Vega/Navi Rumors (Updated)

[EXCellR8]

wish AMD would stop sending me promotional emails telling me to buy the damn cards when I fackin can't yet...

 

[tviceman]

Cred huh?

I'll leave that for you to worry about. I really don't care what you think of me.

I think you're confused of the notion that I might "care." We're discussing products, you got caught defending a really bad rumor that was debunked the moment you were defending, and now you're reflecting. It's cool; people do it all the time when they're wrong.

Instead of trying to say there might be a diamond lodged in the turd because you heard it from your neighbor's friend's cousin, just say it's a turd and that you're going to get it anyways because XYZ. Treat the situation how Titan buyers should be treated: Good for them for getting what they want even when they know it's terrible value (Titan) because they want the best, or like in your case, good for you wanting freshly minted crap (Vega) because your freesync monitor or Nvidia is evil in your eyes. I don't hate as long as reality isn't trying to be twisted.

 

[Paratus]

That is not what Ryan reported at all.

Ryan reported that DSBR was not enabled for Vega FE. That is Frontier Edition. Here is what he wrote:

"The short answer is that no, the DSBR is not enabled in Vega FE’s current drivers."

Vega FE, NOT Vega RX. I am not sure why people are having so much trouble spotting the difference.


Also the endnotes for the architecture slides, with DSBR tests, mentions the results for two different DSBR tests were done on 17.20 and 17.30 drivers.

See:
https://www.techpowerup.com/reviews...cture_Technical_Overview/images/slides-49.jpg
Notes for DSBR test on page 42, uses Radeon 17.30 drivers
Notes for DSBR test on page 43, uses Radeon 17.20 drivers

So yeah, it is there in both.

Yep I missed the same point in another discussion, mistaking FE for RX.

The confusion here wasn't that it was disabled in FE, it was that Ryan stated it would not be available until RX launch:

The short answer is that no, the DSBR is not enabled in Vega FE’s current drivers. Whereas we have been told to expect it with the RX Vega launch

It seems to be referring to DSBR which won't be available until RX launch.

But based on the slide deck from AMD which seems to show data with DSBR for RX already It is probably referring to DSBR for Vega FE not being available until RX launch.

That's the problem with the pronoun game.

 

[3DVagabond]

I think you're confused of the notion that I might "care." We're discussing products, you got caught defending a really bad rumor that was debunked the moment you were defending, and now you're reflecting. It's cool; people do it all the time when they're wrong.

Instead of trying to say there might be a diamond lodged in the turd because you heard it from your neighbor's friend's cousin, just say it's a turd and that you're going to get it anyways because XYZ. Treat the situation how Titan buyers should be treated: Good for them for getting what they want even when they know it's terrible value (Titan) because they want the best, or like in your case, good for you wanting freshly minted crap (Vega) because your freesync monitor or Nvidia is evil in your eyes. I don't hate as long as reality isn't trying to be twisted.

Take a deep breath. You're getting way too wound up. The world won't end either way.

 

[Konan]

The confusion here wasn't that it was disabled in FE, it was that Ryan stated it would not be available until RX launch:



It seems to be referring to DSBR which won't be available until RX launch.

But based on the slide deck from AMD which seems to show data with DSBR for RX already It is probably referring to DSBR for Vega FE not being available until RX launch.

That's the problem with the pronoun game.

So basically it is highly unlikely that there is the existence of a miracle patch for when the cards go retail. Why even take two weeks on it, seems rushed and last minute if so.

I'm assuming the tech world who will be doing reviews will be doing it from the same version than AMD used in their slides, if not one up from that but in my opinion it looks like DSBR is something that may give improvements over time but shouldn't be looked at as the golden nugget here. Doesn't Nvidia already enable this feature (maybe not in the full extent of AMDs update, but to some capacity already). I'm sure I read somewhere from an enabling perspective it's more or less there.

 

[JDG1980]

AMD is apparently distancing itself from CrossFire and other multi-GPU configurations.

The optimistic interpretation is that Navi will include multiple small GPUs automatically acting together as one, so traditional multi-card configurations can be downplayed. The pessimistic interpretation is that this is part of AMD's retreat from the high-end gaming market altogether. That said, Nvidia would clearly like to get rid of SLI if they could, so multi-card GPU setups obviously don't have much of a future at this point no matter which team you go with.

 

[Muhammed]

So basically it is highly unlikely that there is the existence of a miracle patch for when the cards go retail. Why even take two weeks on it, seems rushed and last minute if so.

