Vega/Navi Rumors (Updated)

[Glo.]

I find VCZ to be far more reliable than TPU or WCCF.
Taking their values, for the 16GB 6864 HBM part, it would come to 820 GB/s which is respectable. But I would suppose this would be the Instinct SKU.

I would suppose a top Consumer card to be 8-12 GB using 4-12 Hi Stacks for 410 - 615 GB/s considering Koduri mentioning considerable flexibility with the Memory for various SKUs.

RX Vega GPUs are designed to work with up to 2048 bit memory bus.

 

[DisEnchantment]

RX Vega GPUs are designed to work with up to 2048 bit memory bus.

Yes, I have been going through the gmc v9.0 sources on 4.13 wip.

I mean its 204.8 GB/s / 4x Hi Stack. So wouldn't that imply the 16GB part to be 820 rite?

EDIT: Now I see so that would be max 409 GB/s

 

[Sweepr]

Not sure if already posted...

RX Vega Core - $ 399 ;
RX Vega Eclipse - $ 499;
RX Vega Nova - $ 599

June 5 for global sales.

Edit: old, and fake?

 

[Magic Hate Ball]

So if we're looking at two stacks of HBM2 4x high, we're talking ~410 GB/s currently?

That's about 60% higher bandwidth than the stock RX480 was, so if memory becomes the bottleneck, they'll have to hope their HBCC memory trickery is enough to make up for it to hit GTX1080 levels?

Unless they had a second 2048bit memory bus built into the Vega 10 chips for future use in Pro-sumer cards that they didn't plan on using for the desktop line they've had to utilize?

If I remember correctly, the HBM memory interfaces are much smaller than the GDDR5 ones, so maybe that's on the die already? Like how Ryzen has server components totally un-used on the consumer CPU's because they're all one piece of silicon?

 

[crisium]

The reddit post that never dies.

 

[Bacon1]

Not sure if already posted...

RX Vega Core - $ 399 ;
RX Vega Eclipse - $ 499;
RX Vega Nova - $ 599

http://news.mydrivers.com/1/532/532370.htm

June 5 for global sales.

LOL the debunked names that keep on giving.... The guy even changed the names in his reddit post because people didn't like the ones he picked initially..

 

[DisEnchantment]

So if we're looking at two stacks of HBM2 4x high, we're talking ~410 GB/s currently?

That's about 60% higher bandwidth than the stock RX480 was, so if memory becomes the bottleneck, they'll have to hope their HBCC memory trickery is enough to make up for it to hit GTX1080 levels?

Unless they had a second 2048bit memory bus built into the Vega 10 chips for future use in Pro-sumer cards that they didn't plan on using for the desktop line they've had to utilize?

If I remember correctly, the HBM memory interfaces are much smaller than the GDDR5 ones, so maybe that's on the die already? Like how Ryzen has server components totally un-used on the consumer CPU's because they're all one piece of silicon?

In the code, there is a calculation for memory width and the max is 2048 in the branch where 16GB 6864 is present.
Could be that they have not updated everything, but lets see if there are new commits.

 

[Magic Hate Ball]

In the code, there is a calculation for memory width and the max is 2048 in the branch where 16GB 6864 is present.
Could be that they have not updated everything, but lets see if there are new commits.

Yeah, it is wild speculation on my part, but based on the fact that they've had to take the Swiss Army Knife approach for silicon in the past, I wouldn't be surprised.

 

[zinfamous]

I hope that, at the very least, the CFO or whomever at the Analysts meeting tomorrow, lets out a fart that smells faintly like Vega.

That would be better than the any info we have gotten over the last several months.

 

[Arachnotronic]

I hope that, at the very least, the CFO or whomever at the Analysts meeting tomorrow, lets out a fart that smells faintly like Vega.

That would be better than the any info we have gotten over the last several months.

Koduri will be doing an RTG-specific preso. I would be shocked if he didn't talk about Vega and Navi.

 

[CatMerc]

In the code, there is a calculation for memory width and the max is 2048 in the branch where 16GB 6864 is present.
Could be that they have not updated everything, but lets see if there are new commits.

Could you link me to that?

 

[Head1985]

Not sure if already posted...

RX Vega Core - $ 399 ;
RX Vega Eclipse - $ 499;
RX Vega Nova - $ 599

June 5 for global sales.

