My take on the next 45 days

[raghu78]

Peak is a blurp, avg is 208 over the course of an entire run.

guru3d doesn't give any wattage to the processor at all and does not reduce consumption based on PSU inefficiency.

you are clutching at straws here. you started off saying Titan draws 170w at stock under load. now that you have been proven wrong you come up with some other flimsy excuse . atleast admit that you always exaggerate positively for Nvidia and negatively for AMD.

if you say that Titan draws 170w under load that would mean the HD 7970 Ghz draws 150w under load. :awe:

http://www.hardocp.com/article/2013/02/21/nvidia_geforce_gtx_titan_video_card_review/10#.UiVkx3_3x8E

http://techreport.com/review/24381/nvidia-geforce-gtx-titan-reviewed/14

http://www.hardwarecanucks.com/foru...orce-gtx-titan-6gb-performance-review-17.html

 

[lavaheadache]

Fighting with him is just going to drag this thread into the garbage. I used to fall into his trap all the time. Now I just sit back and enjoy having the hardware that he argues about.

 

[Silverforce11]

28nm, no. Its unlikely to be more than 30% faster than a 7970 Ghz. Thats just expecting too much since it would HAVE to be a massive die, and AMD has shown no indications they are willing to go down that route.

A small <400mm2 die just does not have the required muscle to challenge Titan.

 

[raghu78]

Fighting with him is just going to drag this thread into the garbage. I used to fall into his trap all the time. Now I just sit back and enjoy having the hardware that he argues about.

yeah I will take your advice and avoid falling for the bait.

 

[BallaTheFeared]

you are clutching at straws here. you started off saying Titan draws 170w at stock under load.

No, actually I didn't. Reading fail just made you think that was what I said. What I said was actually a question.

Let me spell it out in great detail...

Would the 9970 use 170w at stock, and 210w at load with the made up OC potential you created, thus using the same power as the Titan (stock) or 208w while being 20-30% faster at the same power consumption, much like Titan did to the 7970 GHz?

I hope that helps.

 

[3DVagabond]

28nm, no. Its unlikely to be more than 30% faster than a 7970 Ghz. Thats just expecting too much since it would HAVE to be a massive die, and AMD has shown no indications they are willing to go down that route.

A small <400mm2 die just does not have the required muscle to challenge Titan.

Things are pointing towards a larger die. Think "Queen Mary" size.

 

[skipsneeky2]

Wouldn't care if the die size is that of a half dollar,if it can smash a titan and be like half the price at best then its alright in my book.

Maybe i am not as techy as some of you other folks but is it possible that the die size of a gpu or cpu could be as big as a half dollar this way they could do much more with such a large space or does it not really work that way?Why not go bigger and just do so much more over going smaller and having complications?

 

[BallaTheFeared]

Fighting with him is just going to drag this thread into the garbage. I used to fall into his trap all the time. Now I just sit back and enjoy having the hardware that he argues about.

Stay classy :thumbsup:

 

[RussianSensation]

If i'm interpreting that board correctly, it indicates 4096 SPs and a 512bit GDDR5 VRAM bus?

I call BS on the 512-bit memory bus. 7970Ghz has plenty of memory bandwidth. The bottlenecks are all elsewhere. Even then AMD can add 6.5-7 Gbps GDDR5 and end up with more memory bandwidth than the Titan. Why go with a power consuming and more complex 512-bit bus?

 

[Shmee]

Member call outs are not allowed, thanks.

 

[SolMiester]

I always wonder about die size when the process is smaller....we have had bigger dies from both NV and AMD and kept them cool....if the voltage is lower, whats the problem with big dies?

 

[raghu78]

Things are pointing towards a larger die. Think "Queen Mary" size.

yeah thats exactly what it is. my guess is 480 sq mm. let see how close I get with my guess.

 

[raghu78]

I call BS on the 512-bit memory bus. 7970Ghz has plenty of memory bandwidth. The bottlenecks are all elsewhere. Even then AMD can add 6.5-7 Gbps GDDR5 and end up with more memory bandwidth than the Titan. Why go with a power consuming and more complex 512-bit bus?

RS its either 512 bit GDDR5 or 256 bit GDDR6. the memory layout points to one of these two options. the chances of the latter are quite remote as there is no memory manufacturer who is currently shipping GDDR6 in high volume.

 

[lavaheadache]

Things are pointing towards a larger die. Think "Queen Mary" size.

