AMD's Vega architecture previewed at ve.ga

[MorganO]

1 minute ago, Energycore said:

igmur post says found on http://ve.ga

 

word cloud doesn't say much tbh. 8x capacity/stack probably refers to HBM2.

ahh there is sauce for the pasta

[Energycore]

1 minute ago, KOMTechAndGaming said:

ahh there is sauce for the pasta

I don't eat no pasta without sauce.

[MorganO]

1 minute ago, Energycore said:

I don't eat no pasta without sauce.

pasta without sauce is useless

although im genuinely wanting amd to compete

[Energycore]

1 minute ago, KOMTechAndGaming said:

pasta without sauce is useless

although im genuinely wanting amd to compete

What I want is for prices to drop on goddamned RX 480s in my country.

[CostcoSamples]

15 minutes ago, EOZ said:

I wish all the best to AMD Vega but reality is reality i think GTX 1080Ti (GDDR5X) will easily knockout AMD Vega GPU with (HBM2) Memory.

 

The really sad part is that AMD is often a little slow on driver optimization at launch, meaning launch day benchmarks favor Nvidia.  If you do the same tests a couple months later, AMD is often equal or ahead of Nvidia, but by that time they've lost the market share.  Perhaps a lot of the Polaris optimizations will benefit Vega and this time around we'll see a different story.  

[captain cactus]

51 minutes ago, RagnarokDel said:

a word cloud has been leaked if anyone is interested.

So it is a new architecture then, not an iteration on GCN.

 

Ties in pretty much to this Reddit post which I put in a spoiler for you since it's a wall of text (nevertheless a very interesting wall of text):

 

Spoiler

They did. AMD is making it known that Vega is a brand new architecture. They aren't even calling it GCN anymore, but NCU.

GCN refers to a unique layout of the Compute Units, 4 SIMDs, each 16 ALUs, for a 64 ALU Compute Unit setup. This was the core of GCN for Tahiti to Polaris.

Vega being a New Compute Unit, AMD would have done major changes to it's design. 

This definitely has implications. One of which is the AMD FineWine "technology"... AMD optimizing for Hawaii or Fiji or even Polaris benefits older GCN cards because they are the same design.

AMD optimizing for Vega in the future, may not carry through to GCN based GPUs. AMD would have to optimize for both GCN & NCU.

As for die sizes, Polaris is ~232mm2, with the first Vega released rumored to have 4096 SP, it should be around twice that die size, accounting some extra for non-linear increases due to non-SP related transistors to support a higher Sp count. Now, factor in HBM die-space saving by shifting most of the memory controller to the first layer of the HBM stack, we should arrive at ~400mm2 for a Vega 4096 SP HBM2 chip.

Big Vega (if it exists, which I think it does), should be +50%, or right at the 600mm2 mark, matching Fiji and Titan Maxwell & Pascal GP100.

As for what I think Vega's NCU is, there's references to patents for the next-gen GPU architecture from AMD. Some of these patents were discussed prior to Polaris launching, so go back to that time via google if you want the source but it's been covered here and on tech forums already. 

In summary, AMD's next-gen Compute Unit will do away with the identical layout of SIMDs. ie. Their current CU is 4x SIMD of 16 ALU. This design is great for compute workloads, but for gaming, AMD always seem to have trouble fully utilizing those 4x SIMD to the max. This is due to variability in the shaders being processed by games, they don't always saturate or use up all the 16 ALUs of each SIMD fully. Thus you have idle ALUs, dropping efficiency.

AMD's new Compute Unit per their patent is moving to a variable width SIMD, so each Compute Unit may have 1x 16 ALU SIMD and 3x Variable Width ALU SIMD (different ratios of 12,8,6,4 ALU etc), however, each of those SIMDs can combine to process shader that requires 16 width, thus it is dynamic and flexible. Gaming code should in theory better utilize the CU, less ALUs will idle. 

The other patents cover SIMD level clock-gating, which is very fine-grained. Currently it's the entire array of Compute Units that operate at a certain clock speed (for both GCN & up to Pascal). AMD wants to regulate clocks on the SIMD level, so any idled SIMD can down-clock, while active SIMD can overclock to compensate by using the saved TDP of the down-clocked SIMD.

