AMD 20nm GPU Discussion Thread

[ShintaiDK]

I just hope the high end lets me max The Witcher 3 at an insane res, that's all i care about.

Highend on the node may come much later, as seen with Hawaii.

 

[Mondozei]

GCN 1.1 barely offers anything.

lol

AMD 20nm GPUs will sport an improved GCN 2.0 architecture and most likely HBM (stacked memory).

http://electroiq.com/blog/2013/12/amd-and-hynix-announce-joint-development-of-hbm-memory-stacks/

I am guessing 4096 stream processors @850 - 900 Mhz (because TSMC 20nm will be a power constrained node), 64 CU, 64 ROPs and 4096 bit memory (4 HBM stacks running at 1 Ghz). With architectural improvements and improved per core performance a 40% improvement over R9 290X is possible.

http://sites.amd.com/us/Documents/TFE2011_006HYN.pdf

Interesting links, thanks. The GCN 2.0 sounds plausible, and if your ballpark estimates are even close to correct, it would give Maxwell a run for its money. I would still guess that it would lose in absolute performance, but that would mean AMD would drive its prices lower, thereby checking Nvidia's greed if the new cards are not too far apart performance-wise. (And checking Nvidia's greed is Lord's Work). So all in all, a continuation of today. (Of course everything I said applies outside of areas where e-tailers are price gouging for cryptocoinbonanza).

 

[VulgarDisplay]

Sales would drop. Meaning smaller cashflow. And there is only so much you can do on the same node. AMD got GCN 1.1 as their new uarch. nVidia got Maxwell. GCN 1.1 barely offers anything. And Maxwell on 28nm while great performance/watt wise, is still not what a 20nm chip would bring. And people are not upgrading their old cards for that.

And Broadwell with expanded GT3e models on 14nm is also eating its share. Mobile is already a lost cause fo dGPUs.

GCN 1.1 barely offers anything? What?

Nearly 780ti performance in around 100mm2 less die size? XDMA making crossfire blow SLI away in terms of scaling and performance?

We get you don't like AMD but at least acknowledge their engineering feats in GCN 1.1.

 

[parvadomus]

More speculation. But I think the codenames might be right Bermuda, Fiji and Treasure Island.

 

[Ed1]

AMD 20nm GPUs will sport an improved GCN 2.0 architecture and most likely HBM (stacked memory).

http://electroiq.com/blog/2013/12/amd-and-hynix-announce-joint-development-of-hbm-memory-stacks/

I am guessing 4096 stream processors @850 - 900 Mhz (because TSMC 20nm will be a power constrained node), 64 CU, 64 ROPs and 4096 bit memory (4 HBM stacks running at 1 Ghz). With architectural improvements and improved per core performance a 40% improvement over R9 290X is possible.

http://sites.amd.com/us/Documents/TFE2011_006HYN.pdf

I don't think we will see 3d/stacked memory in next batch of cards .
Memory manufactures are just talking about these so I would give it at least a yr or so IMO .

 

[ShintaiDK]

GCN 1.1 barely offers anything? What?

Nearly 780ti performance in around 100mm2 less die size? XDMA making crossfire blow SLI away in terms of scaling and performance?

We get you don't like AMD but at least acknowledge their engineering feats in GCN 1.1.

Compared to GCN 1.0 its a small update. And nothing in terms of what we have seen with Maxwell.

100mm2 less is not due to GCN1.1. Hawaii is essentially just a double Pitcairn. And the skyhigh temperatures is a result of that.

XDMA and trueaudio is the real only benefits of GCN 1.1.

Even Anandtech is quoted for saying:

Ultimately the differences between GCN 1.0 and GCN 1.1 are extremely minor, but they are real.

The only hope is to see if AMD will have GCN 2.0 ready for H2.

 

[Ed1]

lol



Interesting links, thanks. The GCN 2.0 sounds plausible, and if your ballpark estimates are even close to correct, it would give Maxwell a run for its money. I would still guess that it would lose in absolute performance, but that would mean AMD would drive its prices lower, thereby checking Nvidia's greed if the new cards are not too far apart performance-wise. (And checking Nvidia's greed is Lord's Work). So all in all, a continuation of today. (Of course everything I said applies outside of areas where e-tailers are price gouging for cryptocoinbonanza).

AMD already is 512bit bus so memory bus BW is pretty good already , just up the mem speed and that would be enough for next chip .
They would need a much faster chip to be memory staved with 512 and decent memory .

 

[parvadomus]

GCN 1.1 barely offers anything? What?

Nearly 780ti performance in around 100mm2 less die size? XDMA making crossfire blow SLI away in terms of scaling and performance?

We get you don't like AMD but at least acknowledge their engineering feats in GCN 1.1.

