Vega Rumors

[razor1]

True about Fiji, For both Fiji and Vega, I think it was custom made, there are two parts to the controller, one on the HBM memory and one on the GPU.

Hmm was thinking would Infinity fabric be causing the latency issues as well? Its possible if timings are good....

The r600 had similar problems too, well memory timings and channel amounts, nothing to do with the actual interconnect.

 

[Dayaks]

Seems Buildzoid is recommending best route if interested in Vega is watercooled and ignore air variants (unless putting your own waterblock on them), over last few weeks seeing some of the actual spec behaviour starting to feel the same myself.
But I think one area that may be of interest is whether AMD binned the Vega water cooled edition and so possibly a better route for some than air variant modded with say EK waterblock, fingers cross we get enough review information to weigh either option.
I guess the opinion may change when it comes to custom AIB air cooling, but I think the watercooled model will actually be a bit more efficient and with greater flexibility in the Boost mechanism behaviour due to thermals.



Cheers

I'd give it a few months to make sure water doesn't have wide spread issues like the Fury X did.

 

[CSI_PC]

I'd give it a few months to make sure water doesn't have wide spread issues like the Fury X did.

Hints of the Fury X cooling situation did come out in reviews where multiple sites were commenting about whine and the noise traits, but yeah like any new tech waiting for reviews with good analysis makes sense.
Cheers

 

[exdriver]

Well they would require an awkward ROP/MC configuration to end up with 64 ROPs/ 2048b bus divided between six GPCs so that wouldn't really make sense but assuming that is a non-issue and they would have gone for 96/3072 it would provide 50% higher geometry throughput and a paltry ~15% increase in shader throughput thus shifting the balance significantly towards geometry, something that appears to be a major bottleneck for Vega

I gather you're not the only one who feels this way, assuming that it would fit. Maybe next time.

Yeah so we have two problem areas, the geometry throughput and bandwidth issues. bandwidth issues are up in the air, possible theory is increased latency of the longer pipelines, which might be able to be fixed by updated drivers, if its something like that.

Updated drivers accounting for how much of a performance increase in your mind? At $399 if the 56 sees any improvement in its speculated performance I could see it being pretty popular at current market prices.

 

[razor1]

I gather you're not the only one who feels this way, assuming that it would fit. Maybe next time.



Updated drivers accounting for how much of a performance increase in your mind? At $399 if the 56 sees any improvement in its speculated performance I could see it being pretty popular at current market prices.

not much, the r600 issues were never solved, it really had some issues with its memory controller.

 

[CSI_PC]

Interesting to note buildzoid registered significantly lower than stock power draw while at 1800 mhz LN2, suggests leakage is a major issue for Vega, I wouldn't have been too surprised to find it drawing same power as stock but 100W less as he claimed is just nuts

Yeah I noticed that as well, but tbh Fury X if I remember also had a similar situation although this is rather high like you say.
Just one of the reasons watercooled version IMO will be the most efficient, along with potentially being binned; worth noting Fury X actually had better efficiency figures than the other Fiji models when measured accurately-isolated with a scope - context being when pushed to the same boost spec ceiling.
Cheers

 

[SighTurtle]

Gamer Nexus joins the mining rumor train, says they "independently" confirmed with their sources in industry that RX Vega will be around 70 for mining. (Hes not sure if 56 or 64). I've linked the video to timestamp.

Also, NKD, GN says 56 will be likely shipping reference later in August/Early September, seems like AMD has not been clear about if it comes with the launch date of 14th of August.

 

[Raendor]

Gamer Nexus joins the mining rumor train, says they "independently" confirmed with their sources in industry that RX Vega will be around 70 for mining. (Hes not sure if 56 or 64). I've linked the video to timestamp.

Also, NKD, GN says 56 will be likely shipping reference later in August/Early September, seems like AMD has not been clear about if it comes with the launch date of 14th of August.

Lol, while Nvidia currently has the best gaming cards (full range of Pascal is amazing from bottom to the top, nobody can deny it), all amd could is come up with the best mining card 1.5 years later as a response. That's sad.

 

[Rasterizer]

Gamer Nexus joins the mining rumor train, says they "independently" confirmed with their sources in industry that RX Vega will be around 70 for mining.

