Radeon 8870/8850 spec reveal? Oland chip chart. Also: GTX 780 info?

GoldenTiger

GoldenTiger

Fully [H]
Joined
Dec 2, 2004
Messages
29,669
http://www.techpowerup.com/172312/AMD-quot-Oland-quot-Radeon-HD-8800-Series-SKUs-Unveiled.html

Chart with estimated flops/etc. there. Educated guesses in comments are that it sports 256-bit, 1792 sp's, 112 tmu's, 32 rop's, for the 8870 model. The chart pins the MSRP @ $279, with the Oland chip powering it.

8870chart.jpg


Venus, the next-gen high-end, will be forthcoming.
AMD682B.1 = "VENUS LE" AMD6823.4 = "VENUS PRO" AMD6821.1 = "VENUS XT" AMD6820.2 = "VENUS XTX" Venus XTX = 2x
Click to expand...

No release date info for either series.

In related info on the nVidia front...

GoldenTiger said:
Current rumors are that more powerful cards will start arriving such as the GTX 780 in the first few months of 2013, with the latest rumored specs placing it at 15 SMX units (a GTX 680 has eight) + 384-bit bus width (GTX 680 has 256-bit). Add in higher clocks to that and more ROP's (48 would be the likely # on it vs. the 680's 32) and efficiency optimizations at the hardware level, and you're going to see another huge leap for performance. For the lazy... 15 SMX units would be 2880 stream processor CUDA cores.

GK110/100 is already in usable form and arriving in quantity at their high-end customers as Tesla cards now: http://www.hpcwire.com/hpcwire/2012...h_of_kepler_gpus_for_titan_supercomputer.html

Likely specs appear here from an nVidia presentation: http://www.sweclockers.com/nyhet/15816-nvidia-gk110-pa-vag-till-geforce-och-konsumentmarknaden (English re-post: http://www.legitreviews.com/news/14009/)
Click to expand...
 
Interresting, I wonder how well these cards will perform in multi-monitor gaming. Also, with Nvidia pushing out new AA techniques and adaptive Vsync, I am more interested in new features, not performance.
 
Looks ridiculous to me, a 21 % gain in transistors, and a 5% gain in clockspeed, on the same process node and basic design, and there's a 60-70% performance gap??? That seems highly unlikely.
 
Yoda4561 said:
Looks ridiculous to me, a 21 % gain in transistors, and a 5% gain in clockspeed, on the same process node and basic design, and there's a 60-70% performance gap??? That seems highly unlikely.
Click to expand...

Raw specs do not mean actual gaming performance.
 
I'd consider caution, as far as AMD leaked specs/benchies are concerned. Bulldozer, according to first specs/benchies was supposed to trump Sandy Bridge too. Until I see realy tests from [H], I won't believe the rumors.
 
MorgothPl said:
I'd consider caution, as far as AMD leaked specs/benchies are concerned. Bulldozer, according to first specs/benchies was supposed to trump Sandy Bridge too. Until I see realy tests from [H], I won't believe the rumors.
Click to expand...

Do not compare CPU's with GPU's, AMD has amazing GPU section...
 
Yoda4561 said:
Looks ridiculous to me, a 21 % gain in transistors, and a 5% gain in clockspeed, on the same process node and basic design, and there's a 60-70% performance gap??? That seems highly unlikely.
Click to expand...

the HD 7870 specs are wrong. HD 7870 has a performance of 2.56 TFLOPS at 1 Ghz.

http://www.hardocp.com/image.html?image=MTMzMDkwNDA2ODVLSVJpekFSNFZfMV84X2wuZ2lm

HD 8870 has 1792 shaders at 1.1 Ghz. performance of 1792 x 2 x 1100 = 3942400 MFLOPS or 3.94 TFLOPS. . Thats a 54% performance increase in compute performance. Also understand that 40% of the performance improvement comes from shader count increase. (1792 / 1280 = 1.4) . The remaining 14% comes from a 10% higher clock . (1.4 x 1.1 = 1.54)

Its quite realistic because the chip is identical to Pitcairn with 40% more shader processors. Front end resources are going to be identical - 2 rasterizers, 2 tesselators, 2 ACE. ROPs remain same at 32 and memory controller stays 256 bit with a bit of tweaking for running at 6 Ghz speeds. Definitely you can expect improvements to tesselators as AMD has normally been doing for the past three generations. With such performance and efficiency tweaks improving performance by 35 - 40% and matching a GTX 680 is not difficult.
 
Last edited:
Why do we have these specs first? Aren't they are not going to release the 8970 first? Or does AMD cater to the mainstream first then go to the Higher level?
 
