Review: ASUS Radeon RX 5700 XT ROG STRIX

Where were those arguments when AMD had incomparably higher compute power?

Both companies make choices. And they matter. Imagine company works into GPU additional HW capable to do very high quality compression of 12 streams at 4K.
Great for game streaming servers, right? Imagine it costs 20% traditional gaming performance due to die space occupation.
Or imagine that company adds FP16 when nobody uses it. Or INT4/8. Tensors, RT, whatever.

Reason why company puts something into GPU does not matter. In the end of the day extra HW parts mean higher production cost that is put on customer's shoulders. If one has good use for it, good, if not, bad. Turing definitely did not do well for traditional rendering.
If Pascal was scaled to Turing's 2080 Ti size, AMD could not touch it even with RDNA. Pascal was pinacle of efficient design. Turing is fast way down. And unless nVidia changes their RT extra HW strategy, next generation will not have any better reception than Turing.

 

So, a chip that is twice as large, with 30% more transistors, on an older process, is faster and has many more features in hardware, is using the same power as a midrange chip on a smaller process.

What do you think happens when the people who make that chip, decide to go on the same process?

 

I don't get this. 2070 Super is based on 13.6B silicon because its just what Nvidia cut it down to. It's not like they can't physically make a 10.8B 2070 out the gate. In which case by your own admission it's beating the 5700XT for essentially the same transistor amount.

Both companies have tons of patents that aren't implemented.

Edit: I see - the 2070 is a TU106 super is TU104 - but the point still is the 2070 Super is down an GPC - so cut that out and what's the transistor cost? Nvidia could still build that - probably more than 10.8B but still only 10% or so higher than a 5700XT.

I posted in the other thread, obviously after this post, but AMD's patents on RT are definitely in response to Nvidia. Nvidia's been openly working on GPU ray intersect since 2016 (Optix supported GPU intersects in 5.0 released in 2016) and the patents for it stretch back as far as 2007. The BVH intersect acceleration (RT Core) was patented in 2015. It's basically a decade worth of raytracing on GPU R&D being wrapped into one package with RTX. And it's not like Nvidia isn't improving their tech either - people keep assuming that Nvidia can only scale it's RT cores up to improve performance but there is definitely tons of IPC gains they can make via architecture improvement.

RDNA as a whole is an extremely good start though. I think over the next few generations it's going to take market away from Nvidia. Especially because Samsung is not only building their own GPUs with it but going to contribute back to it's development. Samsung is going to work hard to get the power consumption down and that's going to benefit AMD in a bunch of different markets.

It's going to be interesting to see where this goes 3-4 years from now.

 

I basically agree with you. Chip sizes only muddle the picture, and with so many variables, functions, features that do things others may not, its basically a pointless exercise.

 

Well, it seems that now you can find 2070 Super at the same price (it's KFA2 implementation) so Asus version of 5700XT will be a tough sale given the fact that you can get a superior performance card at the same price and with some added technologies + 2 new games...
500 pounds it's too much for this card even with asus strix premium quality build...on the other hand seeing the horrible "drug lord likeable" implementation of MSI if you want the best custom cooled RX 5700 XT then the card reviewed here it's the way to go...

Also the well known AMD game: "wait for our next generation because it will rock!" continues: https://www.digitaltrends.com/computing/amd-graphics-card-nvidia-killer/

 

Poor Ampere.

 

I honestly wish that this Navi is just the first little test, and that they really have an "Nvidia killer" in their hands.

 

And how much would it cost? 2 grands CAD? 3 grands CAD? I think there's way too many guys who refuse to admit there's like no mid range RTX card for people outside of USA. There's the 2060 Super and then there's too expensive cards. The basic two fans models of the 2060 Super are 525-550$ CAD. That's the price of a 5700XT ... BTW Spencer just said the next XBox will focus on fps and smoothness of gameplay.

 

No, it's larger (ie. using more transistors) because it's a different architecture on a different node.

There's no available evidence to suggest that Nvidia's RT/Tensor cores use any significant die real estate and even if there were that would merely be part of how Nvidia chose to allocate their transistor budget. As there's no TU102/TU104/TU106 dies without these features it's a pointless thought exercise with no bearing on reality.

