First retail 6870 photo; now includes full Chinese leaks and Guru3D review discussion

[Silverforce11]

GPUZ doesn't tell how many SP these GPU have, but i've asked a few ppl this a few weeks ago and they said Barts XT has ~1280 SP.

That's reasonable given the performance numbers.

But if it turns out to be 960 SP performing near a 1600 SP cypress XT, that's a huge boost for the new architecture. I'm a bit doubtful since that's a lot better than expected.

 

[3DVagabond]

Sorry, AtenRa. Your math seems fuzzier than mine. +20% SPU = +5% performance? Then by doubling the mem bus, and we've no evidence that the 5770 was Memory bandwidth limited, will net it ~15% increase? It's not adding up to me.

 

[AtenRa]

The idea is simple,

Take a 5850 and OC it to 850MHz, that’s the default frequency of 5870. You will find that the deference in performance between these cards will be les than 5% and the difference in SPs is 1440 vs 1600 (160 SPs).

5830 has 1120 SPs and 800MHz but 16 ROPs vs 32 ROPs and 725MHz for the 5850. Clearly 5830 losing a lot of performance from those 16 les ROPs and not that much from the SPs or clock frequency.

4890 has 800 SPs at 850MHz, 256bit memory with 16 ROPs and it’s a little faster than 5830 witch has 320 more SPs at 800MHz. '

4890 vs 5770, everything is the same (SPs, Clock frequency, ROPs) except 4890 has a 256bit memory vs 128bit. HD4890 is 10-20% faster than HD5770 only because it has more memory bandwidth.

If you take a 5770 and add 2 more SIMDs (12 vs 10) it will add 160 SPs + 8 more Texture Units, add 2 more 64bit memory controllers (128bit vs 256bit) 16ROPs vs 32ROPs and this I Believe will raise the performance to 20-30%. Add and a 4th generation Tesselator engine (for better DX-11 performance) and we have BARTS

 

[Ares1214]

I just found a 6850 listed in a website known as Krex Computers(never heard of it). Reseller ratings seems mixed, but any way.. a 6850 is priced at $218.44. Link

Specs:

EAH6850 DirectCU/2DIS/1GD5; Graphics Engine AMD Radeon HD 6850; Bus Standard PCI Express 2.1; OpenGL OpenGL 4.0; Video Memory 1GB GDDR5; Engine Clock 790 MHz; Stream Processor 800; Memory Clock 4000MHz (1000MHz GDDR5); Memory Interface 256 bit; RAMDAC 400MHz; DVI Max. Resolution 2560 * 1600; D-Sub Max. Resolution 1920 * 1080; D-Sub Output YES X1 (Via DVI to D-Sub adaptor); DVI Output YES X1 (Native Dual-link DVI-I); YES X1 (Native Single-link DVI-D); HDMI Output YES X1 (Native); HDCP compliant YES; DisplayPort YES X1 (Native regular DP); Adapter/Cable Bundled; 1x DVI to D-Sub adaptor

Those prices are over the usual by 15-20%, and therefore it would look more like $175-185 for the 6850 and $216-229 for the 6870.

 

[Ares1214]

GPUZ doesn't tell how many SP these GPU have, but i've asked a few ppl this a few weeks ago and they said Barts XT has ~1280 SP.

That's reasonable given the performance numbers.

But if it turns out to be 960 SP performing near a 1600 SP cypress XT, that's a huge boost for the new architecture. I'm a bit doubtful since that's a lot better than expected.

Thats exactly what i think. Like i said in an earlier post, if 960 SP=1440 SP, then what will 1920 of Cayman equal?! Way too high.

 

[maddie]

The idea is simple,

Take a 5850 and OC it to 850MHz, that’s the default frequency of 5870. You will find that the deference in performance between these cards will be les than 5% and the difference in SPs is 1440 vs 1600 (160 SPs).

5830 has 1120 SPs and 800MHz but 16 ROPs vs 32 ROPs and 725MHz for the 5850. Clearly 5830 losing a lot of performance from those 16 les ROPs and not that much from the SPs or clock frequency.

4890 has 800 SPs at 850MHz, 256bit memory with 16 ROPs and it’s a little faster than 5830 witch has 320 more SPs at 800MHz. '

4890 vs 5770, everything is the same (SPs, Clock frequency, ROPs) except 4890 has a 256bit memory vs 128bit. HD4890 is 10-20% faster than HD5770 only because it has more memory bandwidth.

If you take a 5770 and add 2 more SIMDs (12 vs 10) it will add 160 SPs + 8 more Texture Units, add 2 more 64bit memory controllers (128bit vs 256bit) 16ROPs vs 32ROPs and this I Believe will raise the performance to 20-30%. Add and a 4th generation Tesselator engine (for better DX-11 performance) and we have BARTS

False reasoning. You are looking at the only variables we can see (not being on the design team), and believing that they are the only ones that matter. What about the internal variations that we know nothing about?

