Radeon 7970 Specifications & Release Date slide leak?

[Will Robinson]

wrote tviceman:

It's not anything like Nvidia's cores though. Nvidia's gtx580 512 cores outperform AMD's 1536 Cayman cores. Simply comparing them 1 to 1 is like comparing a watermelon to a strawberry.

So NVDA's vastly more expensive GTX580 cores outperform the HD6970's.
Amazing...if only I'd known..

 

[tincart]

It's not anything like Nvidia's cores though. Nvidia's gtx580 512 cores outperform AMD's 1536 Cayman cores. Simply comparing them 1 to 1 is like comparing a watermelon to a strawberry.

He wasn't talking about performance, he was talking about utilization efficiency.

 

[tviceman]

He wasn't talking about performance, he was talking about utilization efficiency.

See this does not make sense to me either. Is there substantive proof that AMD GPU's are under-utilized? Are there screenshots showing AMD GPU's stuck at 80% GPU usage in BF3 while Nvidia GPU's are at 99% usage?

 

[tviceman]

wrote tviceman:So NVDA's vastly more expensive GTX580 cores outperform the HD6970's.
Amazing...if only I'd known..

Thanks for continuing to troll. FYI you are now happily on my ignore list. When and if you have something pertinent to say with what people are talking about then hell may have froze over.

 

[jvroig]

Will Robinson, tviceman,

Cut it out guys.

Will Robinson, your remark is completely off-base, and I am hard-pressed to interpret it as nothing more than an inflammatory remark meant to fuel more fanboy wars.

tviceman, this isn't your facebook wall. We don't care who you "un-friend" or put on your ignore list.

I don't care about either camp you think you are fighting against. Rules are rules. No inflammatory posts, no personal attacks. You are both guilty.

Moderator jvroig

 

[(sic)Klown12]

See this does not make sense to me either. Is there substantive proof that AMD GPU's are under-utilized? Are there screenshots showing AMD GPU's stuck at 80% GPU usage in BF3 while Nvidia GPU's are at 99% usage?

During the Cayman release, AMD themselves admitted that their VLIW-5 architecture was only running somewhere around 80% utilization on most graphic workloads. The shift to VLIW-4 wouldn't completely alleviate that under utilization either.

This page from Anadtech's own review of Cayman mentions this:
http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/4

 

[Rifter]

sure didn take the trolls long to screw this thread up.

So what exactly makes this "official"?

 

[GoStumpy]

Am I just a n00b, or am I the first one to notice GDDR3 instead of GDDR5 vram?

 

[toyota]

Am I just a n00b, or am I the first one to notice GDDR3 instead of GDDR5 vram?

it says GDDR5 on the slide so the OP just listed it wrong.

 

[videoclone]

I'm going to make a AMD HD7970 officially official AMD endorsed, 100% accurate, the real deal thread. lol

Until the cards come out or we get real information nothing is Offical yet me thinks

 

[Arachnotronic]

Typo on the GDDR5 on my part, sorry!

 

[Arkadrel]

OP post is obviously a really bad made fake.

 

[Silverforce11]

See this does not make sense to me either. Is there substantive proof that AMD GPU's are under-utilized? Are there screenshots showing AMD GPU's stuck at 80% GPU usage in BF3 while Nvidia GPU's are at 99% usage?

http://www.anandtech.com/show/4061/amds-radeon-hd-6970-radeon-hd-6950/4

They are operating at 3.4/5 for VLIW5 and similar for VLIW4. I recall seeing a figure of 3.2/4 for VLIW4 somewhere but can't find the link. Even if its just 1/5 its measured by software as being in use, so 99% GPU load from one game may vary in load to a different game etc. Bitcoin mining are so efficient on AMD hardware is due to their Instruction Parallelism, taking full use of the SP. Without proper coding, games run poorly on AMD architecture which is why IMO, TWIMTBP hurts AMD since they do not get the same level of developer relation to optimize the game engine. Whereas AMD evolve typically runs well on both NV and AMD gpus, because NV's architecture cannot be messed up (as bad) by poor coding.

Thus, GCN is an evolution to match NV's architecture in utilization.

Your example of VLIW4/5 not matching Fermi cores, exactly. Gtx580 has 512 running at much higher frequency (1600-1700?). Cayman's 1536 clocked at 880mhz, with 3.2/4 in optimized games = ~1228 @ 880mhz = 880mhz/1700 = 0.51 clock ratio = ~635 cuda cores (but also factor in ROP/texture and prims/clk). Its clear some games perform very well on AMD hardware, and a 6970 can match or beat a gtx580, but it tend not to happen often because i suspect very few games actually reach 80% utilization on Cayman's sp.

Edit: Also, depending on the instructions, Cayman's sp are severely crippled in output: transcendentals will bottleneck the sps hard whereas it has no detrimental impact on NV's cuda cores. GCN's SP fixes this by making each SP more flexible to handle every instruction required (at a cost of transistor/die area).

