Shader-deficient Radeon HD 4830 Could be in the Channels

[btarunr]

The ATI Radeon HD 4830, the newest foot-soldier from the red-camp, is AMD's answer to the GeForce 9800 GT, a gap left by the company, that may have been eating into its profits for months now. As usual, TechPowerUp received its review samples from TUL (PowerColor), AMD's long-standing partner, and AMD itself. We reviewed both of them, as soon as the product became official today. During the course of reviewing them, with the card from AMD in particular, our reviewer, W1zzard noticed an anomaly: the sample from AMD was showing an abnormal stream processor count of 560.

W1zzard also authors the GPU-Z diagnostic utility, and it is his routine chore to program the utility to detect a new GPU. The newest build of GPU-Z detected the card from AMD to have as many as 80 stream processors disabled from the original specifications for the Radeon HD 4830. In his article, W1zzard attempts to explain this anomaly.

The RV770LE graphics processor (GPU) is a "lite" version of the RV770 GPU that went into making the Radeon HD 4850. Out of the 800 stream processors (SP) the RV770 has, AMD carved out the new GPU by disabling two blocks of 80 SPs each, resulting in a 640 SP-laden Radeon HD 4830. Earlier, GPU makers would simply make a BIOS edit that disabled blocks of shader units. Both NVIDIA and AMD these days, resort to physically modifying the GPU, making irreversible changes to the amount of SPs available for the GPU to use. On the sample TechPowerUp got from AMD, there were three such blocks (perhaps accidentally) disabled, crippling the SP count further down to 560.

This did reflect in the card's performance against the one PowerColor sent based on the same GPU. The card from PowerColor outperformed the reference card due to its "proper" stream processor count. We urge reviewers and enthusiasts to verify the stream processor counts of their Radeon HD 4830 samples or products using the aforementioned build of GPU-Z.

[alexp999]

Wow, thats not good. I could see no mention of it in any other reviews. Maybe the AMD sample was an old one or a dud, but all the MF's now have the true 640SP version?

If its advertised as 640SP and then one you get is 560SP, you could take it back right?

[W1zzard]

until the launch there was no gpu-z version that could even read the number of shaders. older versions would only show "800" shaders for all rv770 cards including those with 640 or 560 shaders

Originally Posted by alexp999

If its advertised as 640SP and then one you get is 560SP, you could take it back right?

absolutely. that is if you find out

[WarEagleAU]

Hmm, looks like ATI is even prone to mistakes. Hopefully this gets sorted out. ID be pissed if I didnt get what I paid for.

[jbunch07]

So will there be two versions of the card, One being 640SP and the other 560SP or was that an accident like stated above?

[PVTCaboose1337]

I think it will be luck of the draw! Too bad! Some poor guy might get the 560 while another gets 640... That is really not fair.

[Selene]

oh i can see it now, 4830 and 4830 "MAX core 640"

4820 anyone?

[DarkMatter]

Hmm by my standards the 560 SP one is the better one. 3% slower, but it's much more power efficient, and that would lead to a cooler and quieter card.

[W1zzard]

Originally Posted by DarkMatter

Hmm by my standards the 560 SP one is the better one. 3% slower, but it's much more power efficient, and that would lead to a cooler and quieter card.

no, the numbers are misleading. the low power draw is achieved by much much lower 2d clocks. ati has like 150 mhz and power color has like 500 mhz.

the different cooler and fan control settings are what make the card quieter.

the heat output of the additional 80 shaders is not much .. a few watts i'd say

[DarkMatter]

OK thanks for the info about the idle power consumption. That doesn't change my view anyway. Under load it still consumed 18w less according to your own reviews, while the 640 SPs one consumed as much as the HD4850. With only 3% performance difference between them I find it the 560 SP one the best of both.

Also it's common sense that with less SPs working there's going to be less heat and lower fan speeds would be required. I know the fan speed (and the fan itself) is different in the two you reviewed and not based on SP number, but theory is still by my side. The difference might not be huge, but it'd be there for sure.