This would once again appear to suggest RTG has solved at least some of the memory bandwidth issues in Vega, since ETH hashrate is strongly affected by memory bandwidth.

 

[CSI_PC]

This would once again appear to suggest RTG has solved at least some of the memory bandwidth issues in Vega, since ETH hashrate is strongly affected by memory bandwidth.

Then it would mean relying upon compression, because the ideal maximum bandwidth with 2 stacks is the same as Fury X with HBM1 and 4 stacks, in reality the official spec for Vega is with lower BW than Fury X; 480GB/s vs 512GB/s.
There is a lot of unknowns around the 70MH/s (which is lower than the rumoured 75-100MH/s), this could be watercooled and pushed to the limits or one of the models without such extremes (still possibly the watercooled edition).
Cheers

 

[Presbytier]

Then it would mean relying upon compression, because the ideal maximum bandwidth with 2 stacks is the same as Fury X with HBM1 and 4 stacks, in reality the official spec for Vega is with lower BW than Fury X; 480GB/s vs 512GB/s.
There is a lot of unknowns around the 70MH/s (which is lower than the rumoured 75-100MH/s), this could be watercooled and pushed to the limits or one of the models without such extremes.
Cheers

Either way it means forgot owning the card for gaming lol

 

[CSI_PC]

Either way it means forgot owning the card for gaming lol

Not sure.
If it is the watercooled edition having to be pushed it could be debatable because 350W mode is required for sustained 1600Mhz by those that tested the watercooled Vega FE and measured it accurately, now the power demand for 70Mh/s could be even higher.
One of you guys who do mining can let us know the watermark of acceptable power demand for Mh/s.
Cheers

 

[razor1]

Either way it means forgot owning the card for gaming lol

Yep as I said if its over 60mhs, miners will eat them up lol.

 

[exdriver]

 

[Archaea]

That's the frontier edition which is not RX Vega.

 

[exdriver]

Yeah, it is (the FE). And he says he thinks it'd be faster once the drivers are updated... unlocked I think is the word he used. There's a lot of crosstalk, some say BS on 70ish some say 70-100, some say 40 is all we will see. He's outright saying you should buy them for mining though.

I'm not saying he's right or wrong. But this shows where we are now pretty clearly and what they're saying they'll do with a driver update, I don't know. Maybe someone with a few hundred cards can chime in.

 

[Dayaks]

Mining needs to die off again. It's the epitome of humans being stupid. What a waste.

I don't really care personally if you do it... but it's just saddens me for some reason.

 

[JustReason]

Granted I am not stating this as fact but wasn't the FE ~380ish for memory BW and the RX ~480ish? So I don't think you can use the FE to compare, especially for mining where memory is a huge factor. It was speculated when the FE came out that the memory BW was lower than the FuryX because of the 8hi and if the RX was 8Gb it would likely be more akin to the Fury.

 

[CSI_PC]

Yep as I said if its over 60mhs, miners will eat them up lol.

What power demand per GPU would be acceptable for 60Mh/s ?
Just curious.
Cheers

 

[Dan]

Im tempted to buy one to swap out my 1080 for because of freesync, But i dont know if freesync alone would be worth it especially at equal or similar performance. I dont know if adaptive Vsync is that much worse then Freesync.

Also I might need to swap out my 500w PSU because Vega sucks that much more juice X_X

 

[exdriver]

What power demand per GPU would be acceptable for 60Mh/s ?
Just curious.
Cheers

Play with the values here

https://www.cryptocompare.com/mining/calculator/eth?HashingPower=60&HashingUnit=MH/s&PowerConsumption=300&CostPerkWh=0.12

 

[Archaea]

Just watched the GN video linked above.

T

What power demand per GPU would be acceptable for 60Mh/s ?
Just curious.
Cheers

~ <= 300 or 350 watts --- unless you are counting on the concept you can put more power in a single motherboard with these beasts (than you can with something lesser) --- which does count for something.
You can see for yourself. Look at the daggerhashimoto score on the card of your choice with the drop down. It's a good ballpark figure.
https://www.nicehash.com/?p=calc

 

[Dan]

Can someone rich on Hardforum buy these in fking bulk and sell them to us at a slight profit so we all dont need to fight the miners X_X

 

[Archaea]

Im tempted to buy one to swap out my 1080 for because of freesync, But i dont know if freesync alone would be worth it especially at equal or similar performance. I dont know if adaptive Vsync is that much worse then Freesync.

