HD 5870 first impressions

[Davros]

so what should you try and buy the high voltage one or the low voltage one ?

 

[Voxilla]

There appear to be at least three sources of leakage in transistors.
http://www.extremetech.com/article2/0,2845,1860883,00.asp

Gate leakage, drain-source leakage and BTBT leakage, not sure which one is dominant.

 

[Voxilla]

Lightman is 100% right the reason the high leakage part has a lower voltage is because they don't need as high a voltage to reach the same clock speed.

Is there any article I can read about this ? As an electrical engineer this interest me, I've not been doing electronics for a long time, so my IC knowledge may need a brush up.

 

[Lightman]

Is there any article I can read about this ? As an electrical engineer this interest me, I've not been doing electronics for a long time, so my IC knowledge may need a brush up.

This is a bit dated explanation but it does the justice:

hitechjb1
09-14-03, 02:38 AM
I do not know what wafer integrity refers to. There are process variations of a given silicon manufacturing process, as any manufacturing process. As a result, the intrsinic silicon proporties, such as transistor channel length and width, gate oxide thickness, silicon carrier doping, transistor threshold voltage, leakage current, ..., of a wafer vary to certain extent (sigma variation). As further scaling down, statistical variation comes into play, namely, nearby transistors in the same chip/wafer can behave differently.

A more interesting question is what the implications of these wafer properties of lower threshold voltage and shorter channel length due to process variation of a manufacturing process are, as I suspect for the the Tbred B DLT3C. It is being rated at lower Vcore but it can run faster than other Tbred B at same voltage. Even it is manufactured with 0.13 micron like other Tbred B, it is effectively behaving like a chip with less than 0.13 micron, resembling the future generation trend.

As the transistor size (channel length) of future generations of silicon chips are scaled down to, e.g., 90, 65, 45, ... nano-meter, the supply voltage, transistor channel length and threshold voltage will be lowered accordingly. Even the supply voltage is lower, the transistors run faster, both current and power density also increase (actual trend). As the transistors are scaled down, logic gate delay decreases, both the active power density (W/cm^2) and the passive leakage power density (both gate and subthreshold) increase. The passive component increases at an even faster pace.

For more details about the low voltage Tbred B 1700+, refer to
Why the 1700+ can run so fast at low Vcore? (http://forum.oc-forums.com/vb/showthread.php?s=&postid=1734079#post1734079)

Originally posted by hitechjb1
Why the 1700+ can run so fast at low Vcore?

The Tbred B 1700+ DLT3C is based on the same 0.13 micron bulk silicon process as all the other model 8 (Tbred A and Tbred B) from XP 1600+ to 2800+ (recently 3000+). (BTW, Tbred B has one more metal layer than Tbred A, both are 0.13u.)

The hammers (Opteron, Athlon 64) are based on 0.13 mircro SOI process, will go to 0.09 mircro eventually.

The Tbred B 1700/1800+ have the same transistor count, same L1, L2 cache size, same number of metal layers, same chip dimensions, ... as the 1.6 and 1.65 V rated Tbred B.

Side track: Same for Barton, which is also based on the 0.13 micron process. But it is a different chip, different transistor counts, chip dimension and has bigger L2 cache of 512KB instead of 256KB in Tbred.

I think the reason why the Tbred B 1700+ DLT3C can work at rated 1.5V and can be clocked at simliar highest clock frequency (if not better) as all the other 1.6V and 1.65V rated Tbred B is due to the following:

Its transistors have lower threshold characteristics due to process variation which produces transistors with shorter channel length. Shorter channel means lower transistor threshold, runs faster, draws larger leakage current and higher active current (hence higher active power). According to AMD spec, the 1.5V 1700+ has higher rated current than the 1.6V 1700+ (about 7% more).

Threshold voltage of a transistor is the gate voltage above which the transistor will conduct current orders of magnitude higher from source to drain compared to that below the threshold. Chips with lower threshold transistors can perform equally well with a lower supply voltage (Vcore) as those with higher threshold, because the transistors can conduct at a lower gate voltage.

This is normal for a given silicon process (say 0.13u) to have such variation that some transistors in certain chip die have shorter channel length (less than 0.13u) or some have longer channel length. Those that have shorter channel length have faster intrinsic speed and can run as fast when smaller Vcore is applied (pros). On the other hand (cons), due to the lower threshold voltage which draws higher leakage current and generates more heat at the same higher Vcore, these chips can run as fast at a low Vcore as the higher Vcore rate chips, but they will max out at a lower Vcore compared to the higher Vcore rated siblings.

The 1700+ has a run-away current at a lower Vcore compared to the 2100+. Run-away current refers to the leakage current and the heat generated positively feeding each other resulting instability.

The final oc success of the Tbred B 1700+/1800+ DLT3C is a race between its natural, born, intrinsic characteristics, the balance and tradeoff between the smaller channel length, lower transistor threshold, hence faster, and the opposing, negative behaviour of higher leakage current, and heat generated.

SOURCE

Obviously the trick is to get good balance, because if you are targeting certain clock ceiling making channel length too short can produce bad dies (shortcircuit) and too long channel will prevent you from hitting desired clock!

Mobile parts are targeting low leakage and performance parts are produced from higher leakage dies because they have much higher target TDP...

 

[Voxilla]

I needed to sleep a night over this, we are getting way off topic here, but anyway.
The above tells the leakage is due to shorter channel length which also enables less voltage for same switching speed, all of this due to process variation. So this is source drain leakage. As the explanation is from a speculative thread and not a scientific article I'm not sure what to think of it.
MfA suggest the leakage is gate leakage due to a thinner layer in the gate insulation.
The thinner insulating layer reduces switching time, and so higher frequency is possible with lower voltage. As a result part of the higher current can be compensated by lower voltage so to not increase power consumption. Sounds plausible.