This is Ryan's comment on the matter:

From the tone of the conversations I had, while DSBR will improve things, everyone was quick to point out that the gains would be higher on a more resource-constrained card. Those aren't the kind of comments I'd expect if they thought performance would make a huge jump with DSBR.

https://forum.beyond3d.com/posts/1993829/

 

[JDG1980]

AMD seems to think memory bandwidth isn't a bottleneck with RX Vega, or else you'd expect the Draw Stream Binning Rasterizer to show greater gains. Assuming this is true, it would seem to highlight the low ROP count and especially the front end (only 4 shader engines) as the likely remaining bottlenecks in the Vega 10 design.

It will be interesting to see, with Raven Ridge, how DSBR performs in a much more bandwidth-constrained situation. Will it provide substantial clock-for-clock improvements in IGPU over Bristol Ridge (GCN 1.2) or not?

 

[xpea]

The anandtech article suggests that the bundle option is to dissuade miners from buying all the cards... Kind of makes sense to me.

At same time AMD is putting new cryptocurrency instructions in Vega ISA...
hypocrisy at his finest

 

[railven]

At same time AMD is putting new cryptocurrency instructions in Vega ISA...
hypocrisy at his finest

If the miner's snatch up the bundles, I expect Wolfenstein 2 keys on the usual markets for <$20! Sign me up!


On the topic of the monitor, reading a few things over at Reddit, it seems the thing doesn't work very well with FreeSync, at least without some tweaks that apparently seem to be hit or miss. One monitor with the coupon, and it seems to not gel well with FreeSync. Good one AMD.

EDIT:

Okay.

http://downloadcenter.samsung.com/content/UM/201705/20170508092544083/BN46-00572A-Eng.pdf

Page 30.

Ultimate Engine: Enable the FreeSync feature with a higher screen frame rate. Screen tearing (abnormal sync between the screen and content) is reduced in this mode. Note that screen flickering may occur during game play.

(this is the mode where you get 48-100 freesync)

So let me get something straight, AMD doesn't have as much say on the specific ranges monitors support for FreeSync? It seems this particular monitor has 80-100hz range, but you can turn on this ultimate mode option to get 48-100hz support, but the side effect (even documented in the manual) is flickering? For $700-900!?

That seems like a feature that shouldn't have been included from the get go.

 

[Krteq]

17.30 driver is not publicly available. It's the driver with DSRB enabled. Stop hanging to straws.

What? Radeon Software 17.7.2 is 17.30 branch driver.

Re-read that linked article

From what we've gathered, the Radeon RX Vega performance figures AMD revealed yesterday include the expected gains of DSBR.

And yet another representatives of web magazines have been told directly by AMD, that DSBR is NOT enabled in currently available drivers, so it makes perfect sense that they had it enabled for tests in lab, but it's not enabled in publicly available driver

Anyway, DSBR can show some performance improvement, but I'm expecting slightly better perf/watt with this feature active.

At same time AMD is putting new cryptocurrency instructions in Vega ISA...
hypocrisy at his finest

Yes, ISA instructions can be "blocked" by driver. If driver doesn't show support for instruction/instruction set, you are not able to use it.

 

[Muhammed]

Radeon Software 17.7.2 is 17.30 branch driver.

This is not Vega driver, 17.30 for Vega is not released yet.

 

[Krteq]

So what's the deal then? Vega FE release driver is from 17.20 branch and DSBR is not enabled there.

 

[lixlax]

What baffles me the most, is the statement that the bulk of the extra 3,9B! extra transistors compared to Fiji were used to get the chip clock higher. And yet there seems to be a clear regression in how much frames it puts out per stream processor/clock- so in wider picture it means they didn't gain any advantage from these transistors. A 1300-1400MHz shrinked Fiji would have done the job as well, but with smaller die area and faster time to get it on the marked (ofcourse it would've needed some changes for never features and HBM 2).
And the worst thing for AMD is that Volta is around the corner and I expect it to increase Nvidias perfomance at least 20-30% for every given market segment. That leaves Vega to compete with Voltas's equivalent to GTX1070 at best (assuming that the numbers AMD have shown for Vega are the best case scenarios) .

Now that AMD finally has a decent product in CPU world, they'll have to take timeout in GPUs to develop something from the gound-up as well?

