AMD: R8xx Speculation

[Pete]

Speculation off of a German language site:
[...]
-205mm ² (vs RV770 256mm ² & GTX280 576mm ²)
-150-160GB / s
-512bit interface is likely

They go from pad-limited with RV770 to smaller and wider with the next gen?

The speculated bandwidth is a bonus . Dunno how to reconcile ditching GDDR5 for cheaper (G)DDR3 but then possibly ratcheting up chip and board costs (and power draw?) with double the RAM chips.

 

[Squilliam]

They go from pad-limited with RV770 to smaller and wider with the next gen?

The speculated bandwidth is a bonus . Dunno how to reconcile ditching GDDR5 for cheaper (G)DDR3 but then possibly ratcheting up chip and board costs (and power draw?) with double the RAM chips.

So whats the chances of them releasing a tweaked and shrunk RV770 on the 40nm process with a 192bit bus with faster GDDR5 memory?

It seems like one of the simplest ways to get a good product out quickly and cheaply for the mid/low end SKUs.

Btw I can't remember if its been covered but has there been any word on the speculation from the RV770 thread which implied that the shader clocks might be unlocked and faster than the core clock?

 

[CarstenS]

Whoops, yeah, you're right. HD4830 is better off, but not RV730.

Jawed

Judging from this ridiculously perfect scaling from techpowerup.com
http://www.techpowerup.com/articles/other/155
http://www.techpowerup.com/articles/155/images/perlin2.jpg

I seriously doubt any performance impacts solely from the number of SIMD cores in 3DMarks perlin noise test.

 

[Jawed]

So why is HD4670's performance in this test so atrocious? It has more ALU performance than HD3850, yet is considerably slower, in the Techreport chart. Similarly in the PCGH chart, it should be slightly behind HD3870 (same number of ALUs, 777MHz v 750MHz) but is significantly behind.

I'm not saying that core count is the reason, by the way.

Anyone care to extract this shader from 3DMark Vantage?...

Anyone got the performance of HD4550?

Jawed

 

[hoom]

So whats the chances of them releasing a tweaked and shrunk RV770 on the 40nm process with a 192bit bus with faster GDDR5 memory?

Thats my bet. Though, if they chopped out the Crossfire Sideport maybe make it still a 256bit bus?

 

[Mintmaster]

So why is HD4670's performance in this test so atrocious? It has more ALU performance than HD3850, yet is considerably slower, in the Techreport chart. Similarly in the PCGH chart, it should be slightly behind HD3870 (same number of ALUs, 777MHz v 750MHz) but is significantly behind.

It's quite odd. The RightMark shaders never show RV730 behind RV670 unless it's a simple shader that's fillrate limited, and often RV730 is substantially faster (as it should be with twice the texturing ability).

I'm not sure what's up with 3DMark and RV730.

 

[CarstenS]

Possibly because 3DMark Vantage's Engine does everything in "HDR"? Maybe it's just an issues of bandwidth here?

On the contrary, I've checked Vantage's perlin noise results between an HD 4850 and HD 4870 and they are almost to scale with core clock.


edit:
HD4450 (600/800, 64 Bit Mem-Interface - mind you!): 3,72 fps in Vantage's Perlin Noise Test.

edit2:
So, per SIMD per 100 MHz it's netting about 0,31 fps, whereas RV730 only manages 0,242 Fps and RV770 about 0,72 Fps. A bit better is RV670 with 0,782 Fps - so most likely it isn't a matter of bandwith, external bandwith at least.

 

[maniac]

my take

1RV870 die will have 1.5Tflops. I can't speculate on SP's or clocks but it will be in that area

midrage(56xx) - 1die
high-end(58xx) - 2dies in 1 package 3Tfops
enthusiast (x2) - 2GPU 4dies , 6 tflops, will be seen as 1 GPU

 

[Jawed]

edit:
HD4450 (600/800, 64 Bit Mem-Interface - mind you!): 3,72 fps in Vantage's Perlin Noise Test.

edit2:
So, per SIMD per 100 MHz it's netting about 0,31 fps, whereas RV730 only manages 0,242 Fps and RV770 about 0,72 Fps. A bit better is RV670 with 0,782 Fps - so most likely it isn't a matter of bandwith, external bandwith at least.