Edit: old, and fake?

Already posted like 10x and its old fake from some dude on reddit.Funny thing its everywhere now even on guru3d.Most sites are just clickbaits....

Btw i hope names will be like this
AMD vega X
AMD vega
AMD nano
Not some nova crap lol...

 

[Paratus]

https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock

Since I've been having a weird sort of fun estimating where Vega could end up, let's see what 1600 MHz Fury X and Polaris could do.

If the other rumors are right and Vega is:

• TMU 256
• SP 4096
• ROP 64

Then let's look at Fury X, which is the same layout at 1600mhz (52% OC)

According to TPU:
Fury X vs 1080Ti: HD 60.3%, 25x14 57.6%, 4K 56.7%

Apply the 52% OC:
Vega based on 1600mhz Fury X vs 1080Ti:
HD 92%, 25x14 87.7% 4K 86.4%

Not enough to catch the 1080Ti


For Polaris let's use RX 570. It's exactly half of Fury X / rumored Vega

Standard:
RX570 vs 1080Ti: HD 47.2%, 25x14 42.0%, 4K 38.3%

Vega based on 2x RX 570 @ 1600mhz (28% OC) vs 1080Ti:
HD 121%, 25x14 108%, 4K 98.6%

So 2X OC'd Polaris is in the ball park to 20% faster than the 1080Ti.

The estimates for the higher resolutions are probably more wrong due to the 570's performance falling off sharply. Vega with 400+Gb/s of HBM2 shouldn't have that problem at 4K.

What I don't know is how much of an increase Vega will have over Polaris CU to CU.

Comparing Fury X to 2X 570 already shows an increase in performance:

RX570 x2 @1050mhz vs Fury X
HD 132%, 25x14 123%, 4K 114%
(Polaris looks to be a much better balanced design)

2X 570 is about 464mm^2 and 11.4B transistors. That's smaller than the rumored 500+mm^2 Vega. Since the 570 is a harvested 580 there's actually 72CUs instead of the rumored 64CU design. That leaves a budget of at least 1-2 billion transistors for:

• HBCC
• HBM2 mem controler (delta from the Polaris GDDR5 256bit controller)
• HPC features
• Performance enhancements

So after all this rambling I'm still left with the numbers showing Vega likely to be at worst on par with the 1080Ti and more likely 10-20% faster. With an outside chance of it being greater than 20%.

I could be off base, but it doesn't seem likely to me.

 

[raghu78]

Since I've been having a weird sort of fun estimating where Vega could end up, let's see what 1600 MHz Fury X and Polaris could do.

If the other rumors are right and Vega is:

• TMU 256
• SP 4096
• ROP 64

Then let's look at Fury X, which is the same layout at 1600mhz (52% OC)

According to TPU:
Fury X vs 1080Ti: HD 60.3%, 25x14 57.6%, 4K 56.7%

Apply the 52% OC:
Vega based on 1600mhz Fury X vs 1080Ti:
HD 92%, 25x14 87.7% 4K 86.4%

Not enough to catch the 1080Ti


For Polaris let's use RX 570. It's exactly half of Fury X / rumored Vega

Standard:
RX570 vs 1080Ti: HD 47.2%, 25x14 42.0%, 4K 38.3%

Vega based on 2x RX 570 @ 1600mhz (28% OC) vs 1080Ti:
HD 121%, 25x14 108%, 4K 98.6%

So 2X OC'd Polaris is in the ball park to 20% faster than the 1080Ti.

The estimates for the higher resolutions are probably more wrong due to the 570's performance falling off sharply. Vega with 400+Gb/s of HBM2 shouldn't have that problem at 4K.

What I don't know is how much of an increase Vega will have over Polaris CU to CU.

Comparing Fury X to 2X 570 already shows an increase in performance:

RX570 x2 @1050mhz vs Fury X
HD 132%, 25x14 123%, 4K 114%
(Polaris looks to be a much better balanced design)

2X 570 is about 464mm^2 and 11.4B transistors. That's smaller than the rumored 500+mm^2 Vega. Since the 570 is a harvested 580 there's actually 72CUs instead of the rumored 64CU design. That leaves a budget of at least 1-2 billion transistors for:

• HBCC
• HBM2 mem controler (delta from the Polaris GDDR5 256bit controller)
• HPC features
• Performance enhancements

So after all this rambling I'm still left with the numbers showing Vega likely to be at worst on par with the 1080Ti and more likely 10-20% faster. With an outside chance of it being greater than 20%.

