Vega/Navi Rumors (Updated)

[.vodka]

^ Which is what I meant by "people of that caliber"

Yeah.. but I'm quite curious why we haven't seen something from Johan "repi" Andersson (Technical director & "pixel pusher" of Frostbite engine)

Yeah... Maybe he'll comment at some point.

 

[Maverick177]

Question: What would constitute Vega as a failure or a winner? Please taken into account release date, performance, price, adoption etc..

 

[tviceman]

Question: What would constitute Vega as a failure or a winner? Please taken into account release date, performance, price, adoption etc..

I think there is so much stacked against it. At this point it's all but confirmed that it won't match 1080 TI in performance, it won't have Pascal-like efficiency, and again, it's a year late. The chip is supposedly bigger than GP102, has the more expensive HBM memory (how much 8 gigs of HBM2 costs vs. 11 gigs GDDR5X I do not know), and will likely need a more complicated power delivery system. Coupled with the fact it's more than a year late in the high end space, and Volta consumer parts looming ~6-7 months from now, Vega may be one of the shortest lived high end GPU's ever. Of course it's always possible AMD is sandbagging and Vega is going to beat GP102. Almost anything is possible.

Here is my question: If it's $599 and 10% slower than 1080 TI but with little OC headroom and a high power draw from the get go, will those who held out be happy?

 

[Armsdealer]

Here is my question: If it's $599 and 10% slower than 1080 TI ....

If you're right on your prognosis of performance, and I think you are, I think you're going to be surprised at the price: Vega is likely to be a 500 USD (maybe 450) proposition despite the hbm for the full die. It's likely a gp 104 competitor and will be priced accordingly.

 

[Paratus]

I think there is so much stacked against it. At this point it's all but confirmed that it won't match 1080 TI in performance, it won't have Pascal-like efficiency, and again, it's a year late. The chip is supposedly bigger than GP102, has the more expensive HBM memory (how much 8 gigs of HBM2 costs vs. 11 gigs GDDR5X I do not know), and will likely need a more complicated power delivery system. Coupled with the fact it's more than a year late in the high end space, and Volta consumer parts looming ~6-7 months from now, Vega may be one of the shortest lived high end GPU's ever. Of course it's always possible AMD is sandbagging and Vega is going to beat GP102. Almost anything is possible.

Here is my question: If it's $599 and 10% slower than 1080 TI but with little OC headroom and a high power draw from the get go, will those who held out be happy?

The reported clock speeds and die size mean it's almost certain to be faster than Pascal and probably competitive with Volta. Power will be higher however.

 

[tviceman]

If you're right on your prognosis of performance, and I think you are, I think you're going to be surprised at the price: Vega is likely to be a 500 USD (maybe 450) proposition despite the hbm for the full die. It's likely a gp 104 competitor and will be priced accordingly.

I will be shocked if they're able to sell it at $450-500 considering RX 580 is less than half the size, about half the performance, less complicated power delivery, way less expensive vram, and the Vega FE being $999. Typically high end cards do not have the same perf/$ as the sweet-spot priced cards.

 

[tviceman]

The reported clock speeds and die size mean it's almost certain to be faster than Pascal and probably competitive with Volta. Power will be higher however.

How I hope you're right. But given all the rumors, indications, and lack of any benchmarks from AMD in games, I doubt it. Also, given that RX 580 has nearly 50% more TFLOPS of GTX 1060, yet performs within 5% on average and consumes way more power, I don't think Vega will live up to your current expectations in performance.

 

[turtile]

How I hope you're right. But given all the rumors, indications, and lack of any benchmarks from AMD in games, I doubt it. Also, given that RX 580 has basically double the TFLOPS of GTX 1060, yet performs within 5% on average and consumes way more power, I don't think Vega will live up to your current expectations in performance.

But Vega has twice the 16FP per 32FP than Polaris.

 

[exquisitechar]

How I hope you're right. But given all the rumors, indications, and lack of any benchmarks from AMD in games, I doubt it. Also, given that RX 580 has basically double the TFLOPS of GTX 1060, yet performs within 5% on average and consumes way more power, I don't think Vega will live up to your current expectations in performance.

The RX 580 does not have double the TFLOPS of the 1060. Nowhere near double, really.

