As it turns out, the video card wars are going to charge into 2019 quite a bit hotter than any of us were expecting. Moments ago, as part of AMD’s CES 2019 keynote, CEO Dr. Lisa Su announced that AMD will be releasing a new high-end, high-performance Radeon graphics card. Dubbed the Radeon VII (Seven), AMD has their eyes set on countering NVIDIA’s previously untouchable GeForce RTX 2080. And, if the card lives up to AMD’s expectations, then come February 7th it may just as well do that.

Today’s announcement is interesting in that it’s just as much about technology as it is the 3D chess that is the market positioning fights between AMD and NVIDIA. Technically AMD isn’t announcing any new GPUs here – regular readers will correctly guess that we’re talking about Vega 20 – but the situation in the high-end market has played out such that there’s now a window for AMD to bring their cutting-edge Vega 20 GPU to the consumer market, and this is a window AMD is looking to take full advantage of.

At a high level then, the Radeon VII employs a slightly cut down version of AMD’s Vega 20 GPU. With 60 of 64 CUs enabled, it actually has a few less CUs than AMD’s previous flagship, the Radeon RX Vega 64, but it makes up for the loss with much higher clockspeeds and a much more powerful memory and pixel throughput backend. As a result, AMD says that the Radeon VII should beat their former flagship by anywhere between 20% and 42% depending on the game (with an overall average of 29%), which on paper would be just enough to put the card in spitting distance of NVIDIA’s RTX 2080, and making it a viable and competitive 4K gaming card.

AMD Radeon Series Specification Comparison
  AMD Radeon VII AMD Radeon RX Vega 64 AMD Radeon RX 590 AMD Radeon R9 Fury X
Stream Processors 3840
(60 CUs)
(64 CUs)
(36 CUs)
(64 CUs)
ROPs 128
64 32 64
Base Clock ? 1247MHz 1469MHz N/A
Boost Clock 1800MHz 1546MHz 1545MHz 1050MHz
Memory Clock 2.0Gbps HBM2 1.89Gbps HBM2 8Gbps GDDR5 1Gbps HBM
Memory Bus Width 4096-bit 2048-bit 256-bit 4096-bit
Single Precision Perf. 13.8 TFLOPS 12.7 TFLOPS 7.1 TFLOPS 8.6 TFLOPS
Board Power 300W? 295W 225W 275W
Manufacturing Process TSMC 7nm GloFo 14nm GloFo/Samsung 12nm TSMC 28nm
GPU Vega 20 Vega 10 Polaris 30 Fiji
Architecture Vega
(GCN 5)
(GCN 5)
Transistor Count 13.2B 12.5B 5.7B 8.9B
Launch Date 02/07/2019 08/14/2017 11/15/2018 06/24/2015
Launch Price $699 $499 $279 $649

Diving into the numbers a bit more, if you took AMD’s second-tier Radeon Instinct MI50 and made a consumer version of the card, the Radeon VII is almost exactly what it would look like. It has the same 60 CU configuration paired with 16GB of HBM2 memory. However the Radeon VII’s boost clock is a bit higher – 1800MHz versus 1746MHz – so AMD is getting the most out of those 60 CUs. Still, it’s important to keep in mind that from a pure FP32 throughput standpoint, the Vega 20 GPU was meant to be more of a sidegrade to Vega 10 than a performance upgrade; on paper the new card only has a 9% compute throughput advantage. So it’s not on compute throughput where Radeon VII’s real winning charm lies.

Instead, the biggest difference between the two cards is on the memory backend. Radeon Vega 64 (Vega 10) 2 HBM2 memory channels running at 1.89Gbps each, for a total of 484GB/sec of memory bandwidth. Radeon VII (Vega 20) doubles this and then some to 4 HBM2 memory channels, which also means memory capacity has doubled to 16GB. And then there’s the clockspeed boost on top of this to 2.0Gbps for the HBM2 memory. As a result Radeon VII has a lot memory bandwidth to feed itself, from the ROPs to the stream processors. Given these changes and AMD’s performance estimates, I think this lends a lot of evidence to the idea that Vega 10 was underfed – it needed more memory bandwidth keep its various processing blocks working at full potential – but that’s something we’ll save for the eventual review.

