Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Traditionally new generation gpu matches performance of previous gen gpu that was one tier higher
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
RUMOR: Radeon 480 to be priced to replace 380, similar output as 390.
- Thread starter Zion Halcyon
- Start date
Traditionally new generation gpu matches performance of previous gen gpu that was one tier higher
TaintedSquirrel
[H]F Junkie
- Joined
- Aug 5, 2013
- Messages
- 12,763
Yeah the 470/480 seem like they will perform exactly as you would expect, unfortunately AMD is missing the other half of the line-up...
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Yeah, I'm genuinely disappointed
Yep. AMD doesn't have a recent equivalent, but the GTX970 matched the GTX780ti at release and today with driver optimizations I have no doubt pulls substantially ahead. Why wouldn't the fully-enabled Polaris 10 do the same to the Fury X? We simply don't know either way, but it's at least plausible.
Hopefully the 480 is a binned Polaris 10. Guess we'll see soon enough.
Hopefully the 480 is a binned Polaris 10. Guess we'll see soon enough.
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Maxwell is an exception, they gutted Kepler's dp to hit the performance targets, I expect theyll do the same for Volta
AMD kind of did that with Fiji
Sweet! SLI here I come!
No, the "barely makes a difference" GPU updates we've seen recently are due to being stuck on the same process node for ages. More often than not a new GPU generation would decimate the previous one.
It looks like AMD is going to be 8 months behind Nvidia in performance, that's not going to be good. We all know what Nvidia does when running unopposed. AMD can't afford to cut the Fury to $300.
D
Deleted member 93354
Guest
wccftech said:
<facepalm>
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
NVIDIA GeForce GTX 680 Review: Retaking The Performance Crown
Geforce GTX 680 review
You're right in thinking this FF generation will decimate the previous one, just not this early in it's lifecycle, NV still has density to spare looking at GP100, and they have a bunch of fp64 units they can move out.
The GPU that is to Pascal as Maxwell is to Kepler will be the one that truly decimates
Basically it looks like we're looking at a repeat of the Radeon 4870 and Radeon 4850.
They were both marketed as mid-level cards, but when they came to market they proved to be much more potent. nVidia was out first with the GTX260 and GTX280 at $450 and $650 respectively, both poised as the best of the best, and fastest cards avaialble. AMD responded with "mid-tier" cards with the 4870 and 4850 at $300 and $180 respectively. The big win was the fact that the 4870 was a bit quicker than the GTX260 and in certain games(GRID) even outperformed the GTX280. It was so bad that nVidia had to send refund checks to early adopters of their latest cards. Basically AMD was providing about 85% of nVidia's top end cards for half the price.
nVidia was on top of their game after their 9xxx series cards, AMD was hurting after their overpriced, underperforming series GPU in the 2900XT, which was cleaned up a bit and turned into a 3-series 3870 and 3850 which sold respectively and because they were realistically priced proved to be a decent alternative. Very similar to today with nVidia being dominant with their 9xx series cards, and AMD loosing market share during the 2xx series cards and then cleaning them up a bit and making the 3xx series cards which today are recommended.
So lets take that situation and apply it today. I believe that nVidia's Pascal will be out the gate first. They will have 3 performance levels.
-Geforce x70 - GTX980 performance, $400, 6GB GDDR5
-Geforce x75 - GTX980-Ti performance, $600 8GB GDDR5x
-Geforce x80 - Titan-X performance +10-20%, $800 8GB GDDR5x
So, at those prices and performance levels, one can argue they are relative bargains, compared to the current $500 GTX980, $650+ GTX980-Ti and $1000 Titan-X. We went through this before, when the GTX970 was cheaper than the GTX780 and was faster, and the GTX980 was also cheaper and faster than the GTX780-Ti.
Well, AMD will wait a couple weeks then release:
-R9-470 - R9-380 performance, $160, 4GB GDDR5
-R9-470x - R9-380x performance $220 - 4GB GDDR5
R9-480 - R9-290x/R9-390 performance $270 - 4GB GDDR5
R9-480 - R9-390x/Fury performance $320 - 8GB GDDR5x
R9-480x - Fury/Fury-X performance $450 - 8GB GDDR5x
This would make it a repeat of the 4870 v GTX280. With DX12 already being stronger on AMD's cards, and the fact that Multi-GPU support will get better. It may make more sense to get a couple R9-480x cards over a single Geforce x80 for a significant performance at only spending an extra $100. Basically we're looking at AMD possibly again offering 85-105% of the performance of nVidia's top card for about half the price.