 

[tviceman]

The RX 580 does not have double the TFLOPS of the 1060. Nowhere near double, really.

You're absolutely right, it's closer to 50% more TFLOPS. That was a brain fart. To be precise, the RX 580 has 42% more TFLOPS than the GTX 1060 (6.17 vs. 4.35). I edited my original post to fix that error.

 

[senseamp]

I suspect they will price it at $700 regardless of the performance, even if it means almost no demand. Because they don't care if they sell any below that. Remember, their opportunity cost to making a Vega is using the allocated wafers to make 2-3 Ryzen cores, which will sell for a lot more without having to spend money on HBM, board, power delivery, oem profit, etc. The only reason to sell for a very low gross margin to drive volume would be to preserve mind share among diehard fans, developers, and partners. Basically a very expensive fig leaf. Or if Ryzen demand is not strong enough to fill their wafer allocation. Although then, they would still be better off building RX580s for miners.

 

[exquisitechar]

You're absolutely right, it's closer to 50% more TFLOPS. That was a brain fart. To be precise, the RX 580 has 42% more TFLOPS than the GTX 1060 (6.17 vs. 4.35). I edited my original post to fix that error.

No worries.

I suspect they will price it at $700 regardless of the performance, even if it means almost no demand. Because they don't care if they sell any below that. Remember, their opportunity cost to making a Vega is using the allocated wafers to make 2-3 Ryzen cores, which will sell for a lot more without having to spend money on HBM, board, power delivery, oem profit, etc. The only reason to sell for a very low gross margin to drive volume would be to preserve mind share among diehard fans, developers, and partners. Basically a very expensive fig leaf. Or if Ryzen demand is not strong enough to fill their wafer allocation. Although then, they would still be better off building RX580s for miners.

They've promised good price/performance, it has even been compared to the Ryzen in that regard. If Vega's performance is lacking, they know they can't be late, underperforming, and expensive at the same time, precisely because of mindshare.

Info from Carsten Spille (PCGH.de Editor in Chief)Comparison to R300, really? In reaction to a post from another user asking if he is kidding he wrote:

Whoa, that's encouraging. Is it me, or is the follow up reply kind of cryptic, though?

 

[Malogeek]

But Vega has twice the 16FP per 32FP than Polaris.

That doesn't really help for games though. Some future games? Quite possibly yes as some workloads are being moved to FP16 instead.

 

[beginner99]

Here is my question: If it's $599 and 10% slower than 1080 TI but with little OC headroom and a high power draw from the get go, will those who held out be happy?

It's almost certainly not slower than a 1080ti. Simply because Fiji with 1600mhz clocks will almost reach that level of performance while have some serious bottlenecks. If you remove the bottlenecks and add in the GCN to NCU improvements, then Vega should clearly be able to match it. However power use will be higher.

Only thing that hints against this is that the FE has 1600 mhz listed as boost and 13xx mhz as base. So there seem to be some serious power issues. If it could run 1600mhz 24/7 it would hardly be listed as max boost clock. And that is probably where the water cooled version comes in: constant 1600 mhz.

I'm kind of with senseamp. The will price it according to it's performance ignoring high power use. If it's 10% slower than 1080ti it will be priced 10% cheaper. As long as AMD can will the fabs with ryzen, they have little incentive to sell Vega for cheap especially also due to hbm2 shortage and miners (meaning they should take the profit from miners not the retailers). So if it is 10% slower it will cost $649 and if not closer to $749.

 

[Armsdealer]

I will be shocked if they're able to sell it at $450-500 considering RX 580 is less than half the size, about half the performance, less complicated power delivery, way less expensive vram, and the Vega FE being $999. Typically high end cards do not have the same perf/$ as the sweet-spot priced cards.

One thing I've noticed is often on enthusiast forums people forget the biggest cost in projects like this is the sunk cost on the people that engineered / marketed / wrote software for the architecture. The actual cost of manufacturing the card isn't so high that they "aren't able" to sell the card at that price. (Gross profit / unit) * (number sold) has to be greater than development / human costs in order for the project as a whole to be profitable. Even at 400 USD and less I'm sure it's economically rational for amd to sell Vega. They *have* to sell units and, notwithstanding the unusual mining environment, based on the data we're seeing, they will really only sell this product in a meaningful fashion at 500 USD or less given gtx 1080 is (in theory) a 500 USD card now (again notwithstanding the mining environment).

