HD 7000 Series Specs Leaked

[Death_Lord]

Originally Posted by Redemption80

I heard 20% faster on average and very expensive to start with.

Makes sense though, it's not like Nvidia are releasing anything soon, so AMD can do this.

What you said doesn't make much sense, why they would make a product that would only sell for 4 months? they want to sell for the whole time the generation lasts, so the best is to make something very fast that can compete with what Nvidia has to offer.

[Redemption80]

Why would it only sell for 4 months?
It's still going to sell very well after Nvidia release their cards, but they will likely have to change the pricing, but for the first few months they are under no pressure price wise.

It's best to make something realistic that beats the present competition, aiming too high, or thinking about what the competition will bring out in the future is only going to lose them money, or have them selling very expensive cards right now.

Seems to me that AMD just wan't to comfortably beat the competition, but without using excessive amount of power.

[Noisiv]

From Hardocp:

I like how OBR is such a dickhead. He has to be the least self-aware person on the internet. There are constant references to random people that no one has probably ever heard of.

Every post of his is basically "Wasily! I call you out! See this benchmarks? I am the one holding the pan with all the benchmarks before you. I predict AMD is fail again on 22/12/11 and see it is true. Wasily doesn't even know and Chekhov doesn't either. Here are pictures I made in my house with show the true meaning of performance and you can see it is not the greatest. One day Wasily you may be as good as me but actually never because you are worst. Bye!"




Sorry for OT, I thought it was funny enough to share

[Xzibit]

According to our sources, the Radeon HD 7970 is about 30 percent faster in 3DMark when compared to the Radeon HD 6970.

Fudzilla

[Noisiv]

not good

[Xzibit]

Originally Posted by Noisiv

not good

So this is Fudzilla's source

[Noisiv]

Chiphell? yeah and OBR

and then they calculated that 30% on 6970 should put 7970 close to 6990

[crap daddy]

It's totally plausible.

[Noisiv]

Voodoopower Ratings which awards dual cards with -15% of their performance,
because of the microstuttering and non-working games should help, but still...

[Krasus]

Originally Posted by MfA

huh how did i forget to post guru3d..

7900 Series The World's Most Powerful and Advanced Graphics

wait a sec. 1 ghz oc headroom means you can improve clocks about 1ghz? Did I misunderstand? whatta..

@redemption80 thanks.

[Redemption80]

No, it means there is headroom to hit 1ghz, and beyond.

[Isbre]

You got to compare the performance also to "Cypress" since it's the first high perf. chip on 40nm. So then we are looking at about 50% avg perf. gain over previous gen. Perf/watts is not much improved for either "Cypress" or "Cayman" atm it seeems, but same thing we saw when first RV600 chip this is likely only result of that, or maybe also combine with early leaky process.

[Beast]

If the HD 7970 is only 30% faster than an HD 6970 it would make it 16% faster than GTX 580.

[BlackZero]

Originally Posted by Beast

If the HD 7970 is only 30% faster than an HD 6970 it would make it 16% faster than GTX 580.

Going by AMD's slides it's closer to 60-70% performance in games and around 3x tessellation compared to HD 6970 and according to techpowerup/computerbase.de these figures are genuine.




http://www.techpowerup.com/156887/Ra...s-Surface.html


Also AMD yesterday decided to give futuremark direct access to catalyst driver tessellation monitoring so they must be really confident in these card's tessellation performance, futuremark have also released a 3dmark11 update incorporating this fix.

[Xzibit]

1, Tahiti or RADEON HD 7900 with dual ACE or dual geometry engine.
2, under the jurisdiction of each ACE 16 GNC QUAD SIMD 2, under the jurisdiction of each ACE 16 GNC QUAD SIMD
3, a total of 8 Render back-end, or 32 week period to complete the color OP, 128 个 Z-Stencil OP. 3, a total of 8 Render back-end, or 32 week period to complete the color OP, 128 个 Z-Stencil OP.
4, two read-write cache is 768KiB; every 4 CU share 16KiB L1 instruction cache 32 KiB L1 data cache. 4, two read-write cache is 768KiB; every 4 CU share 16KiB L1 instruction cache 32 KiB L1 data cache.
5,384-bit GDDR5 memory bus bandwidth of up to 264GiB / s 5,384-bit GDDR5 memory bus bandwidth of up to 264GiB / s
6, using PCI-Express 3.0 bus 6, using PCI-Express 3.0 bus
7, 4.3 billion transistors. 7, 4.3 billion transistors.