From: https://www.reddit.com/r/Amd/comments/5li8gg/some_extremely_rough_speculation_about_vega/?

 

[MorganO]

12 minutes ago, Energycore said:

What I want is for prices to drop on goddamned RX 480s in my country.

rip, blame the shops/country 

[Energycore]

1 minute ago, KOMTechAndGaming said:

rip, blame the shops/country 

RIP me

 

I'm also waiting for someone to release an unlock 470 to 480 BIOS so I can jump on that

[ivan134]

51 minutes ago, xAcid9 said:

Calm down, this is probably the 1070/1080 contender and not 1080 Ti contender.

I think so too. If the 1080 ti contender was ready, they would have used at the Zen reveal.

[Space Reptile]

9 hours ago, sof006 said:

I wonder how much AMD paid for that domain name.

according to godaddy , between 3grand and 300 bucks 

 

well ve.xyz is 3000€ and ve.ws is 5€ 

[Coaxialgamer]

1 hour ago, FakezZ said:

YES. THANK YOU! Had you written cout << "derp" << endl; I would be all over you about how the using namespace std is bad practise  Worst thing is they teach it at some universities ffs agghhh /rant

well; i usually use namespace std for practicality , but since that was only one line , i had to use std::cout 

[raphidy]

1 hour ago, Scitesh said:

Welcome to the club

Volta will be good if they implement specific hardware to drive dx12 and vulkan api. Pascal is just a patched gpu to make dx12 compatible, like most games in 2016. I guess pascal won't be relevant in the coming years. Better bet on RTG this year for new GPU buyers on the high end. I'll probably sell my 1070 when Vega comes with acceptable price after the early price gauging from the e-tailers. 

[kiwibacon]

cant wait to throw away my 980 tis just doesnt fit my team red build.

[TheFlyingSquirrel]

Been on AMD's high end for 9 years, hopefully they can provide me a reason to buy Vega instead of the 1070.

 

[App4that]

AMD has a bad track record of over hyping their product. If AMD brings a card that matches the 1080 or even 1080ti, Nvidia will just drop the prices and still win. AMD has to forget overpriced crap like HBM2 and use what works. Bring a card that beats the 1080 at a 1070's price. That's how they win.

[Coaxialgamer]

36 minutes ago, Kumaresh said:

When I have to write simple 400 line codes, having to use "std::cin" and "std::cout" seem like such a pain. Much easier to use namespace std. Why is it considered so bad to use it anyway ?

I don't actually know.  I guess there might be conflicts at some point ,  or some errors might be easier to make.  Idk.  Might come down to programming style. 

[CostcoSamples]

57 minutes ago, App4that said:

AMD has a bad track record of over hyping their product. If AMD brings a card that matches the 1080 or even 1080ti, Nvidia will just drop the prices and still win. AMD has to forget overpriced crap like HBM2 and use what works. Bring a card that beats the 1080 at a 1070's price. That's how they win.

That's not exactly accurate.  Nvidia is the king of hype.  It's actually really impressive how good they are at hype.  It's why they consistently outsell AMD even when Nvidia GPUs were worse and more expensive.  AMD could learn a thing or two from Nvidia's marketing strategies.

 

The thing that most worries me is launch day drivers.  AMD is always behind the curve at launch, only to catch up months later, after the initial hype train has been milked.  They need to optimize drivers like crazy and hype Vega like there's no tomorrow.  

[App4that]

Just now, CostcoSamples said:

That's not exactly accurate.  Nvidia is the king of hype.  It's actually really impressive how good they are at hype.  It's why they consistently outsell AMD even when Nvidia GPUs were worse and more expensive.  AMD could learn a thing or two from Nvidia's marketing strategies.

 

The thing that most worries me is launch day drivers.  AMD is always behind the curve at launch, only to catch up months later, after the initial hype train has been milked.  They need to optimize drivers like crazy and hype Vega like there's no tomorrow.  

AMD has been competitive at the low end, and sometime mid range. But AMD has been behind Nvidia at the high end for years. While Nvidia does fluff as much as AMD, Nvidia does it against their own cards. AMD pull crap like the AotS RX 480 test. 