Yeah, I forgot this was a 'leak'. Leaks like this make a person on the fence, like me, stop and reconsider for a while. I have a pair of 2GB 5870's xfired and was considering a 7970. So maybe I will wait for some more "leaks".
 
Well the HD 5870 was a HUGE leap forward compared to the HD 4870. I really do hope this is true.

And the 7870 had a MSRP of $349 when it was released, I believe that is what is usually used in those charts.
 
xIronCrossx said:
Well the HD 5870 was a HUGE leap forward compared to the HD 4870. I really do hope this is true.

And the 7870 had a MSRP of $349 when it was released, I believe that is what is usually used in those charts.
Click to expand...

5870 was also on 40nm vs 55nm.
This will be 28nm vs 28nm. No expectations here
 
ggplay said:
5870 was also on 40nm vs 55nm.
This will be 28nm vs 28nm. No expectations here
Click to expand...

Yes, this will be more like the 5870 to 6970, which was a %15 performance bump with higher power consumption.

Don't expect amazing things until the next process shrink.
 
defaultluser said:
Yes, this will be more like the 5870 to 6970, which was a %15 performance bump with higher power consumption.

Don't expect amazing things until the next process shrink.
Click to expand...

This. Seeing that TSMC missed 32nm completely and struggled with the 28nm transition, we may be a couple of years off from 22/20nm GPUs.

Meanwhile Intel is humming along nicely and is schedule to introduce 14nm Broadwell in 2014...
 
jorojr said:
This. Seeing that TSMC missed 32nm completely and struggled with the 28nm transition, we may be a couple of years off from 22/20nm GPUs.

Meanwhile Intel is humming along nicely and is schedule to introduce 14nm Broadwell in 2014...
Click to expand...

65nm-45nm-32nm-22nm is CPU

55nm-40nm-28nm is GPU

They tend to go down by a ratio of sqrt(2) so i wouldn't expect GPUs to have hit 32n,.
 
ggplay said:
65nm-45nm-32nm-22nm is CPU

55nm-40nm-28nm is GPU

They tend to go down by a ratio of sqrt(2) so i wouldn't expect GPUs to have hit 32n,.
Click to expand...

Except that large, multi-customer foundries like TSMC have a history of offering half-node steps. Here are some examples of the various half-nodes TSMC offered in the past:

AMD's half-steps

RV250: 150 nm (half step)
RV380: 130 nm
RV370: 110 nm (half step)
R520: 90 nm
R600: 80 nm (half step)
RV610: 65 nm
RV770 : 55nm (half step)

They have standardized on the full step starting at 55nm because 45nm was delayed (and thus largely ignored), and 32nm was canceled entirely due to cost. So, FOR NOW AT TSMC the nodes will follow full steps, but that's only because of cost reasons.
 
Originally Posted by defaultluser
Yes, this will be more like the 5870 to 6970, which was a %15 performance bump with higher power consumption.

Don't expect amazing things until the next process shrink.
Click to expand...

If it is only going to be 15% I might as well buy a 7970 now rather than waiting for the 8970. However, if this chart is correct, I do not understand how all of these improvements can only result in a 15% improvement. Gimme a little help here!!
 
hcforde said:
If it is only going to be 15% I might as well buy a 7970 now rather than waiting for the 8970. However, if this chart is correct, I do not understand how all of these improvements can only result in a 15% improvement. Gimme a little help here!!
Click to expand...

Yeah I dont think people are taking the die size increase into account, you can get 15% extra with a decent OC from 7 series, this chart shows massive gains from a roughly 50MHz increase so where does all the other power come from...

But AMD will need this type of improvement to compete with a 680/670 that will probably become the GTX 760 & 760 ti, as I would imagine GK110 will finally show up as the 780 and 770.

assuming any of this is true

I was personally aiming to buy a 7950 but this throws a wrench into my plans as basically the 8850 should be about the same for 199 if not better @ 1080p
 
atticus14 said:
Yeah I dont think people are taking the die size increase into account, you can get 15% extra with a decent OC from 7 series, this chart shows massive gains from a roughly 50MHz increase so where does all the other power come from...

But AMD will need this type of improvement to compete with a 680/670 that will probably become the GTX 760 & 760 ti, as I would imagine GK110 will finally show up as the 780 and 770.

assuming any of this is true

I was personally aiming to buy a 7950 but this throws a wrench into my plans as basically the 8850 should be about the same for 199 if not better @ 1080p
Click to expand...

