AMD had claimed for the longest time that Vega had 4 Shader Engines, with a 2.75x increase per clock in geometry (ROP) performance over Fury X. With a 1600MHz instead of 1050MHz clock, that's 4.19x increased geometry performance.
Well they showed a new die shot yesterday, and it's suddenly different. There's two ways to interpret this: >They still consider it 4 shader engines, but each one is split in two groups, and they get that higher ROP performance per clock partially through this splitting up >It's actually 8, and they were sandbagging If it's the later, Vega actually has 8.48x the polygon drawing performance at 1600MHz as Fury X does at 1050MHz.
Also for the retards: This doesn't mean it'll be 419% or 848% total better performance than Fury X. It just will be in raw polygon drawing performance. And lower polygon drawing performance tended to hold GCN back as they were more compute heavy. Good at screenspace effects, less good at lots of polygons and useless amounts of tessellation. But 175%-275% perf is not unlikely at all.
it was the raw throughout that holded back the gcn their polygon drawing was fine
Josiah Robinson
are they giving shader engines the bulldozer treatment?
Cameron Parker
further analysis: It appears it could be the geometry processors and ROPs as the wide rectangles at the top and bottom of the 8 shader engines. So it would appear the instructions are coming in from the outside, then finishing down into the middle, instead of the top-down of GCN. It also appears there are 2 geometry processors and ROPs per shader engine.
But.. again that doesn't make it clear whether they consider this 4 shader engines or 8. I do think they could be considering this as four, but with the data splitting up into multiple geometry processors on either end for whatever reason.
It also looks like the global data share is on one side, and L2 on the other side, and there is over double the area for asynchronous compute engines and the global data share is fucking massive. Bottom is obviously memory controller. Top right appears to be hardware accelerators. The middle between the Shader Engines... I have no fucking clue.
Na it was not enough ROPs for the compute that meant lots of CUs would be idle as there was much more polygon drawing work needed compared to compute work.
More like the Zen treatment. They can more easily cut this down for mobile and smaller GPUs. GCN was fixed at 4 shader engines and you had various degrees of lopsidedness with ROPs and CUs. Same reason you get issues where the RX580 is hardly better at 1080p than the RX570, but is much better at 1440p in screenspace heavy games.
Nathan Bailey
amd's strategy is to clearly blindside their competition
keller already shitstomped intel and theyre in a complete panic
amd knows exactly what failed with fiji
theyve poached the best parts of maxwell/pascal and are about to get a comparable performance from FLOPs with respect to nvidia
Justin Sanders
Oh. It could actually be the geometry processors and ROPs in the middle, and the skinny rectangles above and below them could be the async compute engines. That actually makes more sense because the CUs should be short and wide. And looking at it that way, I do see 4 pairs of geometry processors+ROPs. But if that's the case, the gap between the row of CUs (looking at it turned 90 degrees) makes no sense.
Parker Powell
it was the culling process that was bottlenecking gcn they literally didnt had one for a long time and tried to force everything to pass via the ACE to parallerize them so that they can gain much perf didnt really end up well for them
Jace James
infinity fabric will be used on vega.. i think that amd will clock the memory lower than expected just so that they showcase the full potential of it
Colton Wright
So did this asshole buy all the Vega from AMD? Is this why we've been waiting?
Adrian Wood
its not the first time tho.. apple always buys the higher bins every single chip they can get and its top notch they will pay insane money for it
Caleb Cox
Yes. Lack of culling really hurt Fury. But more polygon thoroughput on top of culling doesn't hurt.
Zachary Lee
amd always have multiple lineups of preformance chips
Zachary Myers
Aren't the higher bins not saturated until later in the production? I mean yields improve over time right?
Evan Jones
he is fighting with powerpc and intelligent right now so he probably will run to amd lol
lets hope he gets a apu with a seperate igpu and hidpi screen with display port and freesync
>wishful bsd dev
Jaxson Butler
they used the same technique as ryzen we dont know the yields but im pretty sure its way too high
Adrian Hill
Dont regret the wait now. Shouldve never doubted AMD
Parker Lewis
I just found out that nvidia can only use 1 of the wayland compositors
amd supports 7+
once again nvidia wants to create their own shit instead of a standard protocol
just like fucking egl reeeeeeeeeeeeeeeeee
Lucas Butler
Yeah and? Its still slower than a 1080
Hudson Cook
>reeeeeeeeeeeeeeeeee Just sell it second hand.
Lincoln Richardson
>Yeah and? Its still cheaper than a 1080 >Yeah and? Its still freeer than a 1080 >Yeah and? Its still less gimped than a 1080 in 2 years
Landon Collins
Fury X at 1600Mhz can't be slower than a 1080.
Stop shitposting.
Jaxson White
Didn't this happen with Maxwell, splitting the SM into smaller units for better utilization?
>More sandbagging? I doubt it. If they had something impressive they'd show their hand by now. Simple as that. They have nothing to gain at this point by Sandbagging. VEGA is going to disappoint you if you think AMD are still hold back.
