Vega/Navi Rumors (Updated)

[IntelUser2000]

GCN 3 2nd gen DCC :

R9 380X: 45.9GBs/TF
FURY X: 59.5GBs/TF
A12 9800 Bristol ridge APU: 33.9GB/s (shared w/ CPU)

GCN4 3rd gen DCC:

rx 580 41.4GBs /TF
rx 480 44.1GBs/TF
rx 560 42.8GBs/TF

GCN5 3rd gen DCC????

Vega 56 39GBs/TF
Vega 64 air 38.1GBs/TF
Vega 64 liq 35.1GBs/TF

GTX 1080:
1850MHz boost = 9.47TF
320GB/s bandwidth

33.8GB/s/TF

A12 9800 Bristol ridge APU: 33.9GB/s (shared w/ CPU)

Shared, so cut it by half. 17GBs/TF

Usually how it works.

 

[CatMerc]

https://www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_Preview/5.html

TPU posted full slide deck. Couple of interesting notes:
▪HBCC increased FPS in Heaven by 7% without aritifical VRAM limit.

▪Calculating from TDP's, being 295W for balanced and 210W for power saving, we get in BF1 4k:
295W for 75.8FPS~
210W for 72.6FPS~

This is quite extreme power scaling.

In thinking about potential bottlenecks and whatnot I had a bit of a play with bandwidth numbers..

I know.. boost clocks aren't always reached, and TF does not equal frames/s throughput.. but it's the only metric of shader power and as we know, there's not a great deal of difference between GCN revisions in this regard.. only delta color compression has improved. Which brings me to the point of Vega having no mention of any improvement over Polaris.

So when you look at bandwidth per TF of raw shader power it gets interesting: Highlighted the stand-out comparisons:

GCN 3 2nd gen DCC :

R9 380X: 45.9GBs/TF
FURY X: 59.5GBs/TF
A12 9800 Bristol ridge APU: 33.9GB/s (shared w/ CPU)

GCN4 3rd gen DCC:

rx 580 41.4GBs /TF
rx 480 44.1GBs/TF
rx 560 42.8GBs/TF

GCN5 3rd gen DCC????

Vega 56 39GBs/TF
Vega 64 air 38.1GBs/TF
Vega 64 liq 35.1GBs/TF


So at max boost, Vega64 liquid has near half the bandwidth per TF of shader power than Fiji. and barely more than a Bristol ridge APU, notwithstanding the APU shares its bw of course, but an interesting pt of reference perhaps..

thoughts?

Vega has more cache and DSBR, both of which save bandwidth.

 

[imported_bman]

GTX 1080:
1850MHz boost = 9.47TF
320GB/s bandwidth

33.8GB/s/TF





Shared, so cut it by half. 17GBs/TF

Usually how it works.

If Nvidia marketing slides are to believed that 1080 has effectively 75GB/s/TF.

One would think there should be some low hanging fruit for AMD given Nvidia can get ~1.8x effective bandwidth from their compression compared to AMD's ~1.3x for Polaris.

 

[IntelUser2000]

If Nvidia marketing slides are to believed that 1080 has effectively 75GB/s/TF.

Sure, but 75GB/s/TF is only true if the competition has zero. Which isn't the case. If we take AMD's presentation and do the same calculation you did Fury X should be at 59.5 x 1.18 = 70.21

One would think there should be some low hanging fruit for AMD given Nvidia can get ~1.8x effective bandwidth from their compression compared to AMD's ~1.3x for Polaris.

Or, this is an area where greater R&D budgets(or to say it roughly like some do, throwing money at it) pay off. Where having experienced engineers + great tools(especially for simulation) bring you steady gains over time which piles up. Like for example generally Nvidia has better drivers and support because they can spend money and resources at it. Those things usually bring small gains when everything is equal and can be easily overcome if you have a vastly superior design(like AMD had with R300) but otherwise something that will tilt the scale.

I'm not sure why Nvidia claims 1.8x there. This says 1.2x, with 1.4x coming from raw increases. They claimed the same for Maxwell. I don't think they improved it for the 1080 Ti, because 1.2x number is from 1080.

Maxwell: http://images.anandtech.com/doci/8526/BandwidthSavings.png
Pascal: https://techporn.ph/wp-content/uploads/2016/05/Nvidia-GeForce-GTX-1080-Key-Features-6.jpg

Maxwell averages: 25% reduction for 33% increase.
Pascal averages: 20% reduction for 25% increase

 

[parvadomus]

Need to remember one of the reasons Vega has such a large die, and so many transistors is is because it has a lot of compute added on. Something GP104 does not have. So while gaming performance may match up with GP104, overall compute does not. Its closer to GP102 in that regard. Not that it match GP102 in efficiency either, just saying its efficiency cannot be directly compared to GP104, which was designed as a gaming card, and ONLY a gaming card. Vega is trying to be two different cards.

AMD said the added transistors where to achieve higher clock speeds. "Vega achieves ~1.7Ghz vs Fijis ~1.05, thanks to additional circuit paths" they said.
Anyway, Vega is not consistently at least 50% faster than Fiji, and I still believe its due to memory bandwidth.

 

[PeterScott]

Well let's check, because CPU <> GPU.

Ryzen is 4.8B with 195mm^2 for 24.6MTr/mm^2

So maybe CPU does = GPU.

That would suggest Intel is coming no where close the theoretical density of their process which is probably based on the size of an SRAM cell.