The specs for HD 8870 look as follows - 1792 sp, 1.05 Ghz / 1.1 Ghz boost, 112 TMUs, 32 ROPs, 256 bit memory bus, 192 GB/s (6 Ghz GDDR5 ).

This chip is a HD 7950 with all the fat trimmed off. no 1/4 Double precision rather 1/16 double precision, no ECC, memory bus trimmed to 256 bit with 6 Ghz memory. Overall its a perf/watt optimized chip like the GTX 680. We know the HD 7950 can easily compete with the high end chips when overclocked. So it would not be difficult for the HD 8870 to compete with GTX 680.

You can expect performance tweaks to the tesselators and other parts of the HD 8870 chip. What you can expect is at 1080p HD 8870 will match HD 7950 (1.1 Ghz) while drawing less power. At 1440p though I believe the HD 7950 will be faster at the same clocks.

the HD 7950 OC is the best price perf high end card. clock for clock HD 7950 is 3 - 5% slower than HD 7970. So get a card like Sapphire HD 7950 950 mhz or Gigabyte HD 7950 Windforce and just push that chip with voltage tweaking to 1.15 Ghz. Unbeatable value for your money.
 
hcforde said:
If it is only going to be 15% I might as well buy a 7970 now rather than waiting for the 8970. However, if this chart is correct, I do not understand how all of these improvements can only result in a 15% improvement. Gimme a little help here!!
Click to expand...

Because for most modern high-resolution, high MSAA gaming, pixel and memory bandwidth are the most limiting factors. Take for example, the 7970 has 30% more compute performance over the 7950, but only 15% more pixel throughput and 10% more memory bandwidth:

http://www.hardocp.com/image.html?image=MTMyNzkyOTIzNnZNQnUwWGkyd2xfMV8zX2wuZ2lm

Guess which one the performance follows more closely?






HINT: it's not the compute performance!

http://www.techpowerup.com/reviews/Sapphire/HD_7950_Vapor-X/28.html

Why yes of course, (113/96 * 100) = %117 performance with the 7970 versus the 7950 across all resolutions (17% increase). Showing that theoretical compute improvements have little to no bearing on actual card performance in games (so long as there is *enough* performance to max the given memory/pixel throughput).

If you look carefully at the graph linked describing the new card's performance, pixel throughput is only up 10% and memory bandwidth is up 25%, so I'd guesstimate the performance increase is somewhere between those two.
 
Last edited:
raghu78
You can expect performance tweaks to the tesselators and other parts of the HD 8870 chip. What you can expect is at 1080p HD 8870 will match HD 7950 (1.1 Ghz) while drawing less power. At 1440p though I believe the HD 7950 will be faster at the same clocks.

the HD 7950 OC is the best price perf high end card. clock for clock HD 7950 is 3 - 5% slower than HD 7970.
Click to expand...


defaultuser
Why yes of course, (113/96 * 100) = %117 performance with the 7970 versus the 7950 across all resolutions (17% increase). Showing that theoretical compute improvements have little to no bearing on actual card performance in games (so long as there is *enough* performance to max the given memory/pixel throughput).

If you look carefully at the graph linked describing the new card's performance, pixel throughput is only up 10% and memory bandwidth is up 25%, so I'd guesstimate the performance increase is somewhere between those two.
Click to expand...

I went from a 4870 to dual Eyefinity 6 5870's both @ 1920*1200. I just got a 2560*1440 monitor (2 more on the way)and the pixel density does not give me any reason to have AA enabled from what I can see. While I am anticipating getting a 8970 rather than an 8870,(or possibly dual 7970's - one now and another one later) what are your thoughts on the percentage of improvement there, about the same? or are we looking at a possible totally different chip scenario. From what I can see 2 5870's xfired are about the same as 1 7970. That is a significant jump in performance rather than 15% per generation. So this is a part of my confusion here, Is this the 'tock' of the "tick-tock" scenario with the BIG jump coming with the 9xxx series? I guess logic would dictate that!

5xxx TICK
6xxx tock
7xxx TICK
8xxx tock
9xxx TICK


My big question then is should I wait for the '8970 tock card' or just bite on the ASUS DirectCU II 28nm Graphics Card HD7970-DC2-3GD5.

Mostly used for productivity but when the workday is over I want to have as much power available as I can.