Now, nVidia can be smart, and see that history is about to repeat itself. I'd say knock the prices of the upcoming Geforce cards to $300(x70), $450(x75) and $650(x80) with the same performance levels, and it may throw a monkey wrench into AMD's plans. Then again, AMD may have wiggle room, with price so they may be cheaper. The problem becomes with stock of older cards, but I think the stock of R9-2xx and 3xx series is starting to dry, with the price increases I'm seeing on those cards. So that may not be an issue. It may sound like a pipe-dream but Fury-X and Titan-X performance might be at hand soon for less than half the price of those respective cards very soon.
Zion Halcyon
2[H]4U
- Joined
- Dec 28, 2007
- Messages
- 2,108
I think that is a very good assessment on where this could go based on history, and I haven't seen anything from NVidia that makes me think they are taking the AMD threat seriously.
The thing is, lets say history does repeat itself here. The main thing that would appear to be the difference between then and now, hopefully, is the Leadership on the AMD side of things. They have to capitalize and keep the pressure on NVidia. They can't just chalk something like that up to a win, have a party, and go back to business as usual.
At the very least, under the current RTG I have more faith that they have the leadership in place they lacked back then to actually take it to Nvidia. But only time will tell.
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Okay... But nothing so far points to Polaris amounting to anything more than ~390 performance, this thread exists because of this latest rumor outright stating that Polaris 10 will be a 390 level performer at a 380 price
Hmm I don't agree on the pricing structure at all Rauelius, don't expect AMD to undercut nV, shit they didn't do it with the Fury line and they made a modest increase with the r3xx line (you could get the chip cards from the the r2xx line for less). The only way they will price them lower than nV's if there is a significant performance difference per bracket. Yes AMD has changed their stance on being the cheaper alternative even though they didn't perform well on release of Fury line.
What made the 48xx line great was performance for $, but AMD also had the advantage of smaller die sizes with a performance advantage of no DP units and aditionally a VLIW architecture that was well suited for DX10 games, and because of the VLIW architecture they were able to have higher transistor densities to help reduce die size even more.
Added to all this the cost per wafer is high with 14/16nm. Unlike previously with fab tech the savings are there initially with ramp up costs on new Fabs.
Just to add though these cards are evolutions of the Hawaii and Maxwell models, so best history trend would be from around a bit before 970/980 released (they were Sept 2014 with 390/390x June 2015) to now (most recent being Fury/980ti) from both manufacturers.
Pricing model will be tricky for AMD as they were squeezed by dropping the Hawaii price a lot once it had been out awhile, but its rrp was not that low for 290/290x at release - albeit still cheaper than NVIDIA.
We also know Lisa Su wants to protect margins and products better this time round than before, and this makes sense because they did drop way too low and that hurt their business model moving forward.
Cheers
Anarchist4000
[H]ard|Gawd
- Joined
- Jun 10, 2001
- Messages
- 1,659
It still leaves the question of what 2 Vega variants are replacing in less than a year.
I'm not so sure. The issue with HBM was cost and the 4GB limit. 4GB isn't a concern for a mid-level part, and it's possible that capacity doubled if anyone transitioned the memory to FINFET. HBM1 was a 20nm process last I checked. I could see a Polaris 10/490 with 4GB of HBM1 as a nano replacement. 490X could be a small(Polaris 10 equivalent) Vega with 8GB HBM2. That still leaves a big Vega as a Fury replacement.
As I mentioned in another thread, adoption of ASTC might account for some of that. Those bus sizes aren't necessarily horrible for low to mid tier parts. Add HBM1 to the highest SKU out of that and it wouldn't really be a regression. It would also be difficult to overclock parts to match performance for bandwidth constrained SKUs.
As I pointed out above, it seems reasonable a 490 could be Polaris 10 with HBM as a nano replacement. The X variant would arrive later. That could be the equivalent of a $350-400 Nano because the die would still be much smaller.