 

[RussianSensation]

http://4.bp.blogspot.com/-lmM_C5I4lQk/Tuj2x5x2ryI/AAAAAAAABh4/h0P2cr-FyU0/s1600/7970.png

Shaders: 2048
Compute Power: 3.5TFLOPs (SP)
Clock speed: 925MHz
ROPs: 32
Frame Buffer: 3GB GDDR5
Memory Bus Width: 384 bit
Memory Speed: 5.5Gbps

Absolute beastly. I think I'll take two, please (once my AMD stock recovers from the shitter, that is...)

That looks insanely underwhelming for a brand new architecture unless there is some magic wiZZard hidden inside.

HD6970 Compute Performance = 1536 SPs x 880mhz x 2 Ops / clock = 2.70 TFlops

Let's not forget that Cayman already improved ROPs from Cypress.

"The 32 ROPs (the same as Cypress) have been tweaked to speed up processing of certain types of values. In the case of both signed and unsigned normalized INT16s, these operations are now 2x faster. Meanwhile FP32 operations are now 2x to 4x faster depending on the scenario. Finally, similar to shader read ops for compute purposes, ROP write ops for graphics purposes can be coalesced, improving performance by requiring fewer operations."

Sure, they might improve the ROPs more, but that's not enough. The only way too boost pixel fill-rate significantly without blowing GPU clock speeds to 1200-1300mhz is to add more ROPs. 32 ROPs is not enough for a next generation GPU.

Also, if we look at overall theoretical computation performance, it's only up 30% (3.5 Tflops vs. 2.7). GCN is supposed to have a very significant increase in compute performance. The difference between HD5850 and HD5870 is about the same as the "leaked" specs show the difference to be between 5870/6970 and HD7970.

HD5970 Compute Performance = 1600 SPs x 850mhz x 2 Ops / clock = 2.72 TFlops

Surely, AMD isn't serious? Of course they aren't, BECAUSE THOSE SPECS ARE FAKE!!

Mistakes on the slides / questionable things
- Those slides have incorrect Compute Power figure relative to the listed specs.2048 SPs x 925 mhz clock speed x 2 Ops / clock cycle = 3.789 TFlops.
- They list 1x DVI port instead of 2x DVI ports. Questionable given how popular DVI still remains.
- There is an extra space between 32 CUs / 2048 ALUs and there isn't an extra space between 32 ROPs / 128 TMUs.

Photoshop Fail.

 

[Arkadrel]

Let's not forget that Cayman already improved ROPs from Cypress.

And its still the main bottleneck, for the cards.
They need to fix their ROP issue for new gen of cards, so I doubt they ll have 32 ROPs on them, they need more (or they ll end up even more bottlenecked than now by them).


Question is are we being trolled by OP?

To me it reads: *sarcasm* absoulutely beastly (op doesn think so).
I think I'll take two, please (once my AMD stock recovers from the shitter, that is...)

So hes saying hes pro nvidia, and hates amd because he lost money on stock speculations.

 

[Silverforce11]

I see 32 ROPs in leaks, its automatically unbelievable.

384 bit bus (confirmed), hence 48 ROPs minimum.

I suspect 48 ROPs would still be the bottleneck unless they voodoo magic those ROPs.
Just speculating, but 2048 1D shader is nearly 80% more effective than Cayman's 1536 (utilization factor). 32 ROPs on Cayman is already limiting. A 50% increase to 48.. still going to be limiting, prolly more-so.

 

[Arachnotronic]

And its still the main bottleneck, for the cards.
They need to fix their ROP issue for new gen of cards, so I doubt they ll have 32 ROPs on them, they need more (or they ll end up even more bottlenecked than now by them).


Question is are we being trolled by OP?

To me it reads: *sarcasm* absoulutely beastly (op doesn think so).
I think I'll take two, please (once my AMD stock recovers from the shitter, that is...)

So hes saying hes pro nvidia, and hates amd because he lost money on stock speculations.

Negative. I do plan on buying 7970s when they launch. I just sold my 580 to a forum member to pay for my first one. I'm counting on my AMD stock speculation to pay for the second.

 

[RussianSensation]

And its still the main bottleneck, for the cards.
They need to fix their ROP issue for new gen of cards, so I doubt they ll have 32 ROPs on them, they need more (or they ll end up even more bottlenecked than now by them).

My point was that AMD already sped up ROPs in the Cayman design by 2-4x. At this point it makes sense to increase them since you already sped them up significantly. That's not even getting into the technicalities of how 32 ROPs don't work with 384-bit controller.

Besides, I won't settle for anything less than 1 Ghz GPU clocks and 48 ROPs at minimum.

 

[Mistwalker]

So what exactly makes this "official"?

Nothing.

OP, can you please change the thread title? The info in these slides has been shown to be incorrect, and unless it's direct from AMD there shouldn't be an "Official Specifications" topic without citing the source...it's confusing at best, misleading at worst.

