Share:
Notifications
Clear all

Is Navi Good Enough? New video thread  

  RSS

Matthew Connatser
(@matthew-connatser)
Member Moderator
Joined: 1 year ago
Posts: 21
June 14, 2019 11:43  

Jim's released a new video, check it out here.


Kian and muziqaz liked
Quote
Ype
 Ype
(@ype)
Eminent Member Registered
Joined: 1 year ago
Posts: 91
June 14, 2019 14:09  

There is no such thing as a bad product, just bad pricing. 
I am definitely not a marketeer but this release is just as unexciting as Vega on 7 nm to me.

Jim didn't mention it in the video so this might be seen as off topic, but Zen2 + Navi will be at the heart of both Microsoft's and Sony's upcoming consoles. From what I've seen the Navi parts of their custom chips might have dedicated ray tracing. 

With that in mind I would not buy this early Navi release. It's not priced aggressively and it could be outdated feature wise next year. You could even argue that's it's outdated right now when you look at Nvidia's lineup with ray tracing. 

Although you could make the outdated argument for any piece of tech, keep in mind that the next consoles will be around for a long time.

With kind regards,
Ype


Ryan liked
ReplyQuote
eraser1
(@eraser1)
New Member Registered
Joined: 1 year ago
Posts: 6
June 14, 2019 15:30  

I seem to remember Sony talking about 4k@60fps on the upcoming PS5. I think Xbox might have also said something similar (maybe just 120 fps at 1080p/1440p?).

Regardless, we know that the 1080ti and 2080 are just barely sufficient for 4k 60fps. The 2080ti does manage 4k60, however. So that means AMD is expected to put a Navi chip that has to - at least - be better than the 2080/1080ti; if not matching the 2080ti; in every next generation console.

Each console would cost around $500? Hell, getting a GPU for $500 that can do 4k60 would be an amazing feat in and of itself. But if you factor in all the other costs of the console? Seems like a fever dream.

 

I don't know. AMD may have really dropped the ball with Navi - Sony and Microsoft aren't gonna be happy about that if they don't get it together by next year. Frankly, I'm not sure they will.

Or perhaps they have a serious ace up their sleeves. But I really doubt that.


liquid liked
ReplyQuote
liquid
(@liquid)
New Member Registered
Joined: 1 year ago
Posts: 3
June 14, 2019 15:40  

Great Video as always. Jim's points are good. AMD has indeed not caught up Nvidia in performance when one considers that they are a node behind. At least when not guessing about the efficacy of magic tricks like possibly being capable of upscaling with higher visual fidelity and lower performance loss than DLSS.

I do however, think high end Navi 1st gen cards are possible to make without running into power problems (or even manufacturing price/perf problems). 

Big disclaimer: I am assuming that the 5700XT is not bottlenecked by bandwidth.

The main reason I believe higher end Navis are possible is due these factors coming together:

¤ Lower prices than near the 2018 peak for DRAM products, presumably including HBM.
¤ Navi's lower need for memory bandwidth compared to Vega to produce the same or higher performance.
¤ Improvements in HBM and GDDR6.
¤ GDDR6 vs HBM power requirements.
¤ GDDR6 16gGbps or 18Gbps GDDR6 @ 256 GB/s vs 14Gbps @ 384-bit

 

https://www.pcgamesn.com/nvidia/turing-gpu-16gbps-gddr6
https://www.techpowerup.com/240522/samsung-starts-mass-producing-2-4-gbps-8gb-hbm2-stacks
https://www.tomshardware.com/news/samsung-flashbolt-hbm2e-hbm2-memory,38874.html

60 CU Navi:

I'll first address the hypothetical 60 CU version of Navi. 5700 XT ("Navi 40") has 448 GB/s bandwidth.
There are two options for this that are more likely better choices than a 384bit memory bus with 14 Gbps GDDR6.

GDDR6 Option:
I believe that the logical choice would not be to have a 384bit bus, but instead, it would have 16 Gbps or 18Gbps GDDR6 at similar or only slightly higher memory system power draw than the earlier 14 Gbps offering.
18 Gbps = 576 GB/s
16 Gbps = 512 GB/s

576 Gbps should be sufficient for a Navi 60 chip.
512 Gbps should be sufficient for SKU's made from salvaged dies with some CUs disabled.

Such a card would likely still have lower power requirements than R7 and and Vega64. The power would still be quite high unless clocked moderately, however.

HBM Option:
Aquabolt HBM: 2 stacks yielding 307.2 GB/s per stack =  614.4 GB/s
That's ~37% higher than 448 GB/s, and is likely sufficient to carry the 50% increased CU count. Memory power would be low and one could have 16GB VRAM spread over 2 stacks. Clocks could be increased at the cost of some power and it would handily beat the Radeon 7 in gaming at lower power and much lower manufacturing costs. Likely to trade blows with the 2080ti.
Downside: Manufacturing cost higher than the GDDR6 solution.

Navi 80. The highest end:
Flashbolt HBM: 2 stacks yielding 410  GB/s per stack = 820 GB/s
820 GB/s is ~83% higher bandwidth than 448 GB/s. This will probably be enough for the doubling of the CU count. Thermal restraints would mean lower stock clocks than the proposed Navi 60 Aquabolt HBM, like the 1080ti having lower stock clocks than the 1080. It would be a monster overclocker on water.  Salvage SKU's @ 72 CU and 64 CU if one assumes double disable rate from Navi 40->36.  Aquabolt might be sufficient for a 64 CU part.