I could be off base, but it doesn't seem likely to me.

Performance never scales 100% with increased clocks or with shaders. For Vega to even match 1080 Ti the perf/sp has to increase significantly.

 

[Cookie Monster]

In my opinion its very hard to estimate how VEGA might perform based on current GPUs unlike potential Volta based Geforce products.

AMDs GCN architecture first and foremost doesn't clock very high when comparing against their competition. Whether its due to the architecture or process node or both its hard to tell unless there was a 16nm TSMC version of say polaris P10.

If they wanted to reach high clocks for a given SKU almost all their GPUs had to run outside their ideal voltage to clock region as seen by recent Polaris rebrands (and their power consumption figures) and undervolting results/fury nano.

Secondly VEGA will most likely have a TDP or a thermal enevelope target of 250W or maybe even 300W. This is typically normal for a high end GPU. Most of the performance estimates looks at doubling of polaris's performance as a base reference or some % increase over fury X but when one starts considering power consumption its quite hard to imagine VEGA reaching parity with a GP102. And this is without considering the fact that increasing clocks does not equal to a linear increase in performance.

Clearly going to 14nm would provide a healthy reduction in power consumption for VEGA but guess the question is just by how much reudction and what kind of available headroom is left for increasing the clocks. Maybe Paratus could do us a favor and work this one out (Maybe using Tonga -> Polaris as a comparison i dont know).

I look forward to VEGA especially since we'll be able to see the kind of improvements they have made with a direct comparison to Fury X. If it actually performs great then it could potentially be what i get next along with a Ryzen CPU.

 

[Gibbons]

What's the story with Vega 11?

There seems to be very little information on it. Die size? Launch quarter? Number of CUs?

 

[Paratus]

Performance never scales 100% with increased clocks or with shaders. For Vega to even match 1080 Ti the perf/sp has to increase significantly.

If we focus on the linear extrapolation of 2X rx570 at 1600mhz and1080P the extrapolated Vega is significantly faster. Add in performance enhancements to each SP from the transistor budget and it will be faster still.

As you state increasing clocks and CUs doesn't necessarily translate into a 1 to 1 increase. The question then becomes how much performance do we lop off our linear extrapolation to account for performance increases not being 1 to 1 with clock and SP increases?

Historically speaking we've seen a couple of different things:

• Tahiti to Polaris had a roughly 20% increase in CU to CU performance (I think compubase may have a page on this - I know someone linked it in this thread)
• The 5870 was 2X 4890: The 5870 ended up 20-50% faster depending on game
• Going all the way back to the X1800XT and XL vs X850XT, we saw anywhere from 5-50% increase in performance at the same clock and up to 90% when the 25% clock speed boost was included, (depending on game)

Vega is already rumored to have a 50% increase in clockspeeds over Fiji. Polaris already has a 10-20% increase in performance CU vs CU over Fiji. Vega is already rumored to have a 10% die increase over GP102. All of these point to competitive to better than competitive performance with the 1080Ti.

Again I maybe wrong. If I am I maybe trying my first NV desktop card later this year.

 

[OatisCampbell]

My Vega predictions, fresh from crystal ball in my a$$:

Little Vega: About 1080 in DX11, a little better in DX12, MSRP $499.99 to start.

Big Vega: Two of the above on a board, running slower, but still a good deal faster than latest Titan. MSRP $999.99 to start, undisputed single card champ unless NV joins the double card fray.

Why?

Because that all makes sense to me, and allows them to get two high end parts out of one chip. Its also a move they're fond of making.

I also think this would be a good move, supplies them with a halo part and a big market part.

 

[majord]

If you mean Dual GPU in the traditional sense, then sounds like a terrible Idea.

 

[dajeepster]

i don't care what they call them.. just gimme

 

[beginner99]

If the other rumors are right and Vega is:

• TMU 256
• SP 4096
• ROP 64

If...

