/w4v/ - Daily Wait for Vega Thread

Hi all,

This is where we wait for the latest AMD graphics card which AMD will surely release in a timely manner and will definitely blow all of our minds, no questions asked, 100% guaranteed, no chance it'll fail, just you wait.

Last time: VEGA'S GOING ON TOUR. If you're in Budapest or Portland, go take some videos and leak them here, then leak them to VideoCardz so they can then be leaked here too:
radeon.com/en-us/rx-community-meetup/
videocardz.com/70895/amd-radeon-rx-vega-is-going-on-a-tour

The latest:
videocardz.com/70928/amd-radeon-vega-frontier-liquid-edition-is-now-available
videocardz.com/70890/amd-radeon-rx-vega-xtx-vega-xt-and-vega-xl
3dcenter.org/news/erste-informationen-zu-den-einzelnen-modellen-von-amds-radeon-rx-vega (Deutsch)
3dcenter.org/news/der-release-fahrplan-der-herstellerdesigns-zur-radeon-rx-vega (also Deutsch)
videocardz.com/70838/gpu-base-boost-typical-and-peak-clocks-whats-the-difference
gamersnexus.net/guides/2977-vega-fe-vs-fury-x-at-same-clocks-ipc

Other urls found in this thread:

llvm.org/docs/AMDGPUUsage.html#processors
twitter.com/NSFWRedditVideo

bugger off m8

Don't like it? Don't respond. Better yet, tell us how we're all in for Another Massive Disappointment.

>3 posts
>2 IP addresses

got you now

Nice overclocks on gamer nexus' hybrid watercooled Vega FE.
Plenty of headroom for the RX to be higher.

Actual good news about wega? No wai.

If vega blows I'm going to buy threadripper

Kinda, a very stable and constant 100mhz+ on core and memory but also another 100w's. 400w average from the cables alone.

Water cooling also drops the leaking though, which means these might not be great cores/memory and a more mature process will gain a lot for this setup.

What would be the price ceiling for this?

Don't know pricing on RX Vega yet. FE Vega is $1500 for the watercooled and $1000 for aircooled. Keep in mind this is not "gaming ready" GPU at the moment. It's for prosumer market.

How much RX Vega must be for it to be "too much"

Reminder

You can already get 1080's for $500. I'd say they need to price it at that or lower.

You forgot to mention that he found that it was drawing 400W through the PCIe power cables alone, plus whatever it sucks from the PCIe slot (another 75W). Close to 500W total power draw to just about edge out a GTX 1080 drawing a third of that. Or to put it another way, more power consumption than a 295X2. For a single GPU.

Great job, Raja!

>4k ALUs running at INSANE clocks draw fuckton of power
No shit sherlock.

>480mm2 die with HBM2
>cheap

Is Vega going to be a loss leader?

Maybe it's time to rethink your architecture at that point, rather than shitting out another underwhelming housefire that will undo all the good will Ryzen has generated.

Wait for Drivers(tm).

>Maybe it's time to rethink your architecture at that point
They did.
And don't worry, every card will be a housefire till we reach MCM.

They can't afford to, there's a reason why all their video card releases since 2011 have been minor iterations upon GCN1.0.

Hopefully with Zen's success, RTG will have the budget to develop a new architecture at last and put graphics core next to rest.

Too bad Vega is not a minor iteration.

That was the lie AMD attempted to peddle, but evidently it isn't too much of a departure. Same old hot, loud, power guzzling GCN with disappointing performance.

>l-lies
Lmao.

Jesus man, no Vega yet? Holy fuck.

There's Vega FE.

Nah nah, I mean the one for video games.

That will be very very very late, hopefully early September for AiB cards.

Oh no, whatever are you going to do???

Reference cards will be out much earlier than that.

Maybe. AiBs get final bios for RX Vega on 2nd of August.

Isn't AMD the inventor of HBM? So they don't have to worry about patent use costs?

They are.

It is a joint venture of AMD and SK Hynix, but HBM is still expensive to manufacture regardless of who invented it.

You say that like 10 high speed GDDR5X modules are cheap.

You're looking at some $20-40 total BoM cost top for the 8GB Vega, 16GB is a different story.

...

10/10 made me chuckle

At least they met Q2 target.

>when real Vega was supposed to release Q3 2017(Q3 2016 tapeout)
>when real Vega will compete with Volta

Vega FE is real Vega.
The drivers are not there.
At all.
Hey at least ROCm works and seems decent.