Anyway how do you read the voltage of the HD 5870 ?
I'm curious what is mine as my GPU has nearly zero overclockability.

 

[dkanter]

Leakage and frequency are related, but somewhat independent.

It's quite possible to get fast/cool parts, slow/hot parts, slow/cool, fast/hot - and everything in between. There are more than 4 corners for process technologies.

What Dave is saying is that because there aren't as many SKUs as CPUs, they try to get the distribution to be narrow to maximize saleable products.

David

 

[Mendel]

I ordered an HIS Radeon HD 5870. I'll try and give you guys some first impressions when the card arrives.

 

[ET]

I'm also interested in getting one, since currently my 4850 can't run compute shader, so I can only test my programs on my GTX 285 (and other G92 cards). But right now the only cards I can find here cost about US$440 (cards from both ASUS and MSI cost about the same).

4850 should have computer shader support in the next driver, I think. Should run the same stuff as the GeForce cards, but a lot less than the 5xx0.

 

[Mendel]

My card arrived in the local post office today, I should be able to go get it in an hour.

 

[Mendel]

edit: this just came in


gpu-z for 4870vs5870


some 3dmark06 feature tests.


I have to say the multitexturing fillrate difference isn't what I expected. (expected double, got more) I'll have to go through some of the reviews and specs again to understand what's going on.

For compare urls:

standard benchmark
http://service.futuremark.com/compare?3dm06=12232084

2560x1600 but other than that, same settings.
http://service.futuremark.com/compare?3dm06=12232211

Score barely drops at all even though there's 4096000 vs 1310720 pixels... 4096000/1310720=3,13
So even with that much higher workload, barely noticeable effect on 3dmark score.

I played a couple of games of RA3 and some Need for Speed Shift... and in situations where I was previously frame buffer limited, the framerate is up tenfold or more. 2560x1600 is very much a resolution I can use now. With 4x or sometimes even 8xAA. Fan noise isn't much different than it used to be, certainly no issue. Thus far this thing is amazing. Tomorrow I shall try to bring it to its knees with 3dmark vantage extreme preset and crysis.

 

[CarstenS]

I have to say the multitexturing fillrate difference isn't what I expected. (expected double, got more) I'll have to go through some of the reviews and specs again to understand what's going on.

Pretty normal, since a decent MT-Test is not bound by bandwidth. So you got 2x scaling from doubled # of execution units and then some from 100 MHz higher engine clock, which should amount to 2.26x the 4870.

 

[Mendel]

I'm getting 2,77x the 4870...

 

[OpenGL guy]

Pretty normal, since a decent MT-Test is not bound by bandwidth. So you got 2x scaling from doubled # of execution units and then some from 100 MHz higher engine clock, which should amount to 2.26x the 4870.

Except he's achieved 2.77x scaling.

The difference is that 4870 is somewhat interpolator limited in this test, 5870 is not.

 

[Dave Baumann]

I'm surprised that we don't remember this discussion.

 

[Mendel]

Anyways, here's a video comparison thingie, upgrading from 4870 512MB to 5870 1GB.

http://www.youtube.com/watch?v=FaNnEeAE_F4

Granted, many of the benchmarks aren't fair because the 4870 512MB runs out of local memory. But so what? This is how I see the difference, and I'm showing it to you as it is



In a nutshell: This is pretty close to being the most awesome hardware upgrade I've ever done.

 

[DuckThor Evil]

Granted, many of the benchmarks aren't fair because the 4870 512MB runs out of local memory. But so what? This is how I see the difference, and I'm showing it to you as it is

Nice video, but I don't believe that is how you actually saw the differences, or you are crazy if you played those games with those settings on your old 4870. Better comparison would be highest playable settings on each card, because that's the real difference between the cards and the difference people are actually going to see when they upgrade.

 

[Lightman]

Nice video Mendel!

I've moved from HD4870 512MB CF to HD5870 and still it is big improvement in more recent games like Crysis series or NFS: Shift even at 1920x1200 res.

 

[Mendel]

Nice video, but I don't believe that is how you actually saw the differences, or you are crazy if you played those games with those settings on your old 4870.

While I could play the first few levels if I took off the AA and switched to gamer shaders... I actually had to stop playing Crysis Warhead at about that train sequence because I could not find a decent playable compromise that looked good enough. Now I can resume playing.

Better comparison would be highest playable settings on each card, because that's the real difference between the cards and the difference people are actually going to see when they upgrade.

Good idea but then I'd have to be able to actually upload higher resolution videos. (Youtube isn't quite there yet.) Either that or you often couldn't see the difference when I switch resolutions between cards.

You'd start seeing differences if I disabled some features or details but then would you disable all the same things and make the same compromises?

 

[Mendel]

ran some tests again and found out that I'm cpu/system limited when recording with fraps.

When not recording, average framerate of 25,40 in test 1 of Vantage.
When recording, average framerate of 15,24.

So, yeah. I guess that would be why people don't make videos like that

 

[Kaotik]

ran some tests again and found out that I'm cpu/system limited when recording with fraps.

When not recording, average framerate of 25,40 in test 1 of Vantage.
When recording, average framerate of 15,24.

So, yeah. I guess that would be why people don't make videos like that

Well, writing uncompressed videodata isn't cheap for HDD performance.
IIRC writing the videofile straight to separate HDD from the windows & game improves performance quite a lot