That is interesting. I remember thinking something like this a while back when AMD was talking about their cache in Ryzen actually being smaller than a similar GB of cache in Skylake, so they had better density than Intel that is always bragging about their process density, but I never dug into more.

I only posted the Intel slide to point out what a Full node should be (.7 Linear sized shrink, double transistor scaling) and then someone pointed out that Samsung said they took two full nodes to get there, which doesn't make sense.

 

[imported_bman]

Yeah, DCC is likely non-trivial and due to limited resources AMD had to choose between implementing tiling or improving DCC. Though it is sad to see a GPU built around HBM hobbled by memory bandwidth.

 

[IntelUser2000]

That is interesting. I remember thinking something like this a while back when AMD was talking about their cache in Ryzen actually being smaller than a similar GB of cache in Skylake, so they had better density than Intel that is always bragging about their process density, but I never dug into more.

It doesn't really matter in reality. So according to AMD Ryzen has 3x the amount of transistors. Heck, the 4 core 8MB module has 1.4 billion transistors, which is much as the entire Haswell 4C GT2 die! Skylake probably still has less than 2 billion transistors. It has the same amount of EUs and caches take the most anyway.

So what are the 4.8 billion transistors doing? Nothing. Ryzen's die is not 3x the size, so it matters zero, whether on die size or performance.

Plus, you can use extra transistors to gain performance. Intel's playing the epeen game as TSMC, Samsung, and Global Foundries are doing.

Though it is sad to see a GPU built around HBM hobbled by memory bandwidth.

Isn't it interesting? Nvidia goes for efficiency, AMD goes for brute force. Neither does both. The results are about on par. It feels like a trade-off. That's why I say a proper cooperation is better than competition.

 

[PhonakV30]

70 to 100Mh/s

https://videocardz.com/71591/rumor-amd-radeon-rx-vega-64-to-be-great-for-mining

RIP Vega Gamers

 

[Karnak]

Holy smokes that's huge.

 

[Yakk]

70 to 100Mh/s

https://videocardz.com/71591/rumor-amd-radeon-rx-vega-64-to-be-great-for-mining

RIP Vega Gamers

That's about what I was anticipating, unfortunately. Demand will be trough the roof!

 

[Snarf Snarf]

70 to 100Mh/s

https://videocardz.com/71591/rumor-amd-radeon-rx-vega-64-to-be-great-for-mining

RIP Vega Gamers

Sweet baby Jesus, it looks like I wasn't wrong when I said AMD would sell all the cards they make

 

[Malogeek]

Good for AMD I guess.

 

[Yakk]

At those hash rates.. Miners area still going to buy a full Ryzen bundle and make ROI quickly... Hot damn!

 

[french toast]

Blimey!

 

[Erithan13]

I'm picturing that 'this is fine' cartoon with the fire made from vega mining rigs (and Skylake X CPUs ayyy....)

Interesting to see the shift from 'Vega sucks it won't sell' to 'Vega sucks it will be sold out' though. Things aren't so black and white now, and just when Vega 56 looked like it was going to put up a good show against the competition. Note that rumor came (indirectly) through Overclockers UK who do try to get GPUs to gamers over miners at least, don't think many other retailers will be so generous with that. If there was ever a time to consider getting a pre-order in this is it.

 

[Snarf Snarf]

Good for AMD I guess.

Indeed, and like everyone thought the FE launch was a total farce considering it was missing a ton of features that are getting enabled with new RX driver . I'd love to sit down 1 on 1 with Raja after a few drinks and find out what really happened to this launch "They gave me a toaster and a box of scrap parts and told me to beat a GTX 1080."

 

[n0x1ous]

Indeed, and like everyone thought the FE launch was a total farce considering it was missing a ton of features that are getting enabled with new RX driver . I'd love to sit down 1 on 1 with Raja after a few drinks and find out what really happened to this launch "They gave me a toaster and a box of scrap parts and told me to beat a GTX 1080."

That reminds of that great series of articles from Anandtech many years ago about Anand talking with Eric Demers and ATI engineers about the backstory to RV770 and Evergreen. Great stuff. I would LOVE something like that about Vega from conception to release.

 

[EXCellR8]

ohhhh great. rumors about Vega being great for mining is not gonna help my chances of unboxing one soon

 

[PeterScott]

Good for AMD I guess.

And maybe gamers who want an NVidia card. Since this might cause the Miners to abandon NVidia.

 

[Head1985]

And maybe gamers who want an NVidia card. Since this might cause the Miners to abandon NVidia.

yeah maybe some cheap 500usd MSRP GTX1080.here in czech republic they cost around 700usd.

 

[CuriousMike]

My guess is even if the hash rate is closer to 50 than 70, they're still going to sell out to miners with great rapidity. (Assuming power consumption is reasonable.)

 

[Face2Face]

70 to 100Mh/s

https://videocardz.com/71591/rumor-amd-radeon-rx-vega-64-to-be-great-for-mining

RIP Vega Gamers

 

[Snarf Snarf]

That reminds of that great series of articles from Anandtech many years ago about Anand talking with Eric Demers and ATI engineers about the backstory to RV770 and Evergreen. Great stuff. I would LOVE something like that about Vega from conception to release.

I volunteer, in my mind it's something that's a combination of those series plus Drunk History. It would be gold, someone get me in touch with Raja.

 

[iBoMbY]

If they are smart they cripple this mining "feature" in RX Vega, via BIOS and drivers (or even better per laser cutting), and only activate it on Vega FE, or even higher priced, products.