Vega/Navi Rumors (Updated)

[T1beriu]

Nope.

If you're not bringing anything to the conversation it's better to abstain from posting.

 

[crisium]

So now is where I have to waste a post asking you for the courtesy of an explanation.

Nope.

Can you explain how 12.5 TFLOPs and 1525MHz+ do not synchronize?

 

[Jackie60]

So is it confirmed no new consumer Vega until sometime possibly soon?

 

[Malogeek]

So is it confirmed no new consumer Vega until sometime possibly soon?

Confirmed on AMD Blog that consumer Vega will be coming after Frontier Edition, which means after June.

 

[Glo.]

So now is where I have to waste a post asking you for the courtesy of an explanation.



Can you explain how 12.5 TFLOPs and 1525MHz+ do not synchronize?

Where do you see that I contradict 12.5 TFLOPs/1525 MHz?

There is nothing dead in 1.2 GHz RX Vega GPU rumors.

 

[crisium]

Where do you see that I contradict 12.5 TFLOPs/1525 MHz?

There is nothing dead in 1.2 GHz RX Vega GPU rumors.

It seems like you're fishing for "gotchyas" with your brief responses devoid of elaboration.

Did you think I was setting a floor on all Vega clockrates to have this minimum? I was not, to be clear.

I'm speaking toward the full fat flagship GPU. The recent leaks were of 64 CU Vega at 1200mhz and 1.4-1.6Gb/s HBM2. That's virtually debunked now from being a released product, even if the Pro and Consumer cards will surely have different clocks. That leak was probably like the 800MHz and 7 Gb/s Polaris 10 leaks from last year.

BTW, there is no precedent for modern (since GCN 1 surely) AMD having that much clock discrepancy between dGPUs on the same architecture - not even 7950 vs 7970 which has never come close to being repeated again. Perhaps a mobile chip could be clocked that low. edit: I could be persuaded to believe a new Nano could be there, but am doubtful.

 

[Malogeek]

And workstation cards are never clocked higher than consumer cards, especially in the range of 300Mhz.

 

[Glo.]

It seems like you're fishing for "gotchyas" with your brief responses devoid of elaboration.

Did you think I was setting a floor on all Vega clockrates to have this minimum? I was not, to be clear.

I'm speaking toward the full fat flagship GPU. The recent leaks were of 64 CU Vega at 1200mhz and 1.4-1.6Gb/s HBM2. That's virtually debunked now from being a released product, even if the Pro and Consumer cards will surely have different clocks. That leak was probably like the 800MHz and 7 Gb/s Polaris 10 leaks from last year.

BTW, there is no precedent for modern (since GCN 1 surely) AMD having that much clock discrepancy between dGPUs on the same architecture - not even 7950 vs 7970 which has never come close to being repeated again. Perhaps a mobile chip could be clocked that low. edit: I could be persuaded to believe a new Nano could be there, but am doubtful.

IMO, the 1.2 GHz is the "Nano" Version of RX Vega.

 

[IEC]

IMO, the 1.2 GHz is the "Nano" Version of RX Vega.

They would be silly not to take advantage of HBM2's space savings and make a capable mini-ITX/HTPC gaming card. At significantly lower clockspeeds it's likely the efficiency will be good as well, making it ideal for SFF builds.

 

[Actaeon]

Anyone else think that AMD's silence regarding performance suggests that Vega won't be faster than GP102? If AMD knows that Vega is be faster I think they would have said something to keep people interested and waiting. With zero products in the high end they aren't risking cannibalize sales by leaking info early, only keeping people away from Nvidia. They used that strategy successfully for Ryzen and kept people from buying Intel until Ryzen was out.

On the other hand if the card was slower, it would be better to not say anything to keep people wondering and not buying. With what they've shared that's what it feels like what they're doing. They're either waiting for the May 31st event or its slower.

I hope they say something on May 31st or I hope that I'm wrong. AMD's done a wonderful job in keeping Intel honest with Ryzen and Vega needs to be successful to do the same to Nvidia. Nvidia has a monopoly on the high end GPU market for way too long now and pricing has gotten pretty ridiculous.

 

[crisium]

They would be silly not to take advantage of HBM2's space savings and make a capable mini-ITX/HTPC gaming card. At significantly lower clockspeeds it's likely the efficiency will be good as well, making it ideal for SFF builds.

Efficiency might be why it would be "downclocked" that low. That might be the equivalent of Fiji Nano running just north of 800Mhz, lower than it normally clocks at (upper 800s to lower 900s). But if Vega uses more juice then it could need that to get around 180W or wherever they target. And we do not know the sweet spot. I could believe it. They have to make sure it stays comfortably ahead of the 1070 ITX though, since Nvidia has a much easier time selling a 314mm^2 GDDR5 chip that cheap.

Didn't that leak come with 1.4GB/s HBM 2 though (around 350 GB/s)? Do we believe that?

I'd like some more thoughts on the topic if a certain user would ever like to explain himself more than 2 words. Discussion can be fun. =P

 

[Valantar]

Reason is extremely simple. In emergin market like Deep Learning there is both very high margin, and very high volume. Focusing your efforts on this market with innovation is very good business decision.

Well, obviously. But that doesn't really affect the potential validity of my hypothesis now, does it?