[-iSynthesis]

1 hour ago, App4that said:

AMD has a bad track record of over hyping their product. If AMD brings a card that matches the 1080 or even 1080ti, Nvidia will just drop the prices and still win. AMD has to forget overpriced crap like HBM2 and use what works. Bring a card that beats the 1080 at a 1070's price. That's how they win.

I don't think AMD overhypes anything, its more the community that does it. Everyone seems to expect AMD to beat NVidia and if they don't it's a bad product because it consumes a few watts more or whatever. 

[App4that]

5 minutes ago, -iSynthesis said:

I don't think AMD overhypes anything, its more the community that does it. Everyone seems to expect AMD to beat NVidia and if they don't it's a bad product because it consumes a few watts more or whatever. 

The community does worse than AMD, but they do too. Look at the Fury X, the Pro Duo, Bulldozer, etc. Look at the Ashes of the Singularity test done with two RX 480 that is a test for one game that uses a feature they used to show what they wanted people to see. Nvidia does it too, bad. But like I said Nvidia does it to their own cards for the most part. 

[CostcoSamples]

1 minute ago, App4that said:

The community does worse than AMD, but they do too. Look at the Fury X, the Pro Duo, Bulldozer, etc. Look at the Ashes of the Singularity test done with two RX 480 that is a test for one game that uses a feature they used to show what they wanted people to see. Nvidia does it too, bad. But like I said Nvidia does it to their own cards for the most part. 

Bulldozer was a disaster.  Fury X suffered a painful death due to poor launch drivers.  A few months after release, once drivers were finally optimized, Fury X was on par with 980 ti.  It's been the same story for the last couple generations.  Go as far back as 2008 and AMD was actually WAY ahead yet hardly sold any cards compared to Nvidia due to failed marketing.  

 

It's all water under the bridge.  Right now we need AMD to sell Vega like freakin' hot cakes to get them back in the game!

[SteveGrabowski0]

2 hours ago, CostcoSamples said:

The really sad part is that AMD is often a little slow on driver optimization at launch, meaning launch day benchmarks favor Nvidia.  If you do the same tests a couple months later, AMD is often equal or ahead of Nvidia, but by that time they've lost the market share.  Perhaps a lot of the Polaris optimizations will benefit Vega and this time around we'll see a different story.  

Yeah, exactly what happened with the 480 vs 1060.

[App4that]

1 minute ago, CostcoSamples said:

Bulldozer was a disaster.  Fury X suffered a painful death due to poor launch drivers.  A few months after release, once drivers were finally optimized, Fury X was on par with 980 ti.  It's been the same story for the last couple generations.  Go as far back as 2008 and AMD was actually WAY ahead yet hardly sold any cards compared to Nvidia due to failed marketing.  

 

It's all water under the bridge.  Right now we need AMD to sell Vega like freakin' hot cakes to get them back in the game!

The Fury X has never caught the 980ti. It has caught a 980ti at reference clock speeds in games that favor the Fury X. If AMD had not used HBM and put that money elsewhere I think that would be different. 

 

Like I said, AMD needs a card that matches the 1080 for the price of the 1070 to win. Match the 1080 and Nvidia has the means to drop the prices past where AMD can survive. To do that AMD needs to stop being foolish and use tech that works. Worry about innovation after they beat Nvidia. 

[RagnarokDel]

4 hours ago, zMeul said:

no it does not

reputable tech sites do not spread this shit around; the others do it just to get views and hits

that article was talking about a single core OC to 5ghz. FYI

3 hours ago, Carclis said:

I wonder how much they're stretching these figures and what they're comparing them to. I'm assuming 390x which would still be quite impressive.

the wordcloud is about HBM2 vs HBM1 from what I understand.

[RagnarokDel]

2 hours ago, raphidy said:

Volta will be good if they implement specific hardware to drive dx12 and vulkan api. Pascal is just a patched gpu to make dx12 compatible, like most games in 2016. I guess pascal won't be relevant in the coming years. Better bet on RTG this year for new GPU buyers on the high end. I'll probably sell my 1070 when Vega comes with acceptable price after the early price gauging from the e-tailers. 

From what I read on reddit (BEEP BEEP TRUCK OF SALT) Volta is only a refresh of Pascal.