Past that, as this is still a Vega architecture product, it’s the Vega we all know and love. There are no new graphical features here, so even if AMD has opted to shy away from putting Vega in the name of the product, it’s going to be comparable to those parts as far as gaming is concerned. The Vega 20 GPU does bring new compute features – particularly much higher FP64 compute throughput and new low-precision modes well-suited for neural network inferencing – but these features aren’t something consumers are likely to use. Past that, AMD will be employing some mild product segmentation here to avoid having the Radeon VII cannibalize the MI50 – the Radeon VII does not get PCIe 4.0 support, nor does it get Infinity Link support –

The other wildcard for the moment is TDP. The MI50 is rated for 300W, and while AMD’s event did not announce a TDP for the card, I fully expect AMD is running the Radeon VII just as hard here, if not a bit harder. Make no mistake: AMD is still having to go well outside the sweet spot on their voltage/frequency curve to hit these high clockspeeds, so AMD isn’t even trying to win the efficiency race. Radeon VII will be more efficient than Radeon Vega 64 – AMD is saying 25% more perf at the same power – but even if AMD hits RTX 2080’s performance numbers, there’s nothing here to indicate that they’ll be able to meet its efficiency. This is another classic AMD play: go all-in on trying to win on the price/performance front.

Accordingly, the Radeon VII is not a small card. The photos released show that it’s a sizable open-air triple fan cooled design, with a shroud that sticks up past the top of the I/O bracket. Coupled with the dual 8-pin PCIe power plugs on the rear of the card, and it’s clear AMD intends to remove a lot of heat. Both AMD and NVIDIA have now gone with open-air designs for their high-end cards on this most recent generation, so it’s an interesting development, and one that favors AMD given their typically higher TDPs.

Vendor performance claims should always be taken with a grain of salt, but for the moment this is what we have. If AMD manages to reach RTX 2080 performance, then I expect this to be another case of where the two cards are tied on average but are anything but equal; there will be games where AMD falls behind, games where they do tie the RTX 2080, and then even some games they pull ahead in. These scenarios are always the most interesting for reviewers, but they’re also a bit trickier for consumers since it means there’s no clear-cut winner.

All told then, the competitive landscape is going to be an interesting one. AMD’s own proposition is actually fairly modest; with a $699 price tag they’re launching at the same price as the RTX 2080, over four months after the RTX 2080. They are presumably not going to be able to match NVIDIA’s energy efficiency, and they won’t have feature parity since AMD doesn’t (yet) have its own DirectX Raytracing (DXR) implementation.

But what AMD does have, besides an at least competitive price and presumably competitive performance in today’s games, is a VRAM advantage. Whereas NVIDIA didn’t increase their VRAM amounts between generations, AMD is for this half-generation card, giving them 16GB of VRAM to RTX 2080’s 8GB. Now whether this actually translates into a performance advantage now or in the near future is another matter; AMD has tried this gambit before with the Radeon 390 series, where it didn’t really pay off. On the other hand, the fact that NVIDIA’s VRAM capacities have been stagnant for a generation means that AMD is delivering a capacity increase “on schedule” as opposed to ahead of schedule. So while far from guaranteed, it could work in AMD’s favor. Especially as, given the performance of the card, AMD intends for the Radeon VII to be all-in on 4K gaming, which will push memory consumption higher.

Finally on the gaming front, not content to compete on just performance and pricing, AMD will also be competing on gaming bundles. The Radeon VII will be launching with a 3 game bundle, featuring Resident Evil 2, Devil May Cry 5, and The Division 2. NVIDIA of course launched their own Anthem + Battlefield V bundle at the start of this week, so both sides are now employing their complete bags of tricks to attract buyers and to prop up the prices of their cards.

Speaking of pricing, perhaps the thing that surprises me the most is that we’re even at this point – with AMD releasing a Vega 20 consumer card. When they first announced Vega 20 back in 2018, they made it very clear it was going to be for the Radeon Instinct series only. That the new features of the Vega 20 GPU were better suited for that market, and more importantly as a relatively large chip (331mm2) for this early in the life of TSMC’s 7nm manufacturing node, yields were going to be poor.

So that AMD is able to sell what are admittedly defective/recovered Vega 20s in a $699 card, produce enough of them to meet market demand, and still turn a profit on all of this is a surprising outcome. I simply would not have expected AMD to get a 7nm chip out at consumer prices this soon. All I can say is that either AMD has pulled off a very interesting incident of consumer misdirection, or the competitive landscape has evolved slowly enough that Vega 20 is viable where it otherwise wouldn’t have been. Or perhaps it’s a case of both.

Shifting gears for a second, while I’ve focused on gaming thus far, it should be noted that AMD is going after the content creation market with the Radeon VII as well. This is still a Radeon card and not a Radeon Pro card, but as we’ve seen before, AMD has been able to make a successful market out of offering such cards with only a basic level of software developer support. In this case AMD is expecting performance gains similar to the gaming side, with performance improving the more a workload is pixel or memory bandwidth bound.