>Q3 2016
This is what bugs me, tapeout was way too late for just a mere Pascal competitor.
And I personally know AMD didn't redirect funds from RTG to Zen until middle of 2016, which would not impact Vega's tapeout at all.


There's something way to fishy with Vega, and considering AMD has done literaly ZERO damage control for the terrible 'gayming' reviews of FE, they're hiding something

New shader stage is alone worth discussing, shame RTG gives us ~zero actual low-level info about Vega.
At all.

Of course because they haven't posted it yet, there's no press slides or presentation.

>met
>"""""""met"""""""

Maybe they'll tell us more at SIGGRAPH.
As for now, they are totally silent.
And they were for the whole Vega R&D cycle.

Duh. They "met" their Q2 target.

Hey, silicon's out.
Deadline met, investors won't sue.

The driver or worst case a hardware defect are a entirely separate case.

I don't think they'd sell defective chips at ebin prosumer cards.
If that was the case they'd dump it to someone much like Intel did with Early Purley.

>still not /w4n/

Why was AMD trumpeting about Zen from the rooftops in the leadup to release, but are barely acknowledging the existence of Vega?

Navi is pretty far away and we've yet to see properly working Vega.

Here's a compilation of Vega FE issues that seem completely abnormal.


>pro and gaming mode don't do anything
>power gating/AVFS seem disabled
>geometry performance worse than Polaris
>stream binning rasterizer not working
>total usable bandwidth much worse than Fiji, leading to bottlenecks above 1400MHz
>texture performance FAR worse than Fiji


Assuming drivers are the cause for all of these, you're not looking at a mere 20-30% increase in performance from drivers, not even close, this is a gutted chip at the moment, the texture performance and bandwidth alone are fatal.

They were?
Zen had like 2 (or even one, the one with 3ghz ES) demos before New Horizon event.

>they disabled power gating
>even fucking Primitive Discard Accelerator from Polaris
WHY?

agree Itc clear that the chip wasn't finished at all make all the excuses you want this is just dumb form a company point of view.

Oh, I forgot.

>core voltage stays 1.2V stays regardless of clockspeed

If you haven't noticed, RTG has a different marketing team, and take a look at their previous release which they praised to the moon.
It's probably radio silence until they figure out when they can get the drivers working, I guess the AM4 APUs and late release of Raven Ridge(compared to Q1 Summit Ridge) are a pretty good indication of the driver situation

The silicon was finished like ages ago.

its because its still runni9ng with modified fiji and that chip didn't have those features.

we are not taking about cpu's dummy

If you look at the radeon source code, you'll know power management is the largest part of the code, and considering Vega is hinted to use a lot of Zen's power management, it's totally not ready.

>chip didn't have those features
What? Its full Vega10 XTX silicon.

the bugs where clearly not fixed on silicon and it looks rushed.
this is not a software bug its clearly just defective.

>>>core voltage stays 1.2V stays regardless of clockspeed
It gets better, you can achieve stable clocks (1400MHz) at 1.0V
Vega Fe is overvolted to hell, hopefully the same is not true for RX because there's no way they'll be able to meet the 12TFLOP at 225W they said last year.

If the reference RX VEGA can match or beat the Ti the AIB RX's will absolutely obliterate it.

>Please don't just be slightly faster than a 1080. PLEASE!

It it was defective it would never be sold as a prosumer card, dummy.
They'd dump the chips as a compute boards to chinks, Amazon or Google, much the same way Intel did with Early Purley.

GN Hybrid's got considerably closer to 1080 Ti, so not having Radeon Pro software is gonna make things even better.

Could FE be the ultimate sandbagging?

We know nothing about Vega right now.
Hell, DOOM demo from January performed better than Vega FE.

It's was the only way to meet Q2 target.
At least ref looks nice and PCB design is stellar.
On pure compute?
Hilarious.

If they turn on TBR and don't use retarded stock voltages that alone would lower the power draw significantly, leading the card to boost to 1600MHz all the time.
The others are just icing on the cake, the TMUs being like 40% slower than Fiji even at 1400MHz is really worrying though, could Vega not even use TMUs, some other transform stage?

>TMUs being like 40% slower than Fiji
What the fuck?
Like really, I've never seen launch this confusing.

With threads like that, you're setting yourself up for one.

Is Vega still GCN in anything but ISA?
Because all this shit working literally backwards is weird.

llvm.org/docs/AMDGPUUsage.html#processors


Shit's very different, that can only explain the driver situation.
Pic related.