Consider any other feature of any random GPU architecture, like Rapid Packed Math, added USB-C or audio functionality, being first to a new memory architecture or display output, and improved codec block etc. Why doesn't any of these features invalidate any comparison with previous architecture?

Because those features are simply evaluated as part of the product and any die space used by said functionality is still something the vendor has to pay for in the end.

Anyway, I'm not going to go any further down this rabbit hole unless you give me a compelling reason to, as it's not only off topic for the thread overall but outside of the scope of the argument I'm making myself.

Which, for reference, is this; Die shrinks are not a panacea, the foundries stated improvements on the transistor level aren't analogues for what we see in the end products.

As such the criticism leveraged against the perceived shortcomings of Navi in general, and the 7nm process in particular, is misplaced at best.

Once again you're ignoring absolutely everything I wrote and restating the same nonsense again.

I'll make one last effort to make my point here but if it doesn't get through after this I guess we're done here.

As for your point above the 2080(S) is using more power than the 5700XT but yeah, sure. I could put the same statement in a very different light with a slightly different highlight though, let's take a look;

"So, a midrange chip on a smaller process reaches 90% of the performance of the competitions high-end while being half the size and using less power."

We're talking about the same chips here, just wording the argument differently.

Die sizes and power usage are targets set by the manufacturers, ie. AMD and Nvidia. I'm sure Nvidia could make a 400W TU104 but there would be challenges both cooling the card as well as pretty solid user concerns about the power draw at that point.

Similarly AMD could definitely have designed the 5700/5700XT for even lower power draw if they wanted to, you can even achieve that after the fact by undervolting the cards yourself. Yet even a relatively minor performance deficit would have changed the positioning vs. the 2060(S)/2070(S) and inevitably resulted in even more fanboy thrashing.

Frankly, once the initial sticker shock receded, I found Navi to exceed my expectations with regard to power/performance/die size by a fair margin. As a new architecture it also has the opportunity to mature and improve well over the coming years, whereas Vega likely won't have much more squeezed out of it at this point.

My only concern is the pricing but that's not really related to either the process or the architecture.

With this in mind, and the rather extensive examples I provided in my previous post, do you understand why your faith in the power of node shrinks is misplaced? The architecture is quite solid, you're just focusing on the things that make in poor in your mind rather than it's inherent strengths.

 

@Exodite , I dont think this debate is worth investing as much time, thought or energy as you seem to done in your response. Just not that important to me.

 

Not where i live you can't. Less expensive 2070 Super on Newegg.ca is 682$ CAD (it's on sale the regular price is 709$ CAD). The non blower 5700XT are not released yet but the less expensive blower model is 529$ CAD and Sapphire non blower card is supposed to be 10$ US more so it will be in the neighborhood of 550$ CAD. That's a pretty big difference if you ask me.

 

Actually, the 1660Ti was performing just fine before it was removed from the system. The reason it was underperforming initially was due to an issue with the chipset drivers Once that was corrected, the card performed as expected.

The FX5700 (non-XT) is functioning just fine.

I'm just not that easily impressed.

When people have no real, valid argument....this is where they go.... Meaningless data.

If people stuck to threads concerning products they are actually interested in, everyone would be better off. Unfortunately, too many people are here simply to start arguments and try to force their opinions of various products and companies on other people.

 

The sad issue with all this is that there can be no real price war, because the end products are too disparate from each other. The only reason that Nvidia can get away with all of this is that AMD doesn't have really competitive hardware.

 

The custom solutions that everybody was expecting for 5700 XT are kind of a disappointment money wise.
Well, let's wait the next summer when AMD will have an nvidia killer

 

Well, we saw what happened to Intel - so who really knows?!

 

Yes - but so have all custom versions been for the past several years now. There's so little OC headroom that it doesn't matter.

 

That is very true, but unlike Intel, Nvidia doesn't seem to relax at all.

 

Or do they? They did release some answers to RX 5000 series, but only just. They've only been releasing "just slightly better" products for years and years now. Just like Intel.

 

C´mon, be serious...

Kepler to Maxwell was a big leap and from Maxwell to Pascal even bigger.

 

You can hate them or love them nvidia is indeed bringing some great cards. Year after year amd just cant catch a break.