Crossfire/SLI are largely irrelevant with DX12 as it's up to the game to implement. Profiles are no longer needed. A coherent bus like NVLink might make it trivial as well.
It's not really undercutting if you end up with a substantially cheaper part. If they had a part that outperformed and costed substantially less, pricing it under the cost of the competition where you still make good money would do a lot.
As I pointed out above, I'm not sure BOTH of those would be Polaris 10. That would almost have to leave one of the Vega chips as a purely mobile part needing 8GB+ HBM2, instead of 4GB HBM1, with the other as a halo product to replace Fury.
Ideally they need a ~$200 equivalent to a 290x out of Polaris 10, unless Polaris 11 can pull that off. Stick HBM1 on the top SKU for Polaris 10 and you might pass a Nano while coming short of a full FuryX.
Zion Halcyon
2[H]4U
- Joined
- Dec 28, 2007
- Messages
- 2,108
Per the rumors, 490 and 490x are Polaris 10 variants and part of the Computex launch.
Vega is its own beast and launching in 2017.
Don't shoot the messenger
TBH, I can't wait until Computex happens, and hopefully we can start dealing with facts and not rumors.
lazz
Limp Gawd
- Joined
- Apr 15, 2007
- Messages
- 324
All the rumours so far on 490, 490X as of last week:
AMD Radeon R9 490X and Radeon R9 490 With Polaris 10 GPU Rumored For Late June Release - Paper Launch Expected at Computex
AMD Radeon R9 490X and Radeon R9 490 With Polaris 10 GPU Rumored For Late June Release - Paper Launch Expected at Computex
lucidrenegade
Limp Gawd
- Joined
- Nov 3, 2011
- Messages
- 388
Async computing shall save us all! The world shall be Zen.
TaintedSquirrel
[H]F Junkie
- Joined
- Aug 5, 2013
- Messages
- 12,763
I think all the 490/490X rumors ended up being the 480~470 specs we're seeing today.
If they exist we should get leaks about them in the coming weeks.
If they exist we should get leaks about them in the coming weeks.
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Per the rumors, 490 and 490x are Polaris 10 variants and part of the Computex launch.
Vega is its own beast and launching in 2017.
Don't shoot the messenger
TBH, I can't wait until Computex happens, and hopefully we can start dealing with facts and not rumors.
The article YOU posted, as the OP of this thread, directly contradicts what you are saying; the claim here is that Polaris 10 is a ~2560 SP gpu and is set to replace 380/x. If you don't believe the rumor why even post it ?
Last edited:
The only way Pascal will offer 980ti performance at the same price point is if AMD completely fails this generation and NV knows it beforehand. Possible, sure. Unlikely though.
(Yeah, fifty bucks-- you can get a 980ti for <$600 today.)
Well I would had expected that to leak faster than Usain Bolt running the 100m
If that was to happen pretty sure we would had heard some rumour, furthermore what would the die size (PHY interconnect consideration) need to be to integrate with HBM - this is probably one reason the rumours are focusing on the 256-bit bus with GDDR5X that many of us are leery about for both manufacturers.
But then NVIDIA managed to keep their products pretty quiet up to now, although they are now also leaking like a sieve.
Cheers
DooKey
[H]F Junkie
- Joined
- Apr 25, 2001
- Messages
- 13,583
It's called fanboyism.
Why would polaris need HBM or HBM 2? Does it make sense it they have improved their architecture to better utilize the available bandwidth for a chip that probably won't have the performance to use that much bandwidth make any sense? Cost also is more for HBM, how can they slot Polaris with a totally different bus another words an entirely new chip (AKA VEGA, Polaris with more ALU"s) into a line up all the while saving money and not stepping on ones own toes.......
Also how does this align with Raja's statement of only 2 new chips this year?
Also how does this align with Raja's statement of only 2 new chips this year?
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
What if they design two completely different GPUs, one with gddr controllers, one with hbm controllers... asynchronously
Well that is Vega lol,
nice pic lol,
To use HBM and HBM 2 it has to be cost effective and really needs the performance of a chip higher than a 980ti (a lot higher) to use that much bandwidth effectively.
