumours suggest, it may not put up much of a fight versus the green team’s top-performing GPUs.
For most of the time Navi’s been in the public eye – albeit in name only – it’s been widely speculated that the next Radeon gaming graphics cards would consist of a multi-chip module design. That means, rather than a single monolithic GPU, Navi would instead feature multiple GPUs strung together, and working in unison, through a high-speed interconnect.
AMD already has an interconnect ready to go called Infinity Fabric. You might recognise it from AMD’s Ryzen and Epyc CPUs. This silicon ‘fabric’ connects up multiple CCXs, or core modules, within AMD’s current chips and allows them to communicate and function as a single unit. It’s also been used in the current Vega architecture design as an interconnect between the GPU die and the High Bandwidth Cache Controller communicating with the HBM2 memory.
Ex-AMD Radeon chief, Raja Koduri, fueled rumours of a potential MCM GPU design for Navi when he claimed Infinity Fabric was core to all of the red team’s future designs.
“Infinity Fabric allows us to join different engines together on a die much easier than before,” Koduri says. “As well it enables some really low latency and high-bandwidth interconnects. This is important to tie together our different IPs together efficiently and quickly. If 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.”
But, while AMD has the tech to squeeze even more low-performing and high-yield silicon into one graphics card with Infinity Fabric, it’s not as easy to actually get the card recognised by the system as a single graphics processing unit, David Wang, senior VP of engineering at AMD, explains.
“We are looking at the MCM type of approach,” says Wang, “but we’ve yet to conclude that this is something that can be used for traditional gaming graphics type of application.
“Anything’s possible… But you know if you think about this big GPU it’s actually multiple pipelines, multiple rendering interfaces, so in theory you can slice it in half. But the devil’s in the details because when you slice in half and you’ve got to make the interface invisible to a programmer, that means your interface has to be very wide and very, very high-speed so that it can look and feel like one chip. That’s the complexity. Because it’s hard to be done that way, that’s why people do CrossFire because that communication’s centred around a narrow, fast interface.
“Do you make that interface indefinitely wide and fast so it feels like a single die? But then it will become a physical implementation issue, whether that can be done.”
So maybe Navi won’t be the multi-chip monster that punters were hopeful for. If AMD was working on a MCM Navi chip for launch in a mere eight months, you’d expect the chief engineer behind its implementation to be a little more convinced on the MCM concept as a whole.
When AMD has been pushed into a corner in the past by the green team, it’s often released dual GPUs packaged together into a single graphics card to take on the top-end of Nvidia’s lineup. However, Crossfire and multi-GPU systems account for such a small quantity of gamers nowadays that developers just don’t make an effort to support these systems – so there goes that idea.
However, it has recently been discovered that the new 7nm Vega 20 cards will use AMD’s xGMI feature – a peer-to-peer interconnect built form the foundations of Infinity Fabric – which can offer GPU bandwidth higher than PCIe 3.0, rivalling Nvidia’s NVLink.
The 7nm Vega is for an Instinct, machine-learning card, so not a gaming GPU, but if we see anything linking Navi to xGMI in the future we’ll know that AMD is doubling down on multi-GPU. And it has said that multi-GPU compute is one of its aims for the future of Radeon Rays, it’s own real-time ray tracing solution…
At least in the memory department we’re a little more sure of what to expect. Wang confirmed to us at Computex 2018 that AMD would be approaching gamers graphics cards with a focus on the most cost-effective and best performing memory for that segment – and that has to be GDDR6 memory right now.
“In a workstation/datacentre segment they’ll be more than happy potentially to pay the premium” Wang says regarding HBM2 memory. “But that technology may or may not be suitable for the mass majority of casual gamers. So I think different technologies might be more suitable for difference price segments.”
That doesn’t mean HBM memory is entirely relegated to the scrap heap, but GDDR6 looks like the best fit for gamers. In either eventuality, Navi is prepped for use with both GDDR6 and HBM. If Navi is truly set to be the successor to the RX 500-series, GDDR6 will aid AMD in hitting the best balance between price/performance.
Navi is also rumoured to be the last graphics card built with the GCN architecture, in its current form, at its core.
Pricing speculation is a torrid affair this early in the game. AMD’s pricing entirely depends on the market segment Navi is geared toward once it launches. If the next-gen Navi chips really are headed for the mid-range segment as a Polaris GPU replacement, then the top AMD card can’t be any more than $400.
“I think the engineering always shoots for excellency,” Wang says. “…and that’s what I set out to do. For example, not just when it comes to performance, but performance per watt. Not just the highest performance, but performance-per-dollar. I think taking into consideration the power envelope but also the die size or the cost constraints.
“I think performance-per-Watt, in my mind, sets a topline ASP [average selling price] we can charge, and the performance-per-area, or per-dollar, sets a bottom line as our cost. So we need to continue to improve our performance-per-dollar or per-area and push out our performance-per-Watt. That would give our business guys much more room to improve margins – that’s what I’m set out to do.”
Wang’s comments definitely indicate a wish to make Navi competitive on price/performance – as was the focus with Polaris back when it first launched.
Even though we have zero benchmarks to go on at this point, and only a slight idea as to the market AMD are targeting with Navi, one of the few things we do know about this next-gen architecture is that it is going to be built on the 7nm FinFET process node. By GlobalFoundries own numbers, this extremely dense node delivers twice the logic density, which accounts for a huge performance boost of roughly 40% compared to the current 14nm FinFET node /or/ a drastic increase in power efficiency.
Just how much of this 7nm node performance AMD will be able to draw out of the GCN architecture is not yet known, but even if very little were to change in the design, we can still expect a considerable bump in performance.
As for power efficiency, this is something that David Wang has made clear is a top priority with the next-generation of graphics cards. The benefits of which has already been made apparent to team red with its first 7nm GPU, 7nm Vega Radeon Instinct for machine learning.
“With the Vega 7nm we packed more and more stuff,” Wang told us at Computex, “and the power looks fantastic. Definitely a lot lower than the current Vega. That’s no magic, just the geometry scaling.”
GDDR6 memory is also set to improve performance over the last generations GDDR5 and GDDR5X. The memory tech is expected to roughly double the bandwidth on offer, from 32GBs to 64GB/s, and work at much lower voltages.