For that to happen there must be some indication that Vega has dedicated FP16 units, and so far there is nothing that hints at such a possibility.

No, there doesn't. All that's necessary for that to make sense is for packed math to require a noticeable amount of die space. It's obvious we're not talking about dedicated FP16 units (then it wouldn't be packed math at all...), but it makes no sense that adding that capability to the SPs would have no impact on die size. If it's "free", why not add packed math capability to every single GPU out there? And as such, unless packed math capabilities are entirely engrained into the SPs in the NCU, they must then be "removable" for chip designs that don't need the capability. That's all I'm stipulating. Is it likely? I have no idea. But I find it to be an entertaining idea.

 

[Head1985]

Confirmed on AMD Blog that consumer Vega will be coming after Frontier Edition, which means after June.

Jesus more than year later than pascal and probably only with crap reference cooler and another 2 months for good aftermarket cards...

 

[DeathReborn]

Where do you see that I contradict 12.5 TFLOPs/1525 MHz?

There is nothing dead in 1.2 GHz RX Vega GPU rumors.

It could be 2x 8pin for 1.5-1.6GHz and 1x8pin + 1x 6pin for 1.2GHz I suppose.

 

[IllogicalGlory]

And workstation cards are never clocked higher than consumer cards, especially in the range of 300Mhz.

That's not so much the case with Polaris professional cards. Two out of the three are clocked at or above 1.2GHz, very close to the desktop gaming parts.

 

[Kenmitch]

Why hasn't the original post been updated to reflect the recent Vega news? Might not be gaming related card news, but then again neither is the Volta threads recently updated news. Not news worthy?

 

[Mopetar]

I agree.

And if they can manage 1.875ghz HBM2 in an 8hi, how about in 4hi? Could it be easier to hit their target 2.00ghz HBM2 with only 4 stacks?

I was thinking along similar lines as well. I'm not sure how much it matters though.

IMO, the 1.2 GHz is the "Nano" Version of RX Vega.

It seems unlikely that they would reduce the clocks that much, unless the FE chip is pretty much maxed out and draws around 300W, which seems unlikely if it's designed to be a data center card. A ~33% difference in clock speed is quite large though and the Fury Nano was not that much less than the Fury X even.

I could see it making sense if they had some chips that had disabled CUs or some chips that couldn't clock very high so AMD creates a bin for those and targets a 150W TDP or something like that and slots the card up against the 1070. Maybe it's just a minimal bin that sits right above the scrap heap to let AMD recoup some cost, but I can't see it being a money maker for them.

 

[Erithan13]

Something I just seen mentioned on Reddit:

The card has 16GB but only 480GB/s bandwidth. That means either 4 stacks of 4GB, which would mean significantly higher bandwidth. Or they're using two 8GB stacks, which is also unlikely. Neither Samsung nor SK Hynix has shown any working.

Even Ryan Smith at Anandtech doesn't know how AMD might have done that.

Calling it here folks, Vega 16GB is actually 8GB HBM2, and 8GB GDDR5 using the HBCC to use it all.

They don't list the card as having 16GB HBM2, they state and list 16GB HBC - High Bandwidth Cache.

Any thoughts/informed guesses/wild speculation?

 

[Actaeon]

Jesus more than year later than pascal and probably only with crap reference cooler and another 2 months for good aftermarket cards...

I have high hopes for the liquid cooled models but the reference blower will certainly be garbage.

 

[crisium]

Something I just seen mentioned on Reddit:



Any thoughts/informed guesses/wild speculation?

Complete nonsense.

 

[Headfoot]

Complete nonsense.

Agreed.

No way they built both memory interfaces into both dies. Maybe small vega has GDDR5 and big vega has HBM2, but very very doubtful both have both.

Way more likely they have 16GB of actual HBM2 on the pro card, and 8GB of actual HBM2 on the consumer card (maybe 8/16 models both). My bet is HBCC is specifically for interfacing with onboard SSD/NVRAM (and also accessing other memory without a CPU in between)

 

[amenx]

Confirmed on AMD Blog that consumer Vega will be coming after Frontier Edition, which means after June.

Worrying sign that they cant even offer a rough release date anymore.

 

[Snarf Snarf]

Something I just seen mentioned on Reddit:



Any thoughts/informed guesses/wild speculation?

I don't think this idea is as hard to believe as people think. During the presentation they already mentioned the NVMe + HBM2 Cache is going to be released in the pro segment. Adding a GDDR5 controller would definitely cost die space that basically cancels out the gains that HBM gives you. Something tells me that 8hi is what we're looking at, SK Hynix just hasn't announced it because AMD bought all of their stock and will continue to for a while. The HBM2 shortage rumor might actually be about 8hi stacks and not HBM2 in general. Given the roadmaps Hynix put out before 8hi would have barely been completed and in extremely low yields/volume right now. 25k units for the Frontier Edition limited launch sounds really normal for a professional card.

 

[Bacon1]

Confirmed on AMD Blog that consumer Vega will be coming after Frontier Edition, which means after June.

Source for that blog post? It would make sense they want to wait for Computex to release info on the gaming cards.

 

[crisium]

16GB HBM2. I don't see the room for misinterpretation.