News AMD Announces Radeon VII

[Hitman928]

More info to come. . .

 

[Qwertilot]

That wasn’t precisely great either I guess at least with hardware failure you get a full replacement?!

I would, if buying this for gaming, be genuinely mildly concerned whether it’ll have any longetivity in terms of driver support. Such a tiny install base and rather distinctive too.

Just hope they get it a bit more together for Navi when that comes round.

 

[VirtualLarry]

Honestly I wouldn't even pay $10 for anything that needed +3,000rpm fans.

My RX 570 MSI cards have fans that go to 3500RPM. I used to stick them at 2700 minimum, to keep things cool. Haven't needed to, since AMD changed Wattman around recently, and now they max at 75C at defaults, rather than 90C.

 

[Ajay]

So, reading through some of these comments, none of this is surprising. As most of you know, AMD's target for Vega 20 is the MI50/60. I’m guessing that yields were a bit better than expected, so AMD decided to ship the Radeon VII. AMD engineers probably had little time to update the Vega 56/64 drivers - hence their current poor optimization (that, or the engineers had to adapt the Pro drivers). The power usage isn’t surprising for a compute oriented GPU with 16 GB of HBM and 1/2 rate DP64 on an older architecture. The Radeon VII has good, but not great performance and isn’t priced competitively enough for it’s performance. It’s just a placeholder till Navi comes along. Clearly, as most of you already know, AMD has been highly focused on development of their Zen microprocessors - to the detriment of the graphics division (hence Koduri's departure). If successful, Navi will mark a resurgence of AMD graphics, but it won’t be an easy road given Nvidia's current market share and brand dominance.

 

[Shamrock]

These are only 1440p benchmarks, has about 20 or so. Also had problems with drivers, but overall, it's up and down against the 2080 RTX.

 

[pandemonium]

W1zzard really has the clearest picture on the cards efficacy in gaming with his performance/power graphs averaged over all games tested. It's digusting.

I hope drivers can bridge that gap some. As it stands, this card is severely disappointing.

 

[Adawy]

Digital Foundry's analysis.

The performance is exactly as I expected, Pretty competitive with the 1080 Ti and 2080 but the frame rate varies depending on the Game and "in-game location or event".

 

[tviceman]

I’m guessing that yields were a bit better than expected, so AMD decided to ship the Radeon VII.

Interesting; I thought the opposite. Since these are cut down dies and can't make it as MI50/60, AMD had to do something with them. If yields were strong, there may not have been enough left over to justify creating any kind of Vega gaming card.

 

[tviceman]

I dunno man, look at where Vega56 is today and tell me that it was as much a failure as people said it was when it launched. The main failure of Vega56 was finding one for sale at MSRP. Just a few tweaks and Vega56 is a respectable card for the money (though not a 2080Ti by any means).

Without taking your (or anyone else's) words or posts out of context, Vega 56 is exactly where it was at launch as it is today in relation to it's performance vs. it's main competition.

VS. the 1070 at launch (6.4% faster than a 1070)

VS. the 1070 today (7.5% faster than a 1070)

 

[Ajay]

Interesting; I thought the opposite. Since these are cut down dies and can't make it as MI50/60, AMD had to do something with them. If yields were strong, there may not have been enough left over to justify creating any kind of Vega gaming card.

The core config is the same as the MI50, last I checked. Well, plus a clock bump.

 

[IntelUser2000]

So 7nm offers 2x the density and 25% performance improvement over the 14nm process. Making it roughly equivalent to the transition between 28nm and 14nm. The 25% performance gain is again an ideal number that works in parts that are not running at top clocks. For a GPU, 1.6-1.8GHz is probably close to this looking at efforts from Nvidia needed to get Pascal to get such clocks.

This explains why most skipped 10nm, because the gains are not big enough, as it was with 20nm. Historically looking 10nm would be another half node.

Yet the effort to move to "half-nodes" such as 20nm and 10nm can't be ignored. The R&D cost and manpower needed may be just as great as full nodes when scaling was relatively easy.

In regards to competition, the underdog usually needs the competitor to basically screw up so it can catch up. I don't think the graphics side of AMD is clearcut as it was with their CPU side. Nvidia typically executed their plans better than Intel.

The pricing also supports my theory its the market leader that decides how products should be priced.

I'm saying don't get too hyped by Navi. Or even the successor to Turing. The gains are going to be less and harder to come by.