(I need to get Eyefinity 6 cards because of the monitors I am running 3-2560*1440, & 3-1920*1200)
 
Last edited:
the 7XXX series was release beginning of this year right? I'm still running an Xfire 6XXX setup, and it always comes down to that game of "you can wait forever, there's always something better"

My personal opinion is that anything 8 series is at least 6 to 10 months away at the minimum. That's based on nothing but my own way of thinking. Maybe I'm wrong, maybe nvidia is going to drop their 'high performance" 6 series GPU, maybe AMD will counter with the 8 series they've been cooking on for most of 2012.

Either way - delaying one's purchase of necessary hardware based on a rumor is silly to me. I hope to be rocking a 7970 or two here in the near future.
 
These chips have the disabled part on HD 7XXX enabled along with some additions AMD did not put in the final Southern Islands design.

In the future I'll have to buy a new desktop computer, and I may wait until HD 9XXX are released.
 
hcforde said:
So this is a part of my confusion here, Is this the 'tock' of the "tick-tock" scenario with the BIG jump coming with the 9xxx series? I guess logic would dictate that!
Click to expand...

Yes it is, you have the right idea. The biggest problem we have now is that TSMC doesn't have any half-step increments in process technology anymore (due to cost of implementation). This means you can't get any incremental reductions in power from the process, and it all has to come from design optimization for those years when there's no process improvement.

One recent example of a half-node revision was the impressive GTX 275, which offered a 20% stock improvement over the GTX 260 Core 216 and had better overclockability. It also used approximately the same amount of power. We probably won't see that ever again - core revisions now will either feature small improvements at the same power level, or marginal improvements at higher power.


My big question then is should I wait for the '8970 tock card' or just bite on the ASUS DirectCU II 28nm Graphics Card HD7970-DC2-3GD5.
Click to expand...

Get the card today. You won't see much performance improvement with the 8000 series, and it probably won't be released for 6-9 months.
 
Last edited:
I read somewhere that the 8970 would have 48 ROPs, what would that mean for performance?

By the way those upcoming 8800 cards should be great and I would assume they overclock like crazy.
 
rumartinez89 said:
I read somewhere that the 8970 would have 48 ROPs, what would that mean for performance?
Click to expand...
As far as I know, more ROPs let the GPU render more pixels per second which also includes more speed for antialiasing, so the GPU can work with higher resolutions. But if you don't add more TMUs to process the extra texture data needed for rendering to a larger area, then your ROPs become unused.
As for the shaders, they're used to modify geometry and vertices and modify pixel values before the render which I think also includes textures, then TMUs stretch textures as needed to fit that geometry and ROPs make the projection to a 2D surface of all that's visible from the camera. With geometry shaders the load on later stages can be reduced because you can modify the scene in a form that if you know what won't fit in the camera you can take it out before passing it to texture mapping and rendering.

Also, if you know you won't be rendering to higher resolutions you can use more die area to shaders so you can apply more effects to texture areas before they are mapped, and also more effects on geometry and vertices. Of course, if you improve the architecture design with the same number of units you can get more work done so with a determined number of ROPs you can get more pixels per second if you wire them more efficiently, that's what ocurred going from HD6XXX series to HD7XXX series. On consoles the need for higher efficient ROPs is not that big because they usually use eDRAM which has a lot of bandwidth and in some designs some circuitry is added to the eDRAM to allow it to apply AA to the framebuffer.

That's what I understand and probably I have some things partially or totally wrong.
 
Last edited:
Thanks filiprino, still a little confused but that did shed some light. I would assume the 8970 wont really need more shaders but if they increase the ROPs along with unusual it should be a best of a card.
 
Hmm, pcper.com lists the rumored availability window for the 8970 between December to February... So looking like 5 months or less, right on schedule for AMD to release their next Gen cards.
 
I found the link.

http://pcper.com/news/Graphics-Cards/Details-Leak-AMDs-Sea-Islands-HD-8970-Graphics-Card.
4.5 TFlops, Die size 400mm^2


My big question then is should I wait for the '8970 tock card' or just bite on the ASUS DirectCU II 28nm Graphics Card HD7970-DC2-3GD5.
Click to expand...
defaultuser - Get the card today. You won't see much performance improvement with the 8000 series, and it probably won't be released for 6-9 months.
Click to expand...

Good advice, thanks!

...and the first cards will be the 'reference designs' and I need Eyefinity6 cards. The 7970 is rated #9 and the 5870's which I have in Xfire are rated #13 in a card power ranking I saw. On top of that my CPU/MB combo is powerful enough for the 7970. If get the 7970 and wait for the 9xxx Eyefinity 6 series I can make a logical whole system upgrade and that is only about 18 months out. An issue however is Crysis 3. The developer said "It will melt your machine".
 
Last edited:
You must log in or register to reply here.
Back
Top