 

[Qwertilot]

That might be quite hard to do for GPU's I think - Intel have long since reached the point where their increases were quite gradual but NV still seem to be forging ahead very steadily @30%+/annum. You'd have to aim at a hugely ambitious target to do a multi year GPU architecture rebuild.

Suspect that a more realistic hope is that they get enough budget to at least execute what they are building better - with Vega say it'd have looked a chunk more relevant if it had come 6-12 months earlier than it has.

 

[Dribble]

So let me get something straight, AMD doesn't have as much say on the specific ranges monitors support for FreeSync? It seems this particular monitor has 80-100hz range, but you can turn on this ultimate mode option to get 48-100hz support, but the side effect (even documented in the manual) is flickering? For $700-900!?

That seems like a feature that shouldn't have been included from the get go.

And this is the monitor AMD is promoting as a good example of freesync by putting it in the bundle... There is no quality control or min standards - you can label pretty well anything a freesync monitor.

 

[3DVagabond]

At same time AMD is putting new cryptocurrency instructions in Vega ISA...
hypocrisy at his finest

Well if nobody's going to buy it for gaming they have to sell it to someone.

 

[Rannar]

And the worst thing for AMD is that Volta is around the corner and I expect it to increase Nvidias perfomance at least 20-30% for every given market segment. That leaves Vega to compete with Voltas's equivalent to GTX1070 at best (assuming that the numbers AMD have shown for Vega are the best case scenarios) .?

While raising die size (aka GM107 and GM200) and/ or power consumption. 12FFN will probably not be as big improvement as some might think and architecturally pascal is already super efficient - hard to improve on that. we will see.

 

[PeterScott]

So what's the deal then? Vega FE release driver is from 17.20 branch and DSBR is not enabled there.

It's trivial to choose not to deliver a feature to the public, that is still in QA testing internally.

 

[Qwertilot]

While raising die size (aka GM107 and GM200) and/ or power consumption. 12FFN will probably not be as big improvement as some might think and architecturally pascal is already super efficient - hard to improve on that. we will see.

Well we're hardly wildly speculating here.

GV100 vs GP100 is very well known. Die size up a lot for GV100, FP32 up +-50% but the same power budget. Also Xavier (Tegra sized) is due before the end of the year and a big uptick vs their extant Pascal stuff in terms of power efficiency.

The only real constraint on their timing for gaming Volta is when they want to do it which probably relies on the timing for son/grandson of Volta as much as anything else.

 

[PeterScott]

Well we're hardly wildly speculating here.

GV100 vs GP100 is very well known. Die size up a lot for GV100, FP32 up +-50% but the same power budget. Also Xavier (Tegra sized) is due before the end of the year and a big uptick vs their extant Pascal stuff in terms of power efficiency.

The only real constraint on their timing for gaming Volta is when they want to do it which probably relies on the timing for son/grandson of Volta as much as anything else.

I think NVidia is waiting for a mature (read higher yield, more affordable, higher clocking) 10nm process from TSMC. I suspect gaming Voltas performance will largely come from better process (higher density and lower power usage). The 12nm process that the Giant Volta cuda/tensor die used is a stopgap that some say would be better named 14nm+.

 

[Rannar]

I think NVidia is waiting for a mature (read higher yield, more affordable, higher clocking) 10nm process from TSMC.

10nm is soc only. AMD has publicy told they will jump from 14nm directly to 7nm. NVidia will probably do the same from 12nm FFN.

 

[PeterScott]

10nm is soc only. AMD has publicy told they will jump from 14nm directly to 7nm. NVidia will probably do the same from 12nm FFN.

Says who?
http://www.tsmc.com/english/dedicatedFoundry/technology/10nm.htm
"TSMC's 10nm FinFET supports various market segments, including server, graphic processor, mobile, and networking. "

AMD is jumping from 14nm to 7nm because that is what Global Foundries is planning.

I have a lot more faith of getting a successful 10nm process for GPUs from TSMC, sooner than GF can deliver on 7nm.

 

[tviceman]

Take a deep breath. You're getting way too wound up. The world won't end either way.

LOL two deflections with no meat or substance. Like I said, I get that your frustrated. I'm sad for AMD too.

 

[tviceman]

While raising die size (aka GM107 and GM200) and/ or power consumption. 12FFN will probably not be as big improvement as some might think and architecturally pascal is already super efficient - hard to improve on that. we will see.

Given that GV100 is right in line with the same perf/w improvement that Pascal was over Maxwell and Maxwell was over Kepler, you look to be probably off on your prediction.