All we know from the drivers is the number of compute cluster or what they are called at 64. And since in previous AMD design this always had 64 SPs we land at 4096. But it's a new uArch. And number of SP per cluster is one thing that changed from Kepler to maxwell for example. We also know that each core/SP was optimized (eg. the image with the large truck and smaller truck). Anyway each SP will be IMHO significantly smaller but also a bit slower than old SPs. Important part being more efficient (power and die size). So each cluster will have more slower SPs. I would guess 96 per cluster at 64 clusters that would be 6144 SP. That would also explain why 2xtimes polaris 10 doesn't add up with die size and transistor count. It then should not be 500mm2+. But if it actually has more SPs but smaller ones, that then totally makes sense.

 

[piesquared]

My Vega predictions, fresh from crystal ball in my a$$:

Little Vega: About 1080 in DX11, a little better in DX12, MSRP $499.99 to start.

Big Vega: Two of the above on a board, running slower, but still a good deal faster than latest Titan. MSRP $999.99 to start, undisputed single card champ unless NV joins the double card fray.

Why?

Because that all makes sense to me, and allows them to get two high end parts out of one chip. Its also a move they're fond of making.

I also think this would be a good move, supplies them with a halo part and a big market part.

I think a very real possibility is that dual Vega is 2 complexes connected together similar to Zepplin. We know Vega has Infinity Fabric incorporated and so it makes sense that they are meant to connect to each other. The 2048 bit bus makes sense also because 2 chips together would give the dual Vega a 4096 bit bus with 1 TB/s bandwidth and 32 GB HBM2 available as a high speed cache, or HBC, and the HBCC would manage access to more memory practically anywhere in the system to stream in as nesessary.
Hard to say what the product stack would look like, there would be a few possibilities. The tech is very cool, now to see how the design worked out.

 

[Glo.]

https://twitter.com/GChip/status/864334949263986688
https://twitter.com/GFXChipTweeter/status/864347035033522178

It appears that Radeon and Marketing team are preparing some ground. Its 16th May, after all.

 

[Krteq]

Hmm, JFK's new frontiers... Does anyone have a clue what exactly they meant by this?

 

[ryzenmaster]

If we focus on the linear extrapolation of 2X rx570 at 1600mhz and1080P the extrapolated Vega is significantly faster. Add in performance enhancements to each SP from the transistor budget and it will be faster still.

As you state increasing clocks and CUs doesn't necessarily translate into a 1 to 1 increase. The question then becomes how much performance do we lop off our linear extrapolation to account for performance increases not being 1 to 1 with clock and SP increases?

Historically speaking we've seen a couple of different things:

• Tahiti to Polaris had a roughly 20% increase in CU to CU performance (I think compubase may have a page on this - I know someone linked it in this thread)
• The 5870 was 2X 4890: The 5870 ended up 20-50% faster depending on game
• Going all the way back to the X1800XT and XL vs X850XT, we saw anywhere from 5-50% increase in performance at the same clock and up to 90% when the 25% clock speed boost was included, (depending on game)

Vega is already rumored to have a 50% increase in clockspeeds over Fiji. Polaris already has a 10-20% increase in performance CU vs CU over Fiji. Vega is already rumored to have a 10% die increase over GP102. All of these point to competitive to better than competitive performance with the 1080Ti.

Again I maybe wrong. If I am I maybe trying my first NV desktop card later this year.

If we look at these numbers and the supposed 12.5 TFLOPS; one could expect Vega to land close to the Titan xp. But there's still the chance that even if, given optimal scenario, vega could outperform 1080Ti -we might not see it in practice that often. Perhaps either due to limitations in DX11(majority of game titles) or games being optimized for Nvidia in specific.

For instance lets say the consumer Vega can vastly outperform Pascal in FP16. All Nvidia has to do is to make sure upcoming games don't heavily rely on FP16 and Vega runs underutilized. In fact they don't even have to explicitly take any action since anyone who wants their games to run well on Pascal will have to avoid FP16 anyway. Hopefully Volta will be an improvement in that regard.

All that aside I for one am willing to assume that Vega can theoretically outperform 1080Ti, but at the same time I wouldn't be surprised if what we see in practice is performance that is "only" around 1080 level more often than not. Personally I would be fine with that(if the price is right), but it would also mean they aren't ready to take on Volta and they would end up making fools of themselves with the whole "poor volta" thing..