 

[AtenRa]

HD6970 Compute Performance = 1536 SPs x 880mhz x 2 Ops / clock = 2.70 TFlops


HD5970 Compute Performance = 1600 SPs x 850mhz x 2 Ops / clock = 2.72 TFlops

Surely, AMD isn't serious? Of course they aren't, BECAUSE THOSE SPECS ARE FAKE!!

2048 SPs x 925 mhz clock speed x 2 Ops / clock cycle = 3.789 TFlops.

Photoshop Fail.

Only IF one GCN ALU = one Cypress/Cayman SP.
Are we sure that each Vector ALU + Scalar perform the same as one VLIW-5/4 SP ??

NV Cuda cores are less in the GTX580 and produce less TFlops than HD6970 but we cannot directly compare them

 

[UJ09]

With this underwhelming spec AMD wants us to pay $500 for 7970? I don't really trust this so-called official spec leak.

 

[Riek]

It is already known AMd would release fake slides.. They did the same in the 69 release and it was expected to be worse for this release... So prepare for magnitudes of bs with a grain of truth.

PS: for the 69xx we only knew the correct shader count less than a week before release and at that point it wasn't even confirmed or widely accepted.

 

[RussianSensation]

Only IF one GCN ALU = one Cypress/Cayman SP.
Are we sure that each Vector ALU + Scalar perform the same as one VLIW-5/4 SP ??

NV Cuda cores are less in the GTX580 and produce less TFlops than HD6970 but we cannot directly compare them

I am not saying we should be comparing Tflops in deriving gaming performance. But the math of 2048 ALUs x 925 mhz with 2 Ops / clk (from the slides) doesn't work making me suspect those slides are fake. How do you get 3.5 TFlops single precision from that?

Think about it. Are those 2048 ALUs magically losing 7.5% performance somewhere? (i.e., 2048 x 925 x 2 ops/clock * 92.5% = 3.5 Tflops)

FLOPS = SPs x ALU clocks * Ops / clock. There isn't really any hidden trick here.

Even for NV, to calculate total compute, you just take SPs x ALU clock * ops / clock.

For instance, GTX580 = 512 SPs x (shader ALUs are double-clocked from GPU speed of 772mhz to 1544MHz) * 2 ops / clock = 1.581 Tflops

PS: for the 69xx we only knew the correct shader count less than a week before release and at that point it wasn't even confirmed or widely accepted.

Remember the original HD6970 was supposed to have 1920 SPs? :whiste:

 

[AtenRa]

Think about it. Are those 2048 ALUs magically losing 7.5% performance

It says 3,5TFlops (IEEE754)

IEEE 754-2008

http://www.validlab.com/goldberg/paper.pdf IEEE 754 on Page 189

Dont ask me to calculate it ehehehehe

 

[MentalIlness]

http://www.xbitlabs.com/news/graphi...32_AMD_We_Are_Shipping_Our_28nm_GPUs_Now.html

AMD: We Are Shipping Our 28nm GPUs Now.

Advanced Micro Devices said on Tuesday that it had begun to ship its next-generation graphics processing units (GPUs) made using 28nm fabrication process at Taiwan Semiconductor Manufacturing Company's fabs for revenue. The company is projected to formally unveil its future graphics products at the Consumer Electronics Show early in 2012.
"We are ramping 28nm [products] with TSMC in Taiwan and shipping the products here and now. We are very excited about the products," said Rory Read, chief executive officer of AMD, during IT Supply Chain conference organized by Raymond James.
For AMD, which promised to be the first to market with 28nm GPUs for a number of times this year, it is a good news to start not only producing such graphics processors at TSMC, but also to initiate revenue shipments of the parts. While we do not know for sure, it is very likely that PC makers will be able to start shipping systems featuring Radeon HD 7000-series "Southern Islands" chips shortly after CES next month.
The next generation of standalone graphics processors from AMD and Nvidia will feature considerably improved performance per watt and while they will still belong to the DirectX 11 generation, they will bring support for numerous new features and capabilities only possible with tangibly improved performance and refined architecture.


In fact, the first unofficial photos of the next-generation AMD Radeon HD graphics cards have already emerged. The new boards depictured are marked as “Tahiti” and are believed to belong to the top-of-the-range single-chip next-generation Radeon HD family. Based on AMD’s current nomenclature, the new graphics adapters will be called the Radeon HD 7900-series.
The forthcoming AMD Radeon HD 7900 “Tahiti” graphics cards are very long and carry twelve GDDR5 memory chips, which means that they use 384-bit memory bus. The novelties have two 8-pin PCI Express power connectors, something which points to rather high power consumption. Typically, photos of unreleased graphics adapters emerge shortly before their launch.
AMD’s Tahiti graphics processor belongs to the highly-anticipated Southern Islands family of chips. The new SI family will use new graphics architecture, feature a number of innovations as well as will be made using 28nm process technology.
AMD did not comment on the news-story.