He has droped some bits and pieces before, I guess we'll see how accurate it is in a few months time 🙂
-
We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.
Sapphire AMD HD 9970 News
Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
He has droped some bits and pieces before, I guess we'll see how accurate it is in a few months time 🙂
iMacmatician
Member
Performance projections? Maxwell could be on 28 nm first, later 20 nm. The "big" Maxwell is likely to come out on 20 nm, which would be needed give it the 2x DP FLOPS over "big" Kepler that their roadmap says. (Or are you talking about another type of projection?)
Apple has not used 28 nm (or 40 nm for that matter) for at least their Ax SoCs so far. 20 nm looks to be very different.
Maybe Curacao parts don't quite beat the 7970 GE and so that leaves Tahiti in-between Hawaii (I'm assuming this is the high-end Volcanic Islands chip) and Curacao.
On "incorrect" codenames: From what I've seen in the past year, many if not all of the codenames may be real, it's just that they don't often seem to be matched to the proper chips in the rumors.
Stuka87
Diamond Member
Apple did not buy any 40 or 28nm chips from TSMC. So what he said is kind of invalid. Apple has already made it clear they will be dropping Samsung for all chip manufacturing. And will be moving their A-series chips to TSMC. I have no doubt that Apple will make sure they are first to receive 20nm chips.
blackened23
Diamond Member
^Correct.
Look, let me state this as a preface. I'm sure the 9790 will be a great. IT is also rumored to have a 512 bit bus, which hasn't been seen in a LONG time - if that is true, the card should be a fantastic card with hopefully a value price in comparison to NV's offerings. I'm pretty confident that AMD has learned many hard lessons with the 7970, which I thought was a *great* card but had a lot of points which made it a less attractive buy than GTX 600 series at launch. That changed a few months afterwards (in terms of the value and price/perf) and this all ties in with lessons learned. So I am sure it will be a great card with that in mind.
However, with that said... I will eat my hat if 9790 is 20nm, the timeline is too far off - TSMC is just now in the process of finalizing their fabs and doing prototype runs, but they are a LONG way off from doing any volume production - in fact, when asked at an investor meeting it was firmly pegged at late Q1 2014 or sometime in Q2 2014 for volume production. Volume production is when the beginnings of any product using the process can begin, so it will definitely not happen in 2013 - furthermore, this is an optimistic estimate. You can imagine delays can and will happen, as has always been the case for TSMC - and one has to mention the fact that Intel's 22nm FinFET process is actually superior to TSMC's 20nm. Intel long ago invested in double patterning and FinFET while TSMC has not. They're just now discussing it - so their process is actually substantially worse than intel's. Their 28nm is comparable to Intel's 32nm, just as a point of reference.
This is all beside the fact that Apple, as the above poster mentioned, did not have TSMC produce anything up to this point on 28 or 40nm, it was done by Samsung. That is changing, and Apple and Qualcomm - sad to say they have a lot of cash to throw at TSMC which nvidia and AMD do not. So they will get first grabs at any volume production, which to me would indicate we won't see 20nm GPUs until 2H 2014, plus or minus a few months.
Last edited:
Fastx
Senior member
@Stuka87 & Imacmatican
Ok guys point taken on Apple and I did assume (not knowing as fact) Apple was getting their previous chips from TMSC I should have researched that, so point taken and noted. Also just for the record for anyone reading this post now Hurleybird did NOT allude to Apple (only Qcom) in his quote I posted so I just want to be clear on that and not misrepresent what Hurleybird posted for guys reading this post now.
Ok guys point taken on Apple and I did assume (not knowing as fact) Apple was getting their previous chips from TMSC I should have researched that, so point taken and noted. Also just for the record for anyone reading this post now Hurleybird did NOT allude to Apple (only Qcom) in his quote I posted so I just want to be clear on that and not misrepresent what Hurleybird posted for guys reading this post now.
sushiwarrior
Senior member
I'm still confused at 20nm... TSMC + release date says absolutely no, but the specs increase + TDP say absolutely yes...
boxleitnerb
Platinum Member
^ you mean the 9970? I would think that let's say 2560 SPs@ 1 GHz should be possible at 230-250W. The 28nm process is very mature now and if you lower the voltage quite a bit, especially compared to the 7970 GE, it should be possible to bring consumption down a lot.
Or 20nm risk production. I mean who says it's impossible? Just because it has not been done before doesn't mean it couldn't be done.
Or 20nm risk production. I mean who says it's impossible? Just because it has not been done before doesn't mean it couldn't be done.
Last edited:
boxleitnerb
Platinum Member
How do you figure it's the ROPs/IMC?
sushiwarrior
Senior member
Testing voltage/clock speed combinations with different tests and noting exactly when the crashes happen.
I'm talking about 100% of resources on the card used at the same time efficiently (maximum power draw) while still maintaining 100% stability in all loads with same voltage.
For 2560 to draw less than 2048 (attached to less basic graphics resources) would require HPm, or Moderate Binning, or moving the goalposts (like the fake 375w TDP on the 7990 that's just a throttle point, bad benchmark comparisons), or just comparing to a low ASIC score bin or badly designed card.
Any combination of the aforementioned scenarios would also work.
Depends on what you mean with 7970. Launch benches? 7970Ghz "reference" cards sent to reviewers? With a mature design on a mature node, turbo and better cooling, I wouldn't be surprised if comparing launch benches, even with clocks being close. Recent 1050mhz 7970s with decent aftermarket cooling are getting pretty close to 200W under load, though.
boxleitnerb
Platinum Member
I don't think that is really conclusive. After all this is one card out of millions.