980ti has what 40% less bandwidth over a Fury X but both chips are equal in performance that tells us a chip that even if it comes close to the 980ti performance doesn't need 500gb/sec bandwidth then you add in cost, its not cost effective.
nice pic lol,
To use HBM and HBM 2 it has to be cost effective and really needs the performance of a chip higher than a 980ti (a lot higher) to use that much bandwidth effectively.
980ti has what 40% less bandwidth over a Fury X but both chips are equal in performance that tells us a chip that even if it comes close to the 980ti performance doesn't need 500gb/sec bandwidth then you add in cost, its not cost effective.
Ieldra
I Promise to RTFM
- Joined
- Mar 28, 2016
- Messages
- 3,539
Are there planned upgrades to memory compression on Polaris ? I heard about texture compression, but nothing about improve color compression, would help alleviate added stress from going to 256-bit bus with ~390 level performance
Well low hanging fruit like color compression I expect them to do that. Memory compression, I don't know... possibly but it all comes down to whats in the realm of changes they are doing. According to the polaris slides, they could do more with it yeah but that isn't what held back the current GCN architecture. I would expect the obvious front end bottlenecks to be on the more priority list than memory compression.
Anarchist4000
[H]ard|Gawd
- Joined
- Jun 10, 2001
- Messages
- 1,659
A single stack of HBM will have 1024 lanes just for data and it's ~40mm2. 4x40mm=160mm2 then room for other connections, power, etc which probably won't number close to the 4096 lanes you've already added. Polaris 10 was what, 232mm2? Doesn't seem like it would be a problem, and that still assumes HBM1 is at the limit on spacing. The issue was more the quantity of traces the PCB would require for the HBM interface if you didn't have an interposer providing much denser traces.Well I would had expected that to leak faster than Usain Bolt running the 100m
If that was to happen pretty sure we would had heard some rumour, furthermore what would the die size (PHY interconnect consideration) need to be to integrate with HBM - this is probably one reason the rumours are focusing on the 256-bit bus with GDDR5X that many of us are leery about for both manufacturers.
But then NVIDIA managed to keep their products pretty quiet up to now, although they are now also leaking like a sieve.
Cheers
TaintedSquirrel
[H]F Junkie
- Joined
- Aug 5, 2013
- Messages
- 12,763
So it overclocks by 50%?
No, But 30 % is not a long shot.
Anarchist4000
[H]ard|Gawd
- Joined
- Jun 10, 2001
- Messages
- 1,659
FuryX is 4096 processors, Polaris 10 is allegedly 2560 processors. So could they get a 60% performance increase out of their move to FINFET and architecture tweaks when they publicly stated 2.5x perf/watt if you completely disregard power efficiency? That's assuming FuryX performance when it'd be aimed at a Nano. Maybe some of the bandwidth is unnecessary, but the form factor would still be a selling point.
A nano is ~$500 right now, a 390X ~$430. So does 596mm2 to 232mm2 save $70 to still be profitable? There are already cases where a 390X beats a 980ti with newer games, so doesn't seem unreasonable.
KazeoHin
[H]F Junkie
- Joined
- Sep 7, 2011
- Messages
- 9,109
InB4 Fiji is rebranded as the 490 and 490X.
By the way 2.5 is 250 %
What is the price difference per wafer for 14nm vs 28 nm? and the extra cost of HBM, interposer and assembly, That doesn't sound like it would be any savings at all.
And the 2.5 is over Hawaii........
FinFet gives 60% performance per watt at the same performance or 30% increase performance at the same wattage.
So to if we ignore 2.5 performance per watt, cause with unknown variables architecture, that is what we get max.
D
Deleted member 222586
Guest
I think many of you are missing something very big: the fact that this is the first time in history that price per gate will increase, instead of decrease. Meaning... that if you were to simply downsize any of the current 28nm chips to Fin Fet... they would end up costing more, not accounting for R&D and other crap.