 

[Arkaign]

I think all of that is true, though massive performance could have been possible with the same die size, clock speeds, and process tech of Turing with a design that skipped Tensor/RT.

I'm not totally faulting Nvidia here, it's a bold move for sure. I could imagine many potential professional applications where it might make sense. For gaming so far it's fairly useless, so when you consider (rough back of envelope math) they could have offered a product basically triple the performance of a 1080ti for $1199, and at least double 1080ti for $699, wow.

Of course the only reason they probably went with such a gargantuan die was purely to implement the new bits. What's funny about looking at Turing is that the Gflop of the traditional portion of the GPU lines up perfectly with observed performance.

Also noteworthy is that while I absolutely am let down by the $/perf of RTX models, one cannot deny the scale of them and the ambition.

 

[Qwertilot]

Wouldn't the power draw have run somewhat out of control if they'd done that?

 

[Arkaign]

Wouldn't the power draw have run somewhat out of control if they'd done that?

I don't believe so, I mean basically instead of die space used for Tensor/RT stuff, just use it for additional traditional arch, basically mirror imaged.

There are some interesting deep dives on it out there. Tom's has a good one.

I mean it's all really interesting and pretty impressive in every way except actual gaming as of today. But who knows, maybe it will pay off. I'm a bit cynical of that due to a ton of factors, but I'm also not clairvoyant. Heh.

 

[coercitiv]

In regards to competition, the underdog usually needs the competitor to basically screw up so it can catch up. I don't think the graphics side of AMD is clearcut as it was with their CPU side. Nvidia typically executed their plans better than Intel.

RTG and Vega isn't AMD and Zen: their RTG products will only improve with the full weight of AMD behind them. Keep in mind Vega still lacks features, still needs modern & functional power management.

When even AMD themselves recommend undervolting as part of the overclocking procedure you know they are in dire need of more efficient voltage control:

GPU: Undervolting values of ~0.95-1.0V (with increase max power 20-50%) and a frequency increase ~80Mhz (3-5%)

This is the reason why you see undervolting discussed so heavily when it comes to AMD graphics products, the gains are huge because for some reason AMD themselves haven't been able to tighten voltage margins themselves. In fact the gains are so ridiculous that in the case of Vega 56/64 people went through the trouble of developing registry settings and testing 'safe' voltage limits in order to compensate for AMD software not allowing proper frequency/voltage control on custom Vega cards.

This cannot continue though, they need to address this issue pronto. If hardware based solutions are not available in the near future, the least they can do is make sure drivers work flawlessly when it comes to manual frequency/voltage settings, and never ever launch a card with drivers in this state, especially a high end card.

 

[DrMrLordX]

Without taking your (or anyone else's) words or posts out of context, Vega 56 is exactly where it was at launch as it is today in relation to it's performance vs. it's main competition.

That's stock. The whole point of getting Vega56 today is not to run it at stock. Try again.

I'm saying don't get too hyped by Navi.

If you want a successor to Polaris, it might be really good. If you are looking for something faster than Radeon VII, eh, good luck.

Interesting; I thought the opposite. Since these are cut down dies and can't make it as MI50/60, AMD had to do something with them. If yields were strong, there may not have been enough left over to justify creating any kind of Vega gaming card.

I think neither is the case. AMD took some cards that could probably make it as Radeon mi50s and rebadged them as Radeon VII to maintain a market presence. They disabled some pro features like PCIe 4.0 functionality, IF links, and 1/2 fp64 to prevent the card from threatening the mi50 too much. They're taking a loss on the cards as a way to keep their lineup fresh so they can finally pursue the consumer market in earnest with Navi later this year. Assuming Navi isn't terribad. Which it might be, we don't know yet.

 

[Qwertilot]

That seems incredibly likely to me too. They can afford the publicity because they'll sell the cards so why not?

I don't expect Navi to be terrible - there's a fair bit more leeway for mid range cards to cope with 'excess' power draw & more development funding with the console money etc etc. The real win for AMD would be if they can somehow get it efficient enough to challenge in notebooks, but that would take a huge leap.

 

[DrMrLordX]

That seems incredibly likely to me too. They can afford the publicity because they'll sell the cards so why not?

I don't expect Navi to be terrible - there's a fair bit more leeway for mid range cards to cope with 'excess' power draw & more development funding with the console money etc etc. The real win for AMD would be if they can somehow get it efficient enough to challenge in notebooks, but that would take a huge leap.