@sushiwarrior:
Want to play a game of hot or cold? 😀
40 CUs (2560SP, 160 TMUs)
48 ROPs
7 Gbps memory
Power consumption (not TDP!) between 7970 and 7970 GHz
Performance 7970 GHz +25%
Nvidia-like turbo
MSRP $549-599
How many did I get right?
Last edited:
I don't get it. Why launch next Radeon series with two architectures, ie Sea Islands and Volcanic Islands? Is it cost?
Curacao is from SI, Hawaii is from VI.
With VI you get transition like this (from Jim's interview).
Tahiti -> Hawaii
Pitcairn -> Iceland
Oh and Dave Baumann said Hainan is notebook chip.
Curacao is from SI, Hawaii is from VI.
With VI you get transition like this (from Jim's interview).
Tahiti -> Hawaii
Pitcairn -> Iceland
Oh and Dave Baumann said Hainan is notebook chip.
Curacao was canned. Hawaii is the new Tahiti successor. It was actually the successor of Curacao, but AMD was speed up the developement. Vesuvius is a mobile GPU. Tonga is the Pitcairn successor.
I think the HD 9000 lineup is Hainan->Oland->Bonaire->Tonga->Hawaii without a dual GPU option, because the top Hawaii VGA will be ~1000 dollar.
Hainan, Oland and Bonaire are Sea Islands arch. Hawaii is Volcanic Islands arch. Pitcairn's successor from VI is Iceland (Jim's interview).
boxleitnerb
Platinum Member
I'm still under the impression, that Sea Islands isn't a specific architecture but a collection of GPUs for 2013. A marketing term if you will. Dave Baumann mentioned something like this in the beyond3d forums, and Semiaccurate reported on it as well. Bonaire is Sea Islands, for example, but it's "tech level" is still GCN 1.0. This was discussed over at b3d as well.
Yeah this, except wasn't Bonaire GCN 1.1 already? If you want to refer to the architecture, GCN + a number is what you want, the code names are used for product families which don't necessarily use the same architecture.
boxleitnerb
Platinum Member
I guess AMD could make a seizable jump if they use Pitcairn as the base and go from there. Tahiti is just too inefficient compared to Pitcairn.
I wonder if any of those 7 cooling solutions Sapphire is working on will sport a better quality fan that doesn't die within 6 months of use. My HD6950 that's Sapphire branded had the fan crap itself inside 6 months, I RMA'd it, and the one I received back started making the same noises inside of 30 days. Meanwhile my other one by another brand (I forget which) runs just fine.
It's enough to make me wary of Sapphire's products despite them often having the best price.
It's enough to make me wary of Sapphire's products despite them often having the best price.
RussianSensation
Elite Member
Are you getting 90% scaling with 2 6990s though?
Any reason you want to wait for HD9970 for this title? NV runs this game very well too. I'd consider dual 780s for you instead of waiting until October-November.
Side-note: HD7970GE is 100% faster in Arma 3 over HD6970 and easily hangs with GTX590/HD6990 - last gen's dual-GPU flagships! So much for people saying GCN was a marginal upgrade over 6970. In many new titles coming out, 7970GE is crushing the 6970 by 60-90%. GTX780 is ~ 2.8x faster than HD5870. That means after-market 780 will be > 3x faster. :awe:
In your estimate, shader & texture performance increase 25%, but ROPs go up 50%. How can the performance increase only 25% then? Does not compute. There is also a rumour of the card doubling ACE engines to 4 and adding a 3rd geometry engine. You assigned no value for those. With those specs, it would be way faster than 25% considering GCN is pixel fill-rate limited, assuming the clock rate is still 1050mhz.
Personally, I think your specs sound too good to be true. Take an HD7870 and compare to what you listed.
1000mhz GPU vs. 1050mhz 9970
1280 SPs vs. 2560 SPs (2x)
80 TMUs vs. 160 TMUs (2x)
32 ROPs vs. 48 ROPs (50%)
154 GB/sec vs. 336 GB/sec memory bandwidth (2.18x)
Right now in games where HD7970 is pixel fill-rate limited, the shaders and TMUs are underutilized. If you look at 7950 stock vs. overclocked, once that ROP bottleneck is lifted, an overclocked 7950 flies away from 7870 but at stock speeds they are very close. If you lift the ROP bottleneck in Tahiti XT and add all those units you listed, you could be looking at 85-90% more performance over the 7870. I don't know, sounds too optimistic to me since you also list power consumption below 7970GE.
If AMD doubles the ACEs, the compute performance of the card will increase too.
Since so many modern games are using compute, this card will be even faster than the Titan in games like Tomb Raider, Grid 2, COH2, Hitman Absolution, etc. AMD will charge $549 for such a card? Also sounds too good to be true.
We also have real world testing of HD7970 overclocked that puts your 25% faster claim into question based on the specs you listed.
1328mhz 7970 with 6.818Ghz memory gives us 26% more shader and texture performance. # of ROPs is fixed at 32, but look at the performance against a stock 7970 GE:
21% faster in Batman
19% faster in BF3
20% faster in Crysis 2
If you estimate performance to increase just 25% but real world testing shows that performance goes up 19-21% with ONLY a 26% increase in shaders & texures and just a 13.6% increase in memory bandwidth, that means according to your estimate adding 50% more ROPs, doubling ACEs and adding a 3rd geometry engine and even more memory bandwidth will only increase performance just 5% more??!!! The card you just listed would mop the floor with a 7970GE. It sounds too good to be true unless AMD drops clock speeds and/or there are no changes to ACEs/geometry engines.
Last edited:
sushiwarrior
Senior member
Boxleitnerb is close-ish... but not to the flagship.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 25K
-
-
-