HD4450 should be about 4.3fps (since HD4450 is 9.6% of HD4850's theoretical FLOPs) so it's around 86% of theoretical.

Any chance of extracting the Perlin Noise shader?

Jawed

 

[Jawed]

1RV870 die will have 1.5Tflops. I can't speculate on SP's or clocks but it will be in that area

midrage(56xx) - 1die
high-end(58xx) - 2dies in 1 package 3Tfops
enthusiast (x2) - 2GPU 4dies , 6 tflops, will be seen as 1 GPU

It's nuts, but if RV870 has a 128-bit bus, hmm...

Jawed

 

[igg]

midrage(56xx) - 1die
high-end(58xx) - 2dies in 1 package 3Tfops
enthusiast (x2) - 2GPU 4dies , 6 tflops, will be seen as 1 GPU

According to Fud RV870 is not a MCM.

 

[CarstenS]

Any chance of extracting the Perlin Noise shader?

Jawed

Not for me, not at the moment. For one, I'm lacking the tools to do that.

 

[rpg.314]

According to Fud RV870 is not a MCM.

If you must package 2 chips together, a la core 2 quad/3870x2/4870x2, wouldn't an mcm package give you higher intra chip bandwidth? (for stuff like crossfire sideport).

My understanding tells me yes, "shorter distances tend to have higher per line/pin bandwidth".

If it's not going to be an MCM, I guess other packaging costs forced them to do so.

 

[maniac]

It's nuts, but if RV870 has a 128-bit bus, hmm...

Jawed

yes, 128bit. GDDR5 should easely reach 5Ghz so it's enough. if you think about it it's kinda hard to squeeze a 256bit bus in <200sqmm

i don't know anything, i'm just speculating

 

[Squilliam]

yes, 128bit. GDDR5 should easely reach 5Ghz so it's enough. if you think about it it's kinda hard to squeeze a 256bit bus in <200sqmm

i don't know anything, i'm just speculating

Would there be any practical reason not go go with a 192bit bus?

 

[MfA]

Even with a minimum bump pitch of 250 um putting a 256 bit bus on a 200 mm2 chip just shouldn't be a problem AFAICS (and 250 um is old/cheap tech). That leaves room for 3200 connections!

 

[Jawed]

So far there's not been even a whisper about the explicit RV770 refresh. Is there going to be one?

http://www.techpowerup.com/75685/AMD_to_Give_RV770_a_Refresh_G200b_Counterattack_Planned.htm

[...] to reengineer parts of the GPU to facilitate higher clock speeds.

Hmm, 950MHz, 27% faster?

Article suggests AMD has already attempted to provided a "super-clocked" SKU based upon RV770, but failed:

AMD has already attempted to achieve something similar, with its big plans on the Super-RV770 GPU, where the objective was the same: to achieve higher clock speeds, but the approach wasn’t right. All they did back then, was to put batches of RV770 through binning, pick the best performing parts, and use it on premium SKUs with improved cooling. The attempt evidently wasn’t very successful: no AMD partner was able to sell graphics cards that ran stable out of the box, in clock-speeds they set out to achieve: excess of 950 MHz.

Jawed

 

[Jawed]

Thats my bet. Though, if they chopped out the Crossfire Sideport maybe make it still a 256bit bus?

Hmm, if the CrossFire Sideport is pointless on RV770, then perhaps RV870 won't have it. That alone should lead to a serious drop in die size.

So far I've been assuming that this would be retained - but I'd forgotten about the failure of it so far.

Jawed

 

[Mintmaster]

Even with a minimum bump pitch of 250 um putting a 256 bit bus on a 200 mm2 chip just shouldn't be a problem AFAICS (and 250 um is old/cheap tech). That leaves room for 3200 connections!

Is it that simple, though? Can you put pads wherever you want without consequence? I'm sure substrate complexity has costs, too.

 

[Shtal]

imagination!!


512bit Shared memory controller, 1600x2=3200 shaders, 80TMU's per GPU.