20nm performance and pricing expectations

[tviceman]

What are your expectations for performance with AMD's and Nvidia's first-out flagship video cards on 20nm? I believe Nvidia will lead with a ~300mm^2 die again before a big die (like they did with Kepler) so I believe that will be used in Nvidia's initial "flagship" video card. I also think much of 20nm will hinge on whether/when GDDR6 will be available for mass production, so both companies could wind up waiting/delaying if that is the case. So, relative to Titan performance, what do you think we'll see from AMD and Nvidia's first gen flagship 20nm video cards? And what about prices? Do you think AMD will lead with $550 again, come in cheaper, or go higher? And what about Nvidia?

I'd really like to keep this thread troll free for at least a few pages (is that even possible) and have a legitimate, thought provoking discussion on where we think performance and prices will end up. So please keep it civil and don't attack people for their opinions & expectations for something we know nothing/little about.

 

[aaksheytalwar]

I guess AMD would have an initial flagship 20-30% faster than Titan out of the box for about $550-700ish.

 

[Rvenger]

Well, of course AMD will be first out with Volcanic Islands. I think they will be very quiet about a release of a big die and come out with one over ~ 500mm^2. It will be about 25-30% faster than Titan while consuming the same amount of power. They will price it at ~ $600 and it will have 6gb of 7ghz GDDR5. I believe AMD will attempt to hit the ground running with GCN 2.0. The HD7790 proved to be an efficient part and it can only get better IMO. Drivers should be pretty stable upon release due to it being a similar architecture.

 

[96Firebird]

All I can go on is past node jumps, so looking at GTX 580 --> GTX 680, there was a ~20% jump across all resolutions at launch. Now this also resulted in a smaller die and less power usage. If we take a look at GTX 580 --> Titan, there was a 64% jump in performance across all resolutions at launch. So, for Nvidia, I'd say if they can get the big die out as their flagship for the first generation of 20nm, they may be able to net 40-50% performance gains across all resolutions. Like you said though, this depends on not only GPU power but the VRAM as well. If they can only manage a small die, like they did with the GTX 680, I'd expect equal to slightly better performance compared to the Titan at launch. Of course, as is always the case, drivers will help push these numbers up as time goes on.

As for AMD, I am more in the dark about that. They could change their strategy and shoot for the performance crown right off the bat, designing a big die competes with Nvidia's big die. Or they may go in the complete opposite direction and cut the compute and go pure gaming on a smaller die. I think they are more flexible on their architecture changes as they don't have a reason to head in one direction or the other, while Nvidia will always be designing compute GPUs for their extensive professional market. For the sake of the thread, I'll say AMD will come out with a GPU that will be ~10% faster than the Titan at launch.

As for prices, I don't even want to comment. Who knows what the economy will look like then, and who knows how much each company will be charged for wafers at 20nm.

 

[Erenhardt]

I also think much of 20nm will hinge on whether/when GDDR6 will be available for mass production, so both companies could wind up waiting/delaying if that is the case.

Or they may release with GDDR5 and later when GDDR6 is available they will put it on PCB with chips that have disabled memory controllers, making up for smaller memory bus, so nothing is wasted. High-end and flagships will not follow that rule, but will come with GDDR6 by default, while low end will stay with cheaper GDDR5

 

[Imouto]

I think Nvidia will be late again and will take about 2 years to release another 5xx mm^2. Ppl keep forgetting that both AMD and Nvidia have almost the same gaming power per silicon square mm on their gaming GPUs. About Tahiti vs GK110 in compute you can see a sick value in the former. With everyone going OpenCL (even the F@H Nv stronghold) I really hope Nvidia will stop crippling its OpenCL performance.

Anyway AMD has cool hardware that's likely to get better in the next refresh. Someone said that putting powerful hardware together is cheap and the software side is damn expensive and that's where AMD is failing hard with its OpenCL compiler far away from optimal, sub par consumer drivers and the never-ending lag fixing and developing stuff.

So I think this generation will repeat all over again unless AMD starts investing seriously in the driver/software department. AMD having the best bang for the buck with the cheapness it implies and Nvidia with its smooth sailing feasting on AMD sales while the market shrinks.