>They still consider it 4 shader engines, but each one is split in two groups This how they arrived at the 2.75x increase in the first place obviously. Pulling an extra 2x out of your ass is pretty desperate.
It's a cool die shot though, you can see all the space dedicated to the infinity fabric logic pretty clearly.
Jacob Foster
If it has the same tflop/perf as Tahiti, it will be a monster.
Jose Morgan
>they'd show their hand by now
No, because they learned their lesson on what happens if you over hype something with Ryzen.
Jose Reyes
You meant Polaris and Fiji. If Vega is half a good arch as Zen is, it would be a miracle.
Bentley Edwards
Vega is what Fiji failed. considering Fiji's perf after seeing Vulkan-DX12 and is LN2 overclocking numbers, I am very hyped for Vega.
Carson Bell
Someone fire up paint and mark the hardware components there, please.
Juan Howard
What's that between the shader array? Looks quite significant in size.
Chase Long
But it's still only 64 rops, bottleneck....?
Dylan Ross
see
Caleb Diaz
Is this design more scalable?
Samuel Long
Has AMD got another winner on their hands? Is this their RV700?
Jacob Davis
I want release dates and prices
Tyler Hughes
July. 500$+ starting.
Levi Edwards
>Is this their RV700 it seems so.
Noah Powell
You telling me low end Vega is 500?
Samuel Peterson
What lowend Vega? Thats months away, you got big vega and bigger Vega now.
Luis Williams
we don't know the line up. we know there is 56 CU Vega (big vega) and 64 CU(bigger vega) Vega thanks too Apple. no price, nothing.
Kevin Hill
So and is pulling a fury/Titan meme release? We're they rebrand a third time for everything else?
Jose Russell
I feel this is more Tahiti than Terascale, Tahiti was really good, could OC like mad, no bottlenecks.
Michael Bailey
im pretty happy with all their offerings but i hope they have some lower power consumer stuff with aes and iommu for muh servers
intel only cares about the enterprise
Matthew Edwards
???? Where's the AMD BTFO DEAD BANKRUPT line? No..No. You're not are you. No way. No way, get fucked. No way get fucked cunt, fuck off.
You can't be seriously having a serious GPU discussion on Sup Forums of all fucking places.
Adrian Howard
Fuck off fag, this thread is pure cause summers fags are asleep
Ethan Jackson
agreed vega vulkan is going to be great
Cooper Sanchez
Hold me as I weep tears of joy.
Jaxson Martinez
How does geometry throughput of polsris/fiji compare to Pascal? I think I saw techreport doing the tests, but fuck me If i can find em while at work on a phone.
Juan Reed
bask in thine glory for it is short lived
Luis Thomas
inb4 it ends up like other amd cards with high theoretical power but it will still be shit at gaymen
Connor Sanders
Around two to three times lower.
Josiah Barnes
And so summer awakens
Michael Jenkins
>trying to interpret anything from this """"""die shot"""""" uhh >AMD had claimed for the longest time that Vega had 4 Shader Engines no they didn't in fact they basically said that it would work over more than 4 engines see: anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2
Dylan Young
AMD just needs some good value chips to retain marketshare, it can curbstomp Nvidia once it gets enough money from their CPUs
Jayden Lewis
Yeah. I don't understand when people say AMD support on Linux is bad. Maybe if you're on any early version 3 of the kernel or earlier.
Now days dies usually sensors on them to easily test and bin themselves very quickly and easily.
Even a Fury X with a 50% clock increase beats the 1080 by a lot. Vega is much more than simply just a 50% clock increase. Retard.
A lot happened with Maxwell. Maybe. idk. I don't follow Nvidia architecture much because they are so secretive and you often don't figure out features until YEARS later of reverse engineering.
Yes, I disagree with his analysis, if you read the OP.
IPC per ROP is 275% higher. Also don't know if it really is 64 ROPs.
Not really. Halo products sell the midrange. Consumers are dumb.
Lincoln Reyes
Way too much arch changes to gauge performance off a 1600mhz Fury X.
These frontend changes can be anything from 10% to 40% increase in games alone, the latter for particularly shitty ones lupike Fallout and Dishonored
Nolan Gray
No one is doing so really. It's just the baseline minimum performance increase without arch changes would be ~50% over Fury. But obviously it's higher than that.
Jason Price
The cherry on top would be if it can actually overclock.
Jack Edwards
>they learned their lesson on what happens if you over hype something with Ryzen. How is saying a 40% IPC increase but delivering 52% overhyped? Unless you're one of those people that listened to the shills claiming the world about it just so they can pull this trick, that it was 'overhyped and didn't deliver'.
Anthony Johnson
Well, Samsung shat out a very impressive quantum dot freesync monitor a few days ago, makes me think there's some meat to Vega.
Kevin Gutierrez
Haven't most of Samsung's monitors these days been FreeSync supporting?