GCN 1.1 does not even exist. Its just a refresh of the original GCN with boost and XDMA, and the new chips are configured better than the previous ones. Hawaii is basically 2 Pitcairn GPUs in 1 and offers only a bit better performance due to having more shaders. Its just like Fermi 1 and Fermi 2, just some minor tweaks.
If you think AMD will continue with the same architecture for 3 or more years, you must be silly at best.

 

[VulgarDisplay]

GCN 1.1 does not even exist. Its just a refresh of the original GCN with boost and XDMA, and the new chips are configured better than the previous ones. Hawaii is basically 2 Pitcairn GPUs in 1 and offers only a bit better performance due to having more shaders. Its just like Fermi 1 and Fermi 2, just some minor tweaks.
If you think AMD will continue with the same architecture for 3 or more years, you must be silly at best.

Where did I say they are staying on the same architecture?

 

[parvadomus]

Where did I say they are staying on the same architecture?

It was just a general comment, theres a lot of people that thinks that way.

 

[VulgarDisplay]

Who? If anything AMD has really stepped up their GPU design game and nvidia should be pulling out all the stops to make sure they stay king of the hill.

 

[NTMBK]

Compared to GCN 1.0 its a small update. And nothing in terms of what we have seen with Maxwell.

100mm2 less is not due to GCN1.1. Hawaii is essentially just a double Pitcairn. And the skyhigh temperatures is a result of that.

XDMA and trueaudio is the real only benefits of GCN 1.1.

Even Anandtech is quoted for saying:
The only hope is to see if AMD will have GCN 2.0 ready for H2.

Hawaii also has 1/2 DP rate, which is a very big deal for the compute market. Pitcairn by comparison only has 1/16, and Tahiti 1/4.

 

[TreVader]

Hawaii already has nvidia in a hard place to be on ROP and in performance per mm2, and once they produce a 500mm^2 die nvidia will need 600mm^2 to compete, or they're going to lose their power efficiency and heat characteristics.

The other thing you have to consider is how awful nvidia is at compute. AMD now provides more DP FLOPS on one hawaii die as nvidia does with 2 Titans or Titan Z. The reason AMD isn't re-designing their architecture completely is they've already shown themselves capable of producing quality compute products.

 

[PPB]

Not even mentioning a more granular control over power doubling power states and lowering the ms between each power state change.


Yeah, GCN 1.1 didnt bring anything. Nope, not at all.

 

[Vesku]

If AMD can repeat the density achievement they managed with Hawaii then 20nm will be pretty good for them. Unlike CPUs with GPUs lower clocks, from 20nm being a low power only node, can be offset fairly easily with increasing parallelism.

 

[Headfoot]

Not even mentioning a more granular control over power doubling power states and lowering the ms between each power state change.


Yeah, GCN 1.1 didnt bring anything. Nope, not at all.

Nor did it bring dedicated audio dsps. Or the highest number of ROPS ever featured on a single die

 

[piesquared]

Nor did it bring an extremely efficient 512 bit GDDR5 bus.

So dual Hawaii will bring a 1024 bit bus, which is the HBM interface. So the bus is ready for 20nm and HBM? Or possibly dual Hawaii incorporates HBM.

 

[tviceman]

Hawaii already has nvidia in a hard place to be on ROP and in performance per mm2, and once they produce a 500mm^2 die nvidia will need 600mm^2 to compete, or they're going to lose their power efficiency and heat characteristics.

Your analysis is entirely (and predictably) one sided. Everyone knows GK110 was designed for Nvidia's HPC ambitions first and foremost, hence the reason GK104, 106, and 107 have significantly higher perf/mm^2.

If you want to compare perf/mm^2 look at GCN 1.1 (either Bonaire or Hawaii) and compare it to Maxwell. It took AMD about the same amount of time to upgrade GCN 1.0 to 1.1 as it did for Nvidia to move from the first Kepler to Maxwell. GM107 has a smaller die and less transistors than Bonaire, but is faster and way more efficient. Nvidia is doing way more with less.

But if you want to stick with GK110 vs. Hawaii comparisons, consider power draw, heat, and all the other functions GK110 is capable of that Hawaii isn't. It is very obvious Hawaii is being pushed harder as the r290x than GK110 is as the gtx780ti. Also the exclusive compute specific features in GK110 don't come for free - die space and transistors are used. In fact, insofar as performance is concerned, stock gtx780ti vs. stockr290x is pretty similar to gtx580 vs. hd6970 (the previous node flagship cards). And when looking past that sole comparison, GK110 looks better in every regard vs. Hawaii than GF110 did vs. Cayman.

 

[VulgarDisplay]

Is the gtx780ti 15-20% faster than the 290x?