So... they can't simply release 390/X performance-like as their top dog in Fin Fet. They can't. Simply put: a 390/X-like in Fin Fet will probably be more expensive than it is on 28nm so... why bother? IMO, AMD will start by replacing Fiji first. It also makes some sense because R9 390 already has 8GB (a move that made no sense because Fiji only has 4GB. Its a huge W-T-F situation when the top-tier card has less memory than the next one in line). It also makes sense to start replacing the biggest chip you have... since its the most expensive to produce. And at 600mm^2 + HBM it can't ever be cheap. Ever.
I don't think we have ever seen a node-change in which they kept the top tier cards from last generation.
G80? Nope.
G92? Well they kept the Ultra but... I wouldn't count it.
GT200? Nope
GF100? Nope
GK104?
Well, see my point? Yes, this is NVIDIA.. but AMD did exactly the same. They have never kept the high-end card of the last generation once they start on a newer one. Why? Because on higher-end, big chips... it is profitable to replace them with smaller ones on newer nodes. But it also happens that if the node is far more expensive than it was before... the only place where it makes sense to replace anything is precisely at the high end. Why? High margin cards allow for more things than low margin ones. And the cards with the higher margins are the ones on the higher end of the scale.
I don't know. Economically, it makes 0 sense to release very expensive products to make at the same performance bracket you have already covered. 0 sense. Specially with this FinFET stuff...
My prediction? We should see Fury performance at $400, and Fury X+ at $550. We know that Polaris 11 on its lesser part should have ~GTX950 performance (for what reporters were shown back in january), and thus the XT version should get up to GTX960 no problem. This puts Polaris 10 in 970/980/TI territory without a hiccup. Which completely makes sense, of course.
So... they can't simply release 390/X performance-like as their top dog in Fin Fet. They can't. Simply put: a 390/X-like in Fin Fet will probably be more expensive than it is on 28nm so... why bother? IMO, AMD will start by replacing Fiji first. It also makes some sense because R9 390 already has 8GB (a move that made no sense because Fiji only has 4GB. Its a huge W-T-F situation when the top-tier card has less memory than the next one in line). It also makes sense to start replacing the biggest chip you have... since its the most expensive to produce. And at 600mm^2 + HBM it can't ever be cheap. Ever.
I don't think we have ever seen a node-change in which they kept the top tier cards from last generation.
G80? Nope.
G92? Well they kept the Ultra but... I wouldn't count it.
GT200? Nope
GF100? Nope
GK104?
Well, see my point? Yes, this is NVIDIA.. but AMD did exactly the same. They have never kept the high-end card of the last generation once they start on a newer one. Why? Because on higher-end, big chips... it is profitable to replace them with smaller ones on newer nodes. But it also happens that if the node is far more expensive than it was before... the only place where it makes sense to replace anything is precisely at the high end. Why? High margin cards allow for more things than low margin ones. And the cards with the higher margins are the ones on the higher end of the scale.
I don't know. Economically, it makes 0 sense to release very expensive products to make at the same performance bracket you have already covered. 0 sense. Specially with this FinFET stuff...
My prediction? We should see Fury performance at $400, and Fury X+ at $550. We know that Polaris 11 on its lesser part should have ~GTX950 performance (for what reporters were shown back in january), and thus the XT version should get up to GTX960 no problem. This puts Polaris 10 in 970/980/TI territory without a hiccup. Which completely makes sense, of course.
D
Deleted member 222586
Guest
I can try to find the sources if you like... but the graphs I saw put the FinFET wafer at 2x the price of 28nm. Considering the density... gates end up being more expensive on FinFET than on 28nm. And if we add that yields are probably not good...
Anarchist4000
[H]ard|Gawd
- Joined
- Jun 10, 2001
- Messages
- 1,659
I was under the impression Fury already used HBM and an interposer. The real question is price change from the move to 14nm. Wafer cost is up, but chip size is substantially smaller. If it's actually power efficient they could save a bit on the board as well.
So wafer cost doubles and is replaced by a chip 39% of the size that just needs a <60% performance jump. That still seems reasonable to me. Vega as a late replacement for the higher end fiji chips with HBM2 still makes sense. A mobile Vega for the 490X with 8GB and the big Vega for Fury replacement.I can try to find the sources if you like... but the graphs I saw put the FinFET wafer at 2x the price of 28nm. Considering the density... gates end up being more expensive on FinFET than on 28nm. And if we add that yields are probably not good...