Support DX11.1 Support DX11.1

Similar megatexture a special presentation of Partially Resident Textures (PRT) technology, but this is a hardware implementation, including mapping, filtering and other operations, texture sizes up to 32 TiB (128KiB ^ 3 * 128-bit). Similar megatexture a special presentation of Partially Resident Textures (PRT) technology, but this is a hardware implementation, including mapping, filtering and other operations, texture sizes up to 32 TiB (128KiB ^ 3 * 128-bit).

Tessellation part is clearly up to Gen8 (6900) four times the performance Tessellation part is clearly up to Gen8 (6900) four times the performance

Anisotropic Filtering is basically and 6900, but the new improved high-frequency texture flickering problems. Anisotropic Filtering is basically and 6900, but the new improved high-frequency texture flickering problems.

Double-precision performance of 947 GFLOPS, to achieve the DDR, SRAM full ECC protection. Double-precision performance of 947 GFLOPS, to achieve the DDR, SRAM full ECC protection.

Multimedia instruction, the introduction of QSAD, MQSAD, each cycle can be achieved for each CU 256 pixels perform SAD operation, the entire down the equivalent of 7.6 Tera pixels per second. Multimedia instruction, the introduction of QSAD, MQSAD, each cycle can be achieved for each CU 256 pixels perform SAD operation, the entire down the equivalent of 7.6 Tera pixels per second.

Software, WinZip 16.5 will offer OpenCL support (eg AES), which is also described in the slide Software, WinZip 16.5 will offer OpenCL support (eg AES), which are also described in the slide

PowerTune in PowerTune in

http://translate.google.ru/translate...03470-1-1.html

And here are some specs from store

http://www.salland.eu/product/108271...7970-3gd5.html

[Rich_Guy]

^ Core clocks wrong there, as its only around 900/925mhz now, unless they've changed it again

[S†v0r]

Looking good, can't wait for reviews!

[Rich_Guy]

According to Charlie, the 7800's are not just die shrunk 6900's, they are also GCN's (the new arc).

http://semiaccurate.com/2011/12/16/l...ries-rebrands/

[Noisiv]

Originally Posted by Rich_Guy

^ Core clocks wrong there, as its only around 900/925mhz now, unless they've changed it again

hopefully that's Asus pre-overclocked...

http://www.salland.eu/product/108271...7970-3gd5.html

Measured a total of 12 games, including a sub-run software, but can not play 3, a mapping error.
Name of the game is not being said, the resolution is 1920x1200 all max
12%
35%
20%
n / a
n / a
-8%
17%
31% (mapping error)
19%
-5%
33%
n / a

http://translate.google.com/translat...03805-1-1.html

[Roush9]

I just hope the 2GB 6950s drop dramatically in price so i can upgrade my 6870.

[Redemption80]

Not looking likely at the moment, might be different elsewhere, but a UK shop here says there are no immediate plans to drop the prices of the 6900 cards, and looking at the leaked prices of the 7900 cards, that makes sense.

Unfortunately i don't think anything is going to drop in price until AMD has some competition for the 7k cards.

[Isbre]

I really hope the compute capability and especially f@h perf is gonna be good with this new arch.

If not i will be waiting for Nvidia to see what they have to offer with their upcoming products. Will do this anyway for AMD to drop it's prices.

[kpo6969]

Originally Posted by Rich_Guy

According to Charlie, the 7800's are not just die shrunk 6900's, they are also GCN's (the new arc).

http://semiaccurate.com/2011/12/16/l...ries-rebrands/

So I would take this as that the 7870 would not just be a Cayman shrink but the new deal?
Hope so.

[Laykun]

Originally Posted by jhelsas

If this true? 3.8 Tflops?
it looks like a dream, 3.8 Tflops of OpenCL performance for number crunching.

IF this is REALY true, it means that i could have a cluster grade performance in a single-GPU solution. This is simply ridiculous.

I remember when it was released the 1st Nvidia Tesla, and it had 0.9 Tflops, and the whole pc with 4x gpus reached 3.6.....
It wasn't such a long time ago....

I have to learn CUDA or OpenCL soon.....

BTW, what's the GTX 580 CUDA or OpenCL performance today? Do anyone have this number?

3.8Tflops on Single precision, which is kind of useless for anything worth while. I imagine the double precision performance is cut in half/quaters artificially so ends up being something like 200Mflops in the end, forcing you to buy a high-end version of the card sold for doing compute tasks. It's happened with all previous cards.

[Lane]