Also I might need to swap out my 500w PSU because Vega sucks that much more juice X_X

I'll tell you clearly - it is. Freesync is very nice. I'm trading out my 1080TI to go back to Vega because I miss the smoothness my Fury X offered with Freesync as compared to locked 60Hz with the 1080TI. (1440p HP Omen 32)

 

[Dan]

I'll tell you clearly - it is. Freesync is very nice. I'm trading out my 1080TI to go back to Vega because I miss the smoothness my Fury X offered with Freesync as compared to locked 60Hz with the 1080TI. (1440p HP Omen 32)

I bought the 34" 2560x1080 LG 75hz freesync. I forget how the smothness was with my 390. Maybe once i get my hands on AMD again ill realize what i lost. Right now im just enjoying the raw power the 1080 was over the 390.

 

[Archaea]

I bought the 34" 2560x1080 LG 75hz freesync. I forget how the smothness was with my 390. Maybe once i get my hands on AMD again ill realize what i lost. Right now im just enjoying the raw power the 1080 was over the 390.

Did you do freesync with the 390? Maybe not everyone notices it? But 60hz with vsync vs. 75hz with freesync is a significant difference of smoothness to my perception. So much so that I preferred my former Fury X in Crossfire, to the 1080TI in SLI. And If I hadn't already sold the pair of Fury X to fund the 1080TI - I would switch back in a heartbeat for gaming.

 

[exdriver]

Granted I am not stating this as fact but wasn't the FE ~380ish for memory BW and the RX ~480ish? So I don't think you can use the FE to compare, especially for mining where memory is a huge factor. It was speculated when the FE came out that the memory BW was lower than the FuryX because of the 8hi and if the RX was 8Gb it would likely be more akin to the Fury.

It's an indirect comparison, saying FE has nothing to do with RX isn't quite true either. But yeah, better drivers and better memory speed might be something higher than 28~40?

In any case I think miners are going to be buying.

I'll tell you clearly - it is. Freesync is very nice. I'm trading out my 1080TI to go back to Vega because I miss the smoothness my Fury X offered with Freesync as compared to locked 60Hz with the 1080TI. (1440p HP Omen 32)

I think you do but I'll ask, do you like the Omen? It caught my eye tbh.

 

[razor1]

What power demand per GPU would be acceptable for 60Mh/s ?
Just curious.
Cheers

300-350 watts.

 

[Ieldra]

Granted I am not stating this as fact but wasn't the FE ~380ish for memory BW and the RX ~480ish? So I don't think you can use the FE to compare, especially for mining where memory is a huge factor. It was speculated when the FE came out that the memory BW was lower than the FuryX because of the 8hi and if the RX was 8Gb it would likely be more akin to the Fury.

This is incorrect FE and RX have virtually same rated bandwidth, FE is 484 RX isn't even 5% higher

 

[JustReason]

This is incorrect FE and RX have virtually same rated bandwidth, FE is 484 RX isn't even 5% higher

Then what in the hell was the 380ish BW shown in the reviews? It was stark against the FuryX and garnered huge discussions about how they regressed.

 

[CSI_PC]

This is incorrect FE and RX have virtually same rated bandwidth, FE is 484 RX isn't even 5% higher

Although it is still not clear just how much impact the 8-Hi is having on the 'official' spec and whether this also applies to the 4-Hi stacks as well, in fact I am not sure where the HBM2 memory is even coming from as back in July only Samsung has officially mentioned they are capable of manufacturing 8-Hi.

Another factor we do not know if the issue comes back to some quirk of 2-stack HBM2 implementation that would affect all, worth noting Nvidia came across GDDR5X quirks in the real world.
Or even if it is a side-effect of the HBCC<--->Memory Controller with some quirks in real world but then again that would apply to all.

I guess more waiting for review benchmarks
Cheers

 

[razor1]

Then what in the hell was the 380ish BW shown in the reviews? It was stark against the FuryX and garnered huge discussions about how they regressed.