Oh, that explains Greenland/GFX9 being so late.
It's basically a totally new uarch that simply retains old ISA.

Someone on B3D forums was actually implying that Vega has a bunch more registers than Fiji, leading to the far improved compute performance even at the same clockspeed.
Guess that would make sense, but why isn't it doing anything for gaming?

Drivers?

Some more GFX9 specific improvments

>For GFX9 the appature base addresses are directly available as inline constant registers SRC_SHARED_BASE/LIMIT and SRC_PRIVATE_BASE/LIMIT. In 64 bit address mode the apperture sizes are 2^32 bytes and the base is aligned to 2^32 which makes it easier to convert from flat to segment or segment to flat.

>The global segment can be accessed either using buffer instructions (GFX6 which has V# 64 bit address support), flat instructions (GFX7-9), or global instructions (GFX9).

>The M0 register is not used for range checking LDS accesses and so does not need to be initialized in the prolog.

In any case, Vega seems to be easier to program for at first glance

>Maybe it's time to rethink your architecture at that point

That's what Vega was supposed to be, but they had to put in a legacy graphics path so the card works without 1 year of driver development. Now they only need 6 months to get it to perform good.

>It gets better, you can achieve stable clocks (1400MHz) at 1.0V

How much power does it use that way? I wouldn't mind running it at just that, even if it gets 1070 performance, if it means that the card runs icy cool.

What fucking 1070 performance?
Vega11 is far away.
And they are yet to make Vega10 actually fucking work.
Like, read the thread.

Vega 11 is probably cancelled. We haven't heard shit about it.

Why would they cancel it?
It's a smaller die since you can't bin Vega10 into every niche.

>What fucking 1070 performance?

Wow, that alone triggered you so hard that you didn't even comprehend the rest of my post, didn't you?

Vega FE currently is between 1070-1080 performance.

If undervolting and underclocking to 1.0V 1400MHz means that Vega XL-TX-ABCDX silicon runs icy cool, I'd rather have that, even if it means the card is now reduced to 1070 performance.

I'm running a 3570k on stock reduced to 0.95V too (partially because it couldn't OC for shit).

You won't buy FE anyway and we don't know RX performance yet.
At all.

It's the same fucking chip, only board and drivers will differ. Performance will be, at minimum, Vega FE levels. If we are lucky, it might get more. But like I said, if it performs 1070 level and stays icy cool with some undervolting, that's already fine enough for me (assuming it is priced accordingly and not some bullshit $400 price tag because "lol premium card").

So basically a $100 undercut on price?

That honestly isn't enough. $200 would be, but I have to buy a bigger PSU for Vega and that's an extra $100 right there.

>$100
Where did you get this?

It sounds like the silicon design itself is fundamentally fucked.

There were rumours that Vega had some sort of last-minute design problem which required a new tapeout or whatever and Vega FE is the broken shit that they rushed to market to meet the quarterly deadline, which is why nothing works - the silicon is fucked.

Why would they dump defective chips as prosumer cards when you can give them to meme learning dudes?
Intel did the same with Early Purley.

Nothing that couldn't be fixed with a respin, which would take 3-4 months.

Question is, if the fuckup exists, when did they find out about the fuckup, and when did they start the respin?

Canadian money

so a $10 difference in american is $100 in canadian pesos

or close enough/whatever

>Why would they dump defective chips as prosumer cards when you can give them to meme learning dudes?

dunno, but if they did dump them it was probably to meet the quarterly deadline so they didn't get sued by their investors.

I would strongly argue they found out/knew about it at least around the time of that dual vega shitshow demo, and probably a few weeks earlier than that.

They won't dump defective chips as a. retail prosumer cards. That's lawsuit material.
There's no silicon bugs, since tapeout was ages ago, they had all the time in the world to do a respin.

So Vega will start at $700 if they fix the performance, $600 otherwise?

>They won't dump defective chips as a. retail prosumer cards. That's lawsuit material.

Yes they do, it's called binning.

Binning has nothing to do with selling defective (as in uarch features are not working) silicon.

>tfw you're essentially forced to unironically wait for Vega anyway

Just buy nvidia?

>tfw MxGPU never ever
I don't want a billion VMs, just let me share between the host and one guest pls. I hate to dualboot, but I still like my games.

I was about to buy a 1080, but I don't feel to confident in buying a $540 GPU that's more than 1 year old