I don't want to go too far into Navi since this is a Radeon VII thread. I will say that I wish - for AMD's sake and for everyone's sake, really - for Navi to have as little to do with Radeon VII and Vega as possible. Navi needs to be the replacement for Vega iGPUs, Polaris, and the PS4/Xbone dGPUs. It needs to be able to go head-to-head with A12/A12x GPU and win (200GE iGPU + Vulkan loses to A12 iGPU + Metal, which is disconcerting to say the least). That is what Navi needs to be. Which is a lot of things.

If we are lucky it can also be a successor to Radeon VII. If it tries to do too many things at once, it might be a horrific failure.

 

[tviceman]

Vega 7 is making the 2070 look like a great buy.

The core config is the same as the MI50, last I checked. Well, plus a clock bump.

I didn't bother looking up MI50's specs. You're probably right. In that case, Vega 7 is a market presence product and may never have meaningful amounts of shipments.

 

[tviceman]

That's stock. The whole point of getting Vega56 today is not to run it at stock. Try again.

I'm not familiar with Vega 56 overclocking being amazing - especially after Maxwell. Can you provide some links by reputable sources? And please don't include sources that flash or mod cards; that's neither pervasive nor anymore practical than custom liquid cooling, which is a very nice practice as well.

 

[Ajay]

So 7nm offers 2x the density and 25% performance improvement over the 14nm process. Making it roughly equivalent to the transition between 28nm and 14nm. The 25% performance gain is again an ideal number that works in parts that are not running at top clocks. For a GPU, 1.6-1.8GHz is probably close to this looking at efforts from Nvidia needed to get Pascal to get such clocks.

Well, because of yield concerns, AMD limited Vega 20 to about 330 mm^2. A 500 mm^2 chip, with reduced clocks, would been significantly faster - and would have like been limited to Pro GPUs only (due to the high costs of TSMC 7nm wafers with less dice per wafer). So the gen to gen performance for normalized die size would have better than 25%.

Vega 20 was, for the most part, a port of Vega 10 to 7nm. Navi will be designed from the ground up for 7nm FF. I expect it to be better optimized and have some architectural improvements. I agree with you that Navi won’t have a shattering % performance improvement. The biggest improvement should be perf/$. The drivers will also be substantially better than Radeon VII. Big unknowns are Navi's die size and clocks.

 

[krumme]

Looking at amd roadmap it seems navi is still gcn based?
If so, the question is if we get a fix of the technology in Vega or they just dump the software technology and prioritizes next gen.
If they choose to enable the tech I think we get 25% efficiency improvement and 15% extra perf in many less optimised games (most) vs what could otherwise be expected.
The navi arch would probably end up in millions of apu so its probably worth it. But if it wasn't for the apu I am not sure fixing such an old arch this late makes sense vs allocating the resources for next gen. As a contrast Nv planned it right with Maxwell. Rtx ...less so

 

[Mopetar]

Looking at amd roadmap it seems navi is still gcn based?

That's what I've been assuming, but it is possible for things to change. I suspect that AMD learned enough from Vega to get Navi into workable shape. They'll have the advantage of being on 7nm to help keep them competitive with NVidia for at least another year, but eventually NVidia will have 7nm cards of their own.

Hopefully AMD started working on their new architecture years ago, because if they only recently started it's going to take a while before they have anything truly worth buying. They've shown that with Ryzen they are capable of designing excellent chips. Let's hope that we see the same kind of excellence from their graphics division.

Realistically though, they just need to offer some good products for the mass market. NVidia still doesn't have anything to address the ~$300 market segment that they've been able to dominate for the past few product cycles.

 

[DrMrLordX]

I'm not familiar with Vega 56 overclocking being amazing - especially after Maxwell. Can you provide some links by reputable sources? And please don't include sources that flash or mod cards; that's neither pervasive nor anymore practical than custom liquid cooling, which is a very nice practice as well.

Aww you don't like the V64 BIOS mods? That's what all the cool kids are doing. Anyway :

All they did was change some powerplay table settings. That's it. Massive powahhhhhhh

 

[amenx]

Aww you don't like the V64 BIOS mods? That's what all the cool kids are doing. Anyway :

All they did was change some powerplay table settings. That's it. Massive powahhhhhhh

Massive powahhhhh consumption too.

 

[DrMrLordX]

Massive powahhhhh consumption too.

Exactly! That overbuilt power delivery is good for something. Hell some of the modders have Vega56 pulling 400W.