Something is stopping the FE from using all its bandwidth effectively, we don't know what it is. It could be drivers or something with the architecture, like what we saw with the r600 the ring bus and number of channels the memory had really hurt the r600 when it came to bandwidth.

 

[exdriver]

This is incorrect FE and RX have virtually same rated bandwidth, FE is 484 RX isn't even 5% higher

Thanks. If RX can do 70 on the same memory with new drivers what does that say about FE? At least 70 or more?

Although it is still not clear just how much impact the 8-Hi is having on the 'official' spec and whether this also applies to the 4-Hi stacks as well, in fact I am not sure where the HBM2 memory is even coming from as back in July only Samsung has officially mentioned they are capable of manufacturing 8-Hi.

Another factor we do not know if the issue comes back to some quirk of 2-stack implementation that would affect all, worth noting Nvidia came across GDDR5X quirks in the real world.

I guess more waiting for review benchmarks
Cheers

Well yeah, but waiting for reviews isn't the point of a rumors thread

Something is stopping the FE from using all its bandwidth effectively, we don't know what it is.

I'm agreeing with you, but what are you basing this on?

 

[tungt88]

To be honest, if Vega is indeed grabbed up by the mining community, that would make me very happy with my GTX 1080 Ti purchase from not too long ago (and make me look towards a monitor upgrade around Black Friday or so). A nice big monitor (as close to my current 42" as possible) with 4K@60hz + G-Sync would be a nice bonus indeed.

I still hope that Vega is viable enough for gaming (seems to be ok, like a slightly lesser take on the HD 3850/3870, so to speak); at least for AMD to refine Navi (drivers, etc).

The G-Sync vs FreeSync, as per Kyle, was pretty nice (but, really, just like a one potato chip sample from the bag). Samples are always nice, though

 

[Ieldra]

Then what in the hell was the 380ish BW shown in the reviews? It was stark against the FuryX and garnered huge discussions about how they regressed.

Thats the effective memory bandwidth

 

[razor1]

I'm agreeing with you, but what are you basing this on?

B3D synthetic tests, that test really shouldn't affect the bandwidth like that unless there is some underlying problem.

 

[CSI_PC]

B3D synthetic tests, that test really shouldn't affect the bandwidth like that unless there is some underlying problem.

Yeah and could be anything from driver to hardware such; cache<----->HBCC<---->Memory controller or another hardware aspect.
Worth noting it is not just the effective texture BW that shows anomalies (single black is near same as Fury X) for now but also as shown earlier fill rates that comes back to GPU cache in some ways (possibly the relationship between the cache-HBCC-Memory Controller as I mention above) *shrug*.
For AMD's sake fingers crossed it is drivers.
But this is not impacting all benchmarks, considering just how much faster FE is in some of the updated SPEC performances relative to Fiji, nearly every one of those tests Vega is over 50% faster and in some multiples of X faster than Fury X.

On the plus side with AMD harping on about Vega compared to Titan xP in professional benchmark made Nvidia to finally make the Titan cards Prosumer and enabled to use the SDK related functions now hooray
About freaking time considering it was a Prosumer card in the past just without the same support/guarantees, so got to thank AMD for keeping on about the Titan xP rather than other quadro cards.
Although that now makes the Titan xP very price/performance competitive for such professional workloads.
Cheers

 

[Rasterizer]

Then it would mean relying upon compression, because the ideal maximum bandwidth with 2 stacks is the same as Fury X with HBM1 and 4 stacks, in reality the official spec for Vega is with lower BW than Fury X; 480GB/s vs 512GB/s.

Then what in the hell was the 380ish BW shown in the reviews? It was stark against the FuryX and garnered huge discussions about how they regressed.

I'm agreeing with you, but what are you basing this on?

I'm going to go ahead and re-post the three B3D Suite tests that show bandwidth issues with Vega:

 

[exdriver]

Is this the 1.6 Gbps HBM2?

 

[JustReason]

Is this the 1.6 Gbps HBM2?

That is the confusing part. I thought reviews mentioned it as lower clocked but toms shows 1.8Ghz on the FE but a huge discussion ensued upon its release that based upon the chance that the FE was indeed not clocked the same as the RX speculating the RX would be for gaming and the FE lower for greater stability.