 

[Jaydip]

I said in the other thread but it seems NV has mastered the 28nm process.NV could beat 7979 GHz by architecture superiority alone.Maxwell is a brand new architecture while VI is just a iterative evolution like Kepler.So I wouldn't be too surprised if the current pricing strategy from both camps will continue.AMD will debut a card which will probably outperform Titan by roughly ~5-10% and will be priced @ $749 and then the same story will repeat.

 

[VulgarDisplay]

As others have stated, I fully expect AMD to swing for the fences with a big die. They need to make a splash. Nvidia will continue the trend on 20nm.

I can see the first iteration being released with GDDR 5 and a refresh coming with something new.

 

[DooKey]

I see both camps releasing their flagship at $649 to $699.

 

[SirPauly]

It may depend on market pricing, supply and demand, yields, performance, costs and what company executes first!

 

[tential]

It may depend on market pricing, supply and demand, yields, performance, costs and what company executes first!

The smart answer. Also, I think it'll do with how well the GTX 780 sells. I think everyone can agree, that Titan is a little out there at $1000. However, $650 dollars isn't. If the GTX 780 does very well, compared to the GTX 680, at the $650 price point we could see that, or somewhere close to that as the new standard for the 780/x970.

A lot of people were willing to pay $600+ for the GTX 780, in fact saying they'd buy two. I don't see a good reason why Nvidia and AMD wouldn't release their flagship (I'm excluding Titan unless I'm unaware that Nvidia plans to keep this as a long term product line for years to come), at $600+.

Why charge less than what the majority of people are willing to pay?

 

[tviceman]

Hmmm interesting to see some people think AMD is going to attempt a big die. Haven't they only made one GPU larger than 400mm^2? Regardless, I like to look at performance improvements that are made on gpu's relative to what they're replacing from a physical standpoint. Tahiti replaced Cayman, Pitcairn replaced barts, GK110 replaced GF110, and GK104 replaced GF114.

 

[omeds]

I'm guessing (hoping) AMD will release something faster than Titan, by ~20%, and will be priced around current GTX780 prices, perhaps a little lower.

 

[tviceman]

Here is an interesting bit of data which may (or may not) be indicative of 20nm cards.

HD6970 --> HD7970 (vanilla) = 47% increase @ 25x16
HD6970 --> HD7970GE = 61% increase @ 25x16

Initially, the hd7970 was only 39% faster than the hd6970 when it came out back in Jan. '12. It took some time, but AMD's Tahiti matured into a very worthy successor over the 6970. So for AMD's first 20nm flagship single GPU card, I think a ~45-50% increase over hd7970GE is likely.

So, I think the hd8970 will be about ~10-12% faster than Titan and priced at $599.

The Nvidia lineup is a little harder to discern. GK104 was used in products to replace GF110, but was not specifically GF110's replacement. Therefore I think it is logical to compare GK104 to GF114 to extrapolate performance gains as I think GM104 will be powering Nvidia's GTX880.

GTX560ti --> GTX680 = 93% increase at 25x16, 77% increase at 19x10
GTX560ti --> GTX770 = 112% increase at 25x16, 94% increase at 19x10 (I'm guessing gtx770 will be 10% faster than gtx680)

Nvidia made some absolutely crazy gains with GK104 over GF114, some of which can be attributed to inadequate amounts of vram for the gtx560ti at 25x16 resolution. Still, GTX770's projected 10% increase over the GTX680 still gives GK104 about double the performance over a gtx560 at 1080p or 1440p. I have no idea if Nvidia can make these kinds of gains again though.

 

[Rezist]

I think amd will release the succesor to the 7870 first on a 256 bit bus with faster gddr5 it will probably be 5-10% faster then the 7970ge and retail at 349 to 399 which will be there good money making card it wont beat titan but will be comparable for less money. I honestly think amd will never get consumers to pay over 500 for one of their cards so they need architecture efficiency.

 

[bunnyfubbles]

AMD got bigger with the 7900s, and I'm not sure they'll continue on that same path, that's an arms race that traditionally hasn't worked out particularly well for them. Then there's also the fact that nVidia also increased their gaming efficiency with Kepler.