We don't know exactly how good yields are, so this chip might not be feasible for that reason alone. I am guessing it would be cheaper to manufacture than the R7 was. That the high memory prices at the time combined with having to have 4 stacks of HBM, as well as 7nm being a less mature node that time outweigh that this Navi 80 CU chip would be significantly bigger.
The R7 being dies salvaged from datacenter GPUs probably turns it around and makes the R7 die manufacturing more profitable even if manufacturing costs were higher per card. I still think it would be a great halo product.

AMD might want to wait with this one, or even Navi 60 until 7nm+ EUV, but I wanted to mention it as something that they could make now, if they wanted to.

In closing:
When reviewers tinker with the 5700XT GDDR6 clocks and try to reduce it to see where it starts being bandwidth bottleneck, we might get a clue as to if these 60 and 80 CU Navi are be feasible.

This post was modified 1 year ago 2 times by liquid

ReplyQuote
thehumangerm
(@thehumangerm)
Eminent Member Registered
Joined: 1 year ago
Posts: 71
June 15, 2019 08:06  

The Navi architecture also contains a lot of additional features like RDNA that once adopted by game manufactures will cause performance and or appearance to actually increase over time in game titles. RDNA is ground up different from GCN but is backward compatible with the developer toolkits. Meaning it is backwards compatible with GCN optimized games yet better performance will come on board once game developers optimize for RDNA. Something that will be fairly painless and should be widely adopted given backwards compatibility of the for the tools they are already using.

Radeon Image Sharpening is another example. it is open source so Nvidia can even adopt it and that will give substantial image quality improvements in games at no performance cost. meaning better looking games or the ability to use lower res textures in some instances for a equally good looking result equating to better over all performance on NAVI. NAVI is no chart topper but it just like ray tracing for Nvidia it is introducing new tech into the hands of developers which will see wider adoption over the next few years.

Information source: https://www.youtube.com/watch?v=EdaGdFcDLSk


ReplyQuote
Nortski
(@nortski)
New Member Registered
Joined: 1 year ago
Posts: 3
June 15, 2019 15:39  

Could AMD be sandbagging, waiting for Nvidia to play their next move before unleashing a Navi monster? There are already leaks about Nvidia's 'super' RTX cards. Let's hope this was what AMD were hoping for? #PrayingToTheSiliconeGods


ReplyQuote
Wowsers
(@wowsers)
New Member Registered
Joined: 1 year ago
Posts: 9
June 15, 2019 20:15  
Posted by: Nortski

Could AMD be sandbagging, waiting for Nvidia to play their next move before unleashing a Navi monster? There are already leaks about Nvidia's 'super' RTX cards. Let's hope this was what AMD were hoping for? #PrayingToTheSiliconeGods

Highly doubt they're gonna have cards earlier than Q4 this year, probably not even until late Q1/early Q2 2020.

What I do believe will happen are two things, first of all something like the first x cards purchased being some percentage off to boost first sales, and secondly close to christmas we'll see a price cut of $50-100. This is all pure speculation.

Very intrigued by 7nm+ which will be used on larger Navi though, apparently it's already yielding the same/better than 7nm (yay for EUV). When we'll see the cards really depends on when TSMC is happy with ramping up production, most likely we won't be seeing cards until Q2 next year, they will however likely cause another price drop because of the massively improved yields of the new cards.

Still don't expect the cards to be disruptive, at best tying with the 2080 Ti if they push 285-300W, speculation as usual.


ReplyQuote
thehumangerm
(@thehumangerm)
Eminent Member Registered
Joined: 1 year ago
Posts: 71
June 16, 2019 03:28  

They have no reason to drop prices unless Nvidia does. Their GPU division has been starved for development capital for a long time and navi while not a smoking price is absolutely competitive in their pricing thanks to Nvidia's price gouging.


ReplyQuote
Anthony Mare
(@anthony-mare)
New Member Registered
Joined: 1 year ago
Posts: 14
June 16, 2019 05:04  

I'll wait for actual performance reviews but I doubt I'll hand my AIO Modded V56 reference down to my kid until the next gen gets here and is a bit more compelling. If they drop the price I might bite but I'm a cheap bastard and I typically don't upgrade until I get at least 25%-30% better at the same price point.


ReplyQuote
hseldon
(@hseldon)
New Member Registered
Joined: 1 year ago
Posts: 1
June 16, 2019 09:03  

Is there a new audio technology with Navi gpus? Like similar to what R290 series had. I read a few rumors but haven't seen an information about it yet.


ReplyQuote
liquid
(@liquid)
New Member Registered
Joined: 1 year ago
Posts: 3
June 17, 2019 17:05  

I should add. 320bit + 16Gbps GDDR6 is also an option for a ~60 CU die. Flexibility in manufacturing cost in that one could use 14Gbps for cut down lower CU count SKUs. It really does not need to be 384bit for bandwidth to keep up.


ReplyQuote

Do NOT follow this link or you will be banned from the site!