I have seen tests where there was no improvement in speeds with memory overclocking on the 5770 models.

 

[Vdubchaos]

I don't see any on sale yet? When are they coming out??

 

[Lonyo]

False reasoning. You are looking at the only variables we can see (not being on the design team), and believing that they are the only ones that matter. What about the internal variations that we know nothing about?

I have seen tests where there was no improvement in speeds with memory overclocking on the 5770 models.

That doesn't even matter. There are more conclusive tests of other cards.

http://www.anandtech.com/show/2931/6

This shows that the shaders of the 5 series are slower than the 4 series.
Two cards, exactly the same memory bandwidth (clock and memory bus), same number of shaders (80 vs 80), the only difference is the 5450 is clocked HIGHER, and yet it is SLOWER than the 4550.
The only difference between the two cards is small clock speed difference (in favour of the slower card), and the architecture difference.

The reason the 5770 underperforms the 48xx cards is because it has worse shaders, and the memory bandwidth isn't particularly relevant and wouldn't fix that.

Moving away from the 5450 for a moment, besides the Radeon HD 5770 this is the only other card in the 5000-series that is directly similar to a 4000-series card. In fact it’s the most similar, being virtually identical to the 4550 in terms of functional units and memory speeds. With this card we can finally pin down something we couldn’t quite do with the 5770: clock-for-clock, the 5000-series is slower than the 4000-series.

(From the article conclusion).

 

[Janooo]

GPUZ doesn't tell how many SP these GPU have, but i've asked a few ppl this a few weeks ago and they said Barts XT has ~1280 SP.

That's reasonable given the performance numbers.

But if it turns out to be 960 SP performing near a 1600 SP cypress XT, that's a huge boost for the new architecture. I'm a bit doubtful since that's a lot better than expected.

Yeap, 1280 fits well with RPE rumors. I doubt that the performance shown in the leaks can be achieved with less.

 

[busydude]

HD4890 is 10-20% faster than HD5770 only because it has more memory bandwidth.

Lonyo pretty much nailed it. Also, there was an article in alienbabeltech regarding 5770's bandwidth.

The results generated today seem to disprove the commonly accepted idea that the 5770 is primarily held back by memory bandwidth. In actual fact ATi seems to have equipped the card with enough bandwidth to make it a reasonably balanced part overall.
Also based on the results generated today, if you’re trying to get more performance from your 5770, you should clock the core as high possible.

 

[AtenRa]

So you people say that AMD made les efficient shaders in 5000 series than last generation 4000 series.

From Ryan Smith article of the 5870 , I get the impression 5000 series shaders are more efficient than 4000 series.

“Starting at the SPU level, AMD has added a number of new hardware instructions to the SPUs and sped up the execution of other instruction, both in order to improve performance and to meet the requirements of various APIs. Among these changes are that some dot products have been reduced to single-cycle computation when they were previously multi-cycle affairs. DirectX 11 required operations such as bit count, insert, and extract have also been added. Furthermore denormal numbers have received some much-needed attention, and can now be handled at full speed. “

You have to take in to consideration that when 5450 was introduced, drivers were holding back 5000 series performance vs 4000 series. In Ryans 5450 review , the driver they used was AMD Catalyst 9.9. Im sure if we use the AMD Catalyst 10.3 and afterwards, the 5000 series perform the same or better than 4000 series.

 

[busydude]

From Ryan Smith article of the 5870 , I get the impression 5000 series shaders are more efficient than 4000 series.

To give you an idea, look at the benches of 4870 x2 Vs a single 5870.

Based on the specs, it should be clear that a 5870, with 1600 SP's, should be faster than 4870 x2, with 2x800 SP's, but its not the case and in fact 4870 x2 performs better than 5870 in most cases. Also, a 5870 has higher memory and core clocks.

 

[Skurge]

Guess that wasn't a photo-shop then

http://translate.google.com.na/tran...refox-a&hs=FGY&rls=org.mozilla:en-US:official

 

[AtenRa]

To give you an idea, look at the benches of 4870 x2 Vs a single 5870.

What drivers do they use in this comparison ?

 

[Skurge]

What drivers do they use in this comparison ?

Just so you know, the 10.3s also gave the 4800 series speedups.

 

[Lonyo]

Just so you know, the 10.3s also gave the 4800 series speedups.

And it's vs Crossfire, which typically doesn't scale all that well.

 

[busydude]

What drivers do they use in this comparison ?

Ok, these are the benches after 10.3 was released,

 

[SlowSpyder]

So you people say that AMD made les efficient shaders in 5000 series than last generation 4000 series.

From Ryan Smith article of the 5870 , I get the impression 5000 series shaders are more efficient than 4000 series.