Of course the data could be wrong and the product is a gp102 competitor in which case they'll charge more because they can.

 

[Borealis7]

Tha Vega logo...

i was putting on some diaper paste for my baby and there it was...

UPDATE: they must've changed their logo, their site is showing something else.

 

[Ancalagon44]

My prediction:
Full fat RX Vega with 8GB HBM2 - $650. Will be faster than 1080 but not as fast as 1080Ti, although in some games it will come close.
Cut down RX Vega with 8GB HBM2 - $550.

 

[3DVagabond]

I'm sure given ample cooling it will be 100%.

And with TPU's "hand over the blower" technique?

 

[3DVagabond]

Doubt that. Haven't they been shipping to a cloud gaming provider already back in march? So there were already drivers out then. What can they really do till August? Not like drivers will magically make it 50% faster. At this point anything more than 10% would be weird.

No samples probably means paper launch, so little availability that they could not even sample reviewers.

When was the last time pro cards were sent out to mainstream hardware sites?

 

[Veradun]

My prediction:
Full fat RX Vega with 8GB HBM2 - $650. Will be faster than 1080 but not as fast as 1080Ti, although in some games it will come close.
Cut down RX Vega with 8GB HBM2 - $550.

Uhm. The Pro version has an MSRP of 999$ with 16GB, I highly doubt your prediction here.

Either it has 16GB or it has to have a better price than that.

When was the last time pro cards were sent out to mainstream hardware sites?

I've rarely seen Quadro and Radeon WX reviews at all.

Add in how many Titan Xp review you've seen and you have your answer :>

 

[Ancalagon44]

Uhm. The Pro version has an MSRP of 999$ with 16GB, I highly doubt your prediction here.

Either it has 16GB or it has to have a better price than that.

Why not?

We know that HBM2 is relatively rare. My guess is that AMD will say, if you want 16GB of memory, buy the pro version. If you just want to play games, a $650 8GB HBM2 unit will work just fine.

Let's not forget that AMD calls HBM2 a high bandwidth cache now - my guess is their line will be that 8GB of an HBM2 HBC will be enough memory and bandwidth for 4K gaming for some time.

 

[Veradun]

Why not?

We know that HBM2 is relatively rare. My guess is that AMD will say, if you want 16GB of memory, buy the pro version. If you just want to play games, a $650 8GB HBM2 unit will work just fine.

Let's not forget that AMD calls HBM2 a high bandwidth cache now - my guess is their line will be that 8GB of an HBM2 HBC will be enough memory and bandwidth for 4K gaming for some time.

Point is Polaris based WX cards have the same amount of memory of the respective RX cards but they are priced between 2x and 3x MSRP.

I agree 8GB is more than enough for 4K, anyway.

 

[Glo.]

https://videocardz.com/70526/amd-preparing-radeon-pro-wx-9100-with-vega-gpu

 

[Ancalagon44]

Point is Polaris based WX cards have the same amount of memory of the respective RX cards but they are priced between 2x and 3x MSRP.

I agree 8GB is more than enough for 4K, anyway.

Yes, so? Vega Frontier Edition is not an RX card, it is not a gamer card. It is a prosumer card, and has a price tag to match.

Regular RX Vega for gamers will cost a lot less than $999 but probably will not include 16GB of HBM2 as AMD feels it is unnecessary for almost all gamers.

There will still be WX Vega cards above the Vega Frontier Edition cards that will be even more expensive.

 

[Ancalagon44]

Point is Polaris based WX cards have the same amount of memory of the respective RX cards but they are priced between 2x and 3x MSRP.

I agree 8GB is more than enough for 4K, anyway.

Yes, so? Vega Frontier Edition is not an RX card, it is not a gamer card. It is a prosumer card, and has a price tag to match.

Regular RX Vega for gamers will cost a lot less than $999 but probably will not include 16GB of HBM2 as AMD feels it is unnecessary for almost all gamers.

There will still be WX Vega cards above the Vega Frontier Edition cards that will be even more expensive.