 

[3DVagabond]

Sales would drop. Meaning smaller cashflow. And there is only so much you can do on the same node. AMD got GCN 1.1 as their new uarch. nVidia got Maxwell. GCN 1.1 barely offers anything. And Maxwell on 28nm while great performance/watt wise, is still not what a 20nm chip would bring. And people are not upgrading their old cards for that.

And Broadwell with expanded GT3e models on 14nm is also eating its share. Mobile is already a lost cause fo dGPUs.

Are you listening to what's being said? 20nm is going to have an extremely short life. At this point we haven't even heard a date for 20nm HP/UHP production (as far as I know). It could only end up being 6-12 months before 16nmFF (We all know it's not true 16nm, but it's what they are calling it, so...) comes along. If that's true there is no way either company is going to lose millions designing chips for a dead process. They'll spoon feed us 28nm until 16nmFF comes out.

Damn, I came in here to read about AMD 20nm info, only to find a boatload of Nvidia stuff. Weird.

You aren't really surprised, are you?

 

[el etro]

GCN1.1 is updated in the compute front too, and have better perf/watt(against Tahiti and pitcairn cards) in HD7790 form. See ST review for details.

 

[NostaSeronx]

...4096 bit memory (4 HBM stacks running at 1 Ghz).

Each HBM memory stack is actually connected to one 128-bit DDR channel interface.

500 MHz -> DDR 1 GHz @ 1.2v

The 128-bit channel interface can clamshell up to x1024 I/Os. Each channel of 128-bit can provide at 1 GHz; 128 GByte/s.

128 GByte/s * 4 = 512 GByte/s

Now compare that to GDDR5 in Hawaii; (512 * 5 GHz) / 8 => 320 GByte/s
---
http://i.imgur.com/DFvAZLo.png

--
Ya nevermind, you're right.

 

[tviceman]

Is the gtx780ti 15-20% faster than the 290x?

Don't put words in my mouth. I said the performance delta between comparing gtx580 to hd6970 "was similar" to gtx780ti and r290x.

The performance gaps look pretty similar. If you want to throw in 4k then sure, Hawaii looks way better, but I'm neither interested in using qualifications nor discussing what 1% of the one percenters use.

The other metrics in which GK110 looks better vs. Hawaii than GF110 did vs. Cayman are in thermals, power, and die size. Hawaii had a substantially larger die, power, and thermal increase over AMD's previous node best than GK110 had over GF110. And also like I said, it's universally accepted that a reference r290x has less OC headroom than a reference gtx780ti.

 

[blackened23]

The performance gaps look pretty similar. If you want to throw in 4k then sure, Hawaii looks way better,

I'm consistently seeing 4k resolutions touted as the 290X advantage. Tomshardware revealed a bug with NV drivers which caused performance anomalies at 4k + MST. This was the driver used late last year for 4k testing by various websites. Basically, the bug caused GPU utilization to be limited to 70-80% while in dual screen MST mode in 3d apps. Dual screen MST is essentially 4k resolution, MST "ties" 2 1920 resolutions screen together to create a form of surround.

While the Hawaii is a good GPU considering it's nearing MSRP levels once again, 4k gaming is no longer an advantage. I saw the NV driver bug mentioned on tomshardware - they monitored GPU usage at 4k benchmarks and the NV solutions generally had 70-80% while the Hawaii cards were 99%+.

That was with the NV drivers late last year. GPU usage displayed problems in dual screen MST mode (which is 4k, obviously), but like I said. This driver bug has been fixed. As a result, NV is back in the lead again at 4k.

In fact, legitreviews tested 4k benchmarks with the 780/780ti at 4k resolution versus the Powercolor PCS factory overclocked 290X GPU. This is the triple slot monstrosity card. Anyway, links:

http://www.legitreviews.com/powercolor-pcs-axr9-290x-video-card-review_138422/3

As I said, Hawaii is fine for the price, but with fixed drivers - it is no longer in the lead vs NV solutions at 4k resolution. Then again, AMD can once again play the value and price/performance role here since it is cheaper than the 780ti. The 290X at MSRP is still a good deal cheaper, of course, and it's still fairly close at 4k but not in the lead any longer.

 

[blackened23]

More 4k benchmarks 290X vs 780ti with the fixed NV driver, the most recent WHQL release:

http://www.guru3d.com/articles_pages/asus_geforce_gtx_780_ti_matrix_review,16.html

and yet more:

http://www.guru3d.com/articles_pages/asus_geforce_gtx_780_ti_matrix_review,17.html

The 780ti has a commanding lead in all games tested @ 4k with the most recent WHQL. But the 290X is a better value since it is the cheaper card, and the mining inflation situation is starting to go away. And they're both fairly close in performance, and both very strong at 4k.