:biggrin:

That thread had so much potential. It's too bad that the forums are becoming so strict. At least this thread is geared to be more or less free for open discussion about a few things.

 

[RussianSensation]

I don't think its likely but it would be nice to get some competition in the high end, especially considering the trend I am seeing in games that aren't a big FPS to perform better on the Titan/780 compared to the 690.

There is no such trend. The reviews which you have used to make such a sweeping conclusion use a flawed methodology for dual-GPU testing by virtue of not waiting for updated SLI/CF profiles, which invalidates the compiled data.

A review that uses SLI/CF profiles in modern games shows that GTX690 is 34% faster at 1080p/1440p while 7990 is 39% faster at 1080p and 43% faster at 1440p than after-market 780 cards (and we know after market 780 that boosts to 1150mhz is faster than the Titan):

http://www.xbitlabs.com/articles/graphics/display/geforce-gtx-780-sli-msi-palit_9.html#sect0

RS its either 512 bit GDDR5 or 256 bit GDDR6.

We don't know. Everything right now is based on pure rumours/speculation.

512-bit bus
- consumes a lot more power
- increases die size pin / PCB complexity
- increases cost

Where is the evidence that Tahiti XT with 288GB/sec memory bandwidth is memory bandwidth constrained? If AMD is indeed using a 512-bit bus, if they pair 6Gbps chips, you end up with 384 GB/sec memory bandwidth. That sounds like an HD4890 repeat where all that extra bandwidth was simply wasted. Why not go for 7Gbps chips on 384-bit?

For all we know that data on the whiteboard can be related to a completely different GPU than Hawaii.

guys looks like AMD has gone massive die. :biggrin:

The article says nothing of the sort. There is no indication of die size given and neither is their a comparison to GK110 to conclude it is a massive die. What is massive? Compared to 365mm2, 420mm2 is massive but compared to 561mm2 it is not.

 

[blackened23]

Why go with a power consuming and more complex 512-bit bus?

Because it sounds awesome of course!

Seriously, who knows. OTOH, there are tons of rumors from chiphell indicating 512 bit. Shrug. It's been done before so it isn't exactly unprecedented (and the prior 512 bit cards weren't exorbitant in price I don't believe). Personally I really don't care about 384 bit or 512 bit, I just want to see the final performance figures.

 

[escrow4]

Why is it always high end? What happened to grunty $250 (or sub $300) GPU's that offer value for money? Something next gen that fits this bill?

 

[blackened23]

Why is it always high end? What happened to grunty $250 (or sub $300) GPU's that offer value for money? Something next gen that fits this bill?

Erm, i'm sure AMD will have SKUs ranging from the high end to the low end based on the new architecture with prices in the stratosphere down to your sub 300$ price point - Both AMD and nvidia always do this every generation.

The high end parts will be released first, though.

 

[Jaydip]

512 bit, that reminds me of HD2900XT

 

[skipsneeky2]

Erm, i'm sure AMD will have SKUs ranging from the high end to the low end based on the new architecture with prices in the stratosphere down to your sub 300$ price point - Both AMD and nvidia always do this every generation.

The high end parts will be released first, though.

gtx680 performance at $200 in the 800 series perhaps?The gtx760 is not that far off at $250 right now.

 

[raghu78]

We don't know. Everything right now is based on pure rumours/speculation.

512-bit bus
- consumes a lot more power
- increases die size pin / PCB complexity
- increases cost

Where is the evidence that Tahiti XT with 288GB/sec memory bandwidth is memory bandwidth constrained? If AMD is indeed using a 512-bit bus, if they pair 6Gbps chips, you end up with 384 GB/sec memory bandwidth. That sounds like an HD4890 repeat where all that extra bandwidth was simply wasted. Why not go for 7Gbps chips on 384-bit?

For all we know that data on the whiteboard can be related to a completely different GPU than Hawaii.

RS the question is does a 384 bit memory bus at 7 Ghz (336 Gb/s) draw less power than a 512 bit memory bus at 5.5 Ghz (352 Gb/s). the issue here is higher clocks on narrower bus vs lower clocks on wider bus. its more a question of which draws lesser power.