This one is 144Hz, quantum dot, VA panel with 3000:1 static contrast, and freesync 2, meaning HDR10 and 125% sRGB coverage. If it was 4k, it would have been literally perfect.
And will probably cost a good $700 at least.
Andrew Diaz
Wat is this quantum dot meme
Evan Ward
Doesn't it still have black shift issues since its a VA panel? That stopped me from buying a CF791.
David Barnes
Personally if I had to choose between 1440p with that contrast and HDR and 4k, I'd pick the contrast any time of the day, most 4k high refresh screens are still around 1000:1
Lucas Nelson
>175%-275% Hnnnnggg, give me some budget Vega cards for that sweet altcoin mining
Camden Scott
Wait for tftcentral reviews. I don't think they'd be advertising it as a premium gaymer monitor as is, and IPS has its own issues.
Jacob Brooks
user what altcoins are worth it?
Isaiah Anderson
I don't know any current ones, but there are always coins worth mining. I sold my mining rig a good while ago, I now only trade crypto, but I'm looking into mining again.
Isaac Fisher
You can hype and use fancy marketing crap as much as you want, it still won't change tje fact that it's barely faster than a 1070 a year later
Anthony Wood
No, it's literally RV700 2.0, this time Pajeet made Fiji actually work.
Oliver Davis
No, it's probably Q4 2017-Q1 2018 for Vega11.
Ethan Phillips
Hello, fuccboi. AMD gave you the die shot explicitly to shut up niggers like you.
Nolan Harris
unfortunately those fuckboys will be chimping out until we see vega.
then they go REEEEEEEEEEEEEE
Owen Sanders
You can keep crying as much as you want.
Justin Powell
So what's the term 'sandbagging' in this context. >4 but actually 8 Sounds good?
Jayden Ortiz
It means presenting something worse than it is, probably to slow down the competitors response time.
A good example is Zen, they promised 40% IPC, we got 52%. Sandbagging.
Cooper Garcia
Good question, what is it? Sounds like a marketing gimmick
Luke Diaz
Its just the name of a specific type of pixel structure used in the display.
Thanks. I was confused as in today context it's to stop flooding waters. so comes off as preventing the inevitable. Like they're lying to themselves that everything going to be alright. If anyone wants to drop origins of the phrase you're more than welcome.
Carson Ramirez
It's exactly that. Everyone assumed Vega still had 4 geometry engines, while in reality, as seen in the dieshot, they went for more balanced approach.
Henry Roberts
Thought so. More is usually better. Any good resource that would help the layman look at dieshots and understand what they're looking at. Lurking moar can only get one so far.
Nathan Price
My guess it's alluding to using sandbags as jettisonable weight. So sandbagging would be intentionally downplaying upcoming products so you can then "jettison" the "weight" holding you back and shoot ahead of the competition, taking them by surprise. That's just my understanding though
Josiah Bailey
This tactic won't work much longer, Nvidia knows what RTG is capable of, and they're not slowing down. Unlike Intel who has their head so far up their own ass they only noticed after AMD ate their entire SKU lineup overnight.
Michael Bailey
That's a very good explanation and also reverts back to the beginning of the technology era where the balloon was invented. I'm pretty sure I've herd that expression used in papers at that time. Not that I'm of that age. I've researched many things. Often turns of phrases can be deduced with logical thinking. Thanks again.
Justin Turner
>they're not slowing down. Ye, they bolted fucking ASIC on oversized die in desperate attemp to get into datacenter market.
Sebastian Lopez
Fixed function and specialized hardware is always a good thing. Pretty much everything in large scale is moving from general purpose to specialized hardware.
Parker Thompson
That's why everyone will move away from GPGPUs to FPGAs and ASICs. GV100 move reeks of desperation.
Hunter Price
>Unlike Intel who has their head so far up their own ass they only noticed after AMD ate their entire SKU lineup overnight. Their management, yes. But please don't label their engineers as being complacent in status quo. Well, not because they don't strive for better. We're living in times where such a company has political persuasion. And aiding their products gives them power. That's not to say a smaller company does not have the ability to persuade government. It's just some do it for worse outcomes.
Gabriel Bailey
Intel has one good engineering team: the dudes behind EMIB. That's it.
Bentley Russell
The window is already gone. Ethereum will be mined up before the end of the year. Card prices skyrocketed back up with the BTC boom this year.
Buy a 1070 right now today and it might pay for itself. A vega card would never pay for itself. Frontier edition will be 1000USD, 30 MH/s, 250W calling it now.
Austin Young
Their NICs are good and Optane is really good if it wasn't so unholy expensive.
Luis Murphy
They don't want Nvidia to start selling Titan Xp for 400$ before Vega release.
That's why they are pretty much airtight about it's performance except computational applications.
Elijah Gomez
>Their NICs are good Ah, yeah. >Optane is really good Fuck no it dies overnight. And we're talking R&D.