Wrapping things up, the Radeon VII will be hitting the streets on February 7th for $699. At this point AMD has not announced anything about board partners doing custom designs, so it looks like this is going to be a pure reference card launch. As always, stay tuned and we should know a bit more information as we get closer to the video card’s launch date.

Comments Locked


View All Comments

  • KateH - Wednesday, January 9, 2019 - link

    very curious abt these "new low-precision modes". Does that means FP16? INT8? INT4? How will the throughput compare to the competition in those modes? Also, despite the modest/lack of increase in FP32 perf, DP/FP64 perf is up massively- from 1/16 of SP perf in Vega 10, to 1/2 in Vega 20- which means unless AMD gimps Double Precision compared to the MI50 (something they're not known to do), this card will be good for a whopping ~6.5TFLOPs at FP64 which if you need it is hecking massive.
  • KateH - Wednesday, January 9, 2019 - link

    but really though, I am primarily curious just what the Low-Precision modes implemented on shader cores will look like and how they will compare to the competition's hardware implementation of such (tensor cores)
  • sing_electric - Wednesday, January 9, 2019 - link

    Yeah, it should be interesting to see - it looks like AMD is taking a pretty different ML approach (though part of that may just be due to the legacy of the fact that it turns out that GCN was really pretty good at it without anyone having planned it that way).

    Are there common tasks where memory bandwidth is an issue? If so, I can see the Radeon VII being great, since it's got 60% more than an RTX 2080 (not to mention twice the memory on board).
  • KateH - Wednesday, January 9, 2019 - link

    i couldn't name specific applications that are memory bw bound (cinematographer/content creator and gamer here; as an aside those Davinci Resolve numbers are quite enticing) but they're out there in the HPC sector. Also worth noting is that the big 16GB VRAM pool could do wonders for latency-bound tasks, where the ability to fit a larger dataset into VRAM without having to go back and forth into system memory can produce nonlinear performance increases.
  • Znaak - Wednesday, January 9, 2019 - link

    You are not wrong there, this is what I've been praying for! 16GB means my entire data will fit into GPU memory, which will allow me to come significantly closer to the theoretical maximum FLOPS and use double precision where needed. (mostly just in accumulating results, but still that happens quite often along the way.)

    And where I couldn't convince my boss to shell out for a $ 3000 Titan V or $ 9000 Radeon Instinct or Tesla card (small company and product is not shipping yet) I'm certain I can get him to pay $700 for this card!
  • sing_electric - Wednesday, January 9, 2019 - link

    That's exactly what I was wondering. Being able to move 16GB back and forth at 1TB/s without leaving the card has to offer performance advantages to some people. I've done some GPU compute work but really nothing that major, and its been small datasets so I've just been happy that it's way faster than CPU could be.
  • KOneJ - Friday, January 11, 2019 - link

    Vega Frontier Edition often snuck down into a price range your boss might've gone for. e.g.
  • tipoo - Thursday, January 10, 2019 - link

    Yeah that's exactly it. The 16GB of 60% faster memory will be of more interest to ML researchers, massive CAD modelling, video rendering, etc, more so than is needed for games as of yet. Seems like another Titan or Frontier play, more for professionals wanting a cheaper Pro card than for gamers.

    Wonder if the iMac Pro will be updated with it soon. Never know with them now.
  • mode_13h - Wednesday, January 9, 2019 - link

    First, read this:

    According to this, the new low-precision integer instructions look analogous to Vega 10's fp16 support (i.e. standard packed-arithmetic operations):

    They don't have anything equivalent to Nvidia's Tensor cores, or else their touted performance numbers would've reflected this, in previous Vega 7 nm launch announcements. It's not surprising, either, since the Tensor cores probably caught them off-guard and modern GPU design cycles are too long for us yet to see their response.
  • Santoval - Friday, January 11, 2019 - link

    The Radeon Instinct MI60 supports both INT8 and INT4 (apart from the standard FP16). According to the article I linked below it supports 118 TOPS of INT4 and 59 TOPS of INT8. The MI50 is a bit less performing. I have no idea if these features will be preserved in Radeon VII, but assuming they will I don't think there is a functional difference between them and Nvidia's tensor cores, so I don't see why they couldn't be used for stuff like DLSS and AI denoising. Volta can also do ray-tracing without ray-tracing cores, though it is a (beastly) exception).

    The only difference might be that they cannot be used concurrently with the FP32 shaders, since they are apparently not dedicated. I mean they cannot *all* be used along with the shaders (since they appear to stem from them) but a subset of them certainly could. However lacking support from AMD -at least for DLSS or some analogue of it- I don't see these modes being used at all in gaming.

Log in

Don't have an account? Sign up now