That being said, I really don't expect AMD's 20nm flagship to beat GK110, maybe it will match, but honestly I expect a focus back towards efficiency and competitive pricing.

 

[SiliconWars]

Tahiti is smaller than Cayman.

 

[wand3r3r]

If it's faster than titan (I presume so, unsure big vs. small die) here's my estimated magical price formula.

$550 if around titan or under
+$100 for every 10% faster

If it's not faster then titan, NV can just bump titan clocks up 10% and keep on raping everyone ($1k) and AMD can just fold for being too complacent.

I'll estimate a 50% gain over 7970. So around titan + 20%.

 

[hjalti8]

So, I think the hd8970 will be about ~10-12% faster than Titan and priced at $599.

I really don't expect AMD's 20nm flagship to beat GK110, maybe it will match, but honestly I expect a focus back towards efficiency and competitive pricing.

If the next high end 20nm chip from amd(or nvidia) would be ~400mm^2 it should have 30-45% more transistors than GK110.
If such a chip would only be 10% faster than GK110, which has horrible perf/transistor(when you look at gaming), it would be a disaster.

Hell, a 20nm 300mm^2 chip designed primarily for gaming should outperform the GK110. Just like the 680 easily beat the 580.

 

[ams23]

So, relative to Titan performance, what do you think we'll see from AMD and Nvidia's first gen flagship 20nm video cards? And what about prices?

Well, the measuring stick has just moved, because greater than Titan single GPU performance is available today for $659 USD from the GTX 780 SC ACX (with +45.7% advantage over GTX 680 and +42.3% over HD 7970 GHz Ed. at 25x16 resolution across a wide variety of games): http://forums.anandtech.com/showpost.php?p=35059456&postcount=526 . So a Maxwell flagship worthy of the name GTX 880 will need to be significantly faster than this.

 

[SiliconWars]

Yeah I forgot it's only Nvidia cards that can be overclocked.

 

[ams23]

Yeah I forgot it's only Nvidia cards that can be overclocked.

Are you talking to me? Anyway, while the GTX 780 SC ACX uses higher-than-reference clock operating frequencies, the TPU test results were based on the standard SC clocks, not the max possible clocks! TPU's GTX 780 SC ACX sample was able to achieve ~ 11% higher performance when using max clocks vs. standard SC clocks. Considering that a typical 7970 GHz Ed. overclocked or GTX 680 overclocked would result in approximately 11% higher performance than stock clocks, the GTX 780 SC ACX would be able to approximately maintain it's 42-45% performance advantage when all cards are overclocked to max clocks. So again, the bar is set very high for next gen Maxwell high end parts.

 

[parvadomus]

I think they will release a 360mm2 die 40% faster than GTX780. Later a refresh 60% faster than GTX780 of about 450mm2.
ED : talking about AMD

 

[Silverforce11]

The reason NV lead with gk104 was because TSMC could not make enough functional gk110 (obviously after gk100 went awol), and NV had a much higher % profit segment in HPC to fill as well as a backlog of huge orders for supercomputers. If they could have made gk110 in volume, you would have seen Titan as the flagship GPU.

As to whether TSMC can deliver good yields right away on 20nm, their history suggests they simply can't deliver on a new node. But we wont know how things will be until it will be.

The expectation here is that both architectures will be tweaked or evolved and not a revolution to a new arch, as such, at the same die space, its not unrealistic to expect ~80% performance gains from both companies. AMD have shown with the 7790, that they are able to tweak GCN for more efficiency easily, so it should be a great fight in the 300-400mm2 range.

AMD simply cannot afford to push ahead with a 550mm2 die, because they lack the HPC market support which NV has, where margins are high to offset the lower yields.

 

[Kippa]

I was thinking that a December launch seems so far away, then I realised that we are nearly in June and that it is only 6 months off which isn't that long. As for strategy I think AMD are keeping their cards close to their chest. Not hearing anything much from AMD doesn't mean they're behind or lagging.

If I were a betting person, I would bet that we hear something more concrete in the next month or two from AMD about their up coming products.