“Starting at the SPU level, AMD has added a number of new hardware instructions to the SPUs and sped up the execution of other instruction, both in order to improve performance and to meet the requirements of various APIs. Among these changes are that some dot products have been reduced to single-cycle computation when they were previously multi-cycle affairs. DirectX 11 required operations such as bit count, insert, and extract have also been added. Furthermore denormal numbers have received some much-needed attention, and can now be handled at full speed. “

You have to take in to consideration that when 5450 was introduced, drivers were holding back 5000 series performance vs 4000 series. In Ryans 5450 review , the driver they used was AMD Catalyst 9.9. Im sure if we use the AMD Catalyst 10.3 and afterwards, the 5000 series perform the same or better than 4000 series.

There very well may be some enhancements to improve the speed of the shaders in the 5xxx parts over the 4xxx parts. But, because Cypress was getting to be larger than AMD wanted, if I remember correctly, some fixed-function parts were left out of the core to save space. Those functions were picked up by the shaders. So more work is done by the shaders vs. the 4xxx parts making it slower clock for clock overall.

AMD states themselves that the 4870 and 4890 had more bandwidth than they needed, so that is why the 5770 has 128 bit memory. So we know that shouldn't be too limiting, yet even with an 850MHz core the 5770 is only about equal to the 4870 in speed.

 

[AtenRa]

From AMDs Catalyst 10.3 release notes , we can see that 5000 and 5800 series get an enormous performance leap in most of the games vs 4800 series.

Friday is getting nearer, we will all know soon

 

[dguy6789]

To give you an idea, look at the benches of 4870 x2 Vs a single 5870.

Based on the specs, it should be clear that a 5870, with 1600 SP's, should be faster than 4870 x2, with 2x800 SP's, but its not the case and in fact 4870 x2 performs better than 5870 in most cases. Also, a 5870 has higher memory and core clocks.

Umm, the 4870x2 has a hell of a lot more memory bandwidth than the 5870.

 

[praktik]

Friday is getting nearer, we will all know soon

Is the NDA lift really friday?

I guess I had it in my head it was tomorrow...

 

[RussianSensation]

But if it turns out to be 960 SP performing near a 1600 SP cypress XT, that's a huge boost for the new architecture. I'm a bit doubtful since that's a lot better than expected.

Not sure how you arrived at 960 SP = 1600 SP Cypress.

960 SP Barts XT is clocked at 900mhz, while 1600 SP Cypress is clocked at 850mhz. Plus the 3dMarkVantage score difference is about 17% between them in favour of Cypress. This means 960 SP Barts XT = 1300 SP Cypress. But still this is too simplistic since we are looking at a synthetic benchmark that stresses shaders over textures (Vantage). Considering Barts is expected to have 48 TMUs vs. 80 TMUs for HD5870, real world performance may be totally different. And you said yourself that DX11 performance will improve

Plus 1600 SPs in Cypress were genereally used at 320x4 instead of 320x5 (since the complex shader was rarely used). So again, I don't think you can say 960 = 1300 of old SPs either. hehe.

False reasoning.
I have seen tests where there was no improvement in speeds with memory overclocking on the 5770 models.

Yes, so you would conclude that HD5770 is not memory bandwidth starved? That may or may not be true. For example, you can overclock the CPU in 1 game and you'll see 0 performance benefit. But all of a sudden if you hit fast enough speed, it will appear (Think of it this way, if you have 512mb of textures to fit into videocard VRAM and you have 256MB of ram -> then you get 320MB of ram --> you still need to swap 512mb 2x, you get 384mb of ram -> you still need to swap 2x. Then would you conclude additional ram from 256-> 384mb of ram is not beneficial. It just means it isn't enough to overcome the bottleneck! But say you have 512mb of ram, your same videocard might run another 20-30% faster all of a sudden. Therefore you are Vram bottelecked.)

Example:

So imagine you keep overclock the memory on the HD5770. You won't see any results because 20% memory bandwidth improvement may not be enough to overcome that bottleneck just like in the CPU example above.

But even if we accept that HD5770's low memory bandwidth is sufficient for its 800 SP design, with faster cards like HD6870 the additional bandwidth will be more important. I agree with you that there are other factors involved though. For example we can't assume that HD5000 SP = HD4000 SP in efficiency as Lonyo pointed out.

 

[Martimus]

Is the NDA lift really friday?

I guess I had it in my head it was tomorrow...

I thought that AMD usually did these things on a Tuesday, so that may be how you thought it was tomorrow.

The only thing I have heard was "the week of October 18th", which could be today or any other day this week. All I really know is that it hasn't been released yet, and even then I may have just missed it.

 

[AtenRa]

Is the NDA lift really friday?

I guess I had it in my head it was tomorrow...

I was told the NDA ends on 22 October, I have not seen it my self to verify it but my source is very reliable.

 

[Douglar]

GPUZ doesn't tell how many SP these GPU have, but i've asked a few ppl this a few weeks ago and they said Barts XT has ~1280 SP.

It's a common rumor, but I'd be surprised if they can get that many SPU's into <240mm2 if they keep 32 ROPS. While 1280 would be great, I'm expecting 960 in the XT version--