The article says nothing of the sort. There is no indication of die size given and neither is their a comparison to GK110 to conclude it is a massive die. What is massive? Compared to 365mm2, 420mm2 is massive but compared to 561mm2 it is not.

the article indicates a 512 bit memory bus. RS remember that memory I/O pads require a certain die size. AMD seems to have had enough pad area aroung the core GPU to fit a 512 bit memory bus. all these are indicative of a massive die. remember GTX 285 had a 512 bit GDDR3 memory bus with a die size of 470 sq mm.

http://www.techpowerup.com/reviews/Leadtek/GeForce_GTX_285/images/front.jpg

http://en.wikipedia.org/wiki/Comparison_of_Nvidia_graphics_processing_units#GeForce_200_Series

I am quite sure AMD have ended up with a similar die size of 470 - 480 sq mm.

 

[RussianSensation]

Yeah, it's not like the 9970 will have aftermarket cards such as the lightning 9970. That's not a serious statement is it? If it matches stock 780 and Titan performance, then overclocked 9970s will surely follow.

I wouldn't make such assumptions that 9970 oc could match 780 oc. HD4890/5870/6970 barely overclocked. 9970 could match after-market 780s but have only 5-8% overclocking headroom. GTX780s overclock 20-30%.

the article indicates a 512 bit memory bus. RS remember that memory I/O pads require a certain die size. AMD seems to have had enough pad area aroung the core GPU to fit a 512 bit memory bus. all these are indicative of a massive die. remember GTX 285 had a 512 bit GDDR3 memory bus with a die size of 470 sq mm.

Since when do we rely on videocardz as the end all be all reputable source? Again, the card could be 384 or 512-bit but why are you making such definitive statements that it is 512-bit? I noticed that in all of your estimates so far, you are not accepting any compromises:

1) You think AMD will make a massive die of 470-480mm2
2) You think AMD will have overclocking headroom at least as good as the GTX780
3) You think AMD will suddenly go to 512-bit bus even though all previous rumours up to date called for a 384-bit bus.

In every case from GTX480/580/6970/Titan/780, every time on our forum we have members that tend to take the higher end estimates and these cards end up performing well under those estimates. Based on that, I'll take the same conservative approach with 9970 and assume one of those 3 will not be met by AMD.

 

[3DVagabond]

I wouldn't make such assumptions that 9970 oc could match 780 oc. HD6970 barely overclocked. 9970 could match after-market 780s but have 5-8% overclocking headroom. GTX780s overclock 20-30%.

It really depends on how far they have to push clocks at stock to get the performance they need. It would be nice if it could do Titan+ and have 30% O/C left. I think it will be a smaller die than GK110 though, and therefore won't have that much left in the tank.

 

[raghu78]

Since when do we rely on videocardz as the end all be all reputable source? Again, the card could be 384 or 512-bit but why are you making such definitive statements that it is 512-bit? I noticed that in all of your estimates so far, you are not accepting any compromises:

1) You think AMD will make a massive die of 470-480mm2
2) You think AMD will have overclocking headroom at least as good as the GTX780
3) You think AMD will suddenly go to 512-bit bus even though all previous rumours up to date called for a 384-bit bus.

In every case from GTX480/580/6970/Titan/780, every time on our forum we have members that tend to take the higher end estimates and these cards end up performing well under those estimates. Based on that, I'll take the same conservative approach with 9970 and assume one of those 3 will not be met by AMD.

sushiwarrior was already hinting at a surprise with regards to memory layout.

http://hardforum.com/showpost.php?p=1040092133&postcount=143

RS you talk as of 1.25 - 1.3 Ghz is the exclusive preserve of GK110. every high end and mid range TSMC 28nm chip from Nvidia and AMD has hit those clocks at max OC. Tahiti, Pitcairn, GK104, GK110 have all shown they can hit those clocks with voltage control and good cooling . so why is Hawaii not supposed to be able to do it ?

A question to you - why do you always look at the past ? The past is not always the predictor of the future. If thats the case AMD would never have been able to create Athlon K7 and Athlon 64 K8. You are adamant in your views that Nvidia is unbeatable and this after owning an AMD card. If you don't have faith in AMD atleast lets wait till the results come out.

 

[Cookie Monster]

Hmm, the only downside I can see is that even if the new AMD GPU beats nVIDIA's GK110 or worse just stay competitive with it.. maxwell is just around the corner which also happens to be nVIDIA's next new architecture. Well maybe completely revamped is a better description. Sort of like the "tock" if we use look at it from an "intel" perspective.

But I think it will be good times to go buy video cards since the prices on titans/GTX780s will defiantly start coming down along with a swathe of new goodies from both camps.