In the recent times performance is the only thing that I have found that can't be reduced to theories, you have to test them and looking at your reviews the conclusion is clrear. Clock for clock 560 SP RV770 performs almost on par with the 800 SP one. Even in F@H (real life shader heavy app) there's no difference, so as a personal preference I would take the 560 SP one.

I'm not apologizing for them and I understand there might be many factors that could lead to those differences, but I am pretty much only talking about the two cards you reviewed. Between the two I like the 560 SP one better.

[wiak]

everyone makes mistakes
and review samples are work-in-progress/engineering samples like the one from AMD
and the one from TUL (PowerColor) was a Retail i asume that had correct numb3rs of shaders

[lemonadesoda]

This could be deliberate. Perhaps the 4830 was just a little too good and there was a fear of cannabalising higher performance part sales.

Even with 560 shaders, the performance is great. And they needed a gap between 4830 and 4850 and 4870 to warrant pricing in "different market segments".

I'm sure they werent expecting a genius like w1zzard to spot that so early on.

There is no way this was "a mistake". A GPU manufacturer isnt going to do a simple math mistake about the number of shaders on/off.

[btarunr]

Originally Posted by wiak

everyone makes mistakes
and review samples are work-in-progress/engineering samples like the one from AMD
and the one from TUL (PowerColor) was a Retail i asume that had correct numb3rs of shaders

AFAIK, as many as four reviewers have confirmed to have gotten this 560SP sample, and noticed disparity in scores between AMD-sent cards and others. Make your own conspiracy theory.

[thebluebumblebee]

Originally Posted by Selene

oh i can see it now, 4830 and 4830 "MAX core 640"

If you remember back to the 9800 days, (I'm still running a 9800ProAIW) there were, in descending performance, 9800XT, 9800Pro, 9800, and 9800LE (which sucked). I'm wondering if: 4870=XT, 4850=Pro, 4830=_, and 560SP=LE?

Having worked in inventory for many years, this sounds like a mix up between the sales and shipping/inventory folks.

This will be interesting to see if we're right. But, it may also be an attempt to market a lower power/sound card for the quiet crowd who still want some gaming. $100 card with passive cooling?

[Kreij]

Originally Posted by DarkMatter

With only 3% performance difference between them I find it the 560 SP one the best of both.

Where did you come up with a 3% difference in performance.
From W1z's tests the AMD card got 219.8 fps and the powercolor got 250.5

219.8 / 250.5 * 100 = 87.74% or a 12.26% reduction in performance for the AMD card.

Or am I not understanding what you posted?

[DarkMatter]

Originally Posted by Kreij

Where did you come up with a 3% difference in performance.
From W1z's tests the AMD card got 219.8 fps and the powercolor got 250.5

219.8 / 250.5 * 100 = 87.74% or a 12.26% reduction in performance for the AMD card.

Or am I not understanding what you posted?

LOL What are you saying? You just can't take ONE benchmark at ONE settings and claim the difference is X. I took my numbers from Wizzard's performance summary which is the summary for more than 40 different benchmarks.

[Kreij]

I realize that you cannot gauge performance by a single test, but I thought you were just refering to the numbers from his article about the missing shaders.

Thus my adding the last line to my post.
I was misunderstanding you.
My bad

[DarkMatter]

Originally Posted by Kreij

I realize that you cannot gauge performance by a single test, but I thought you were just refering to the numbers from his article about the missing shaders.

Thus my adding the last line to my post.
I was misunderstanding you.
My bad

LOL I was curious where your numbers came from, because I didn't recognize them. I have very good memory, but I didn't pay attention to the actual numbers in the perlin noise test, only to the bars. I thought: "With such high numbers it must be Far Cry or Quake4", but I was not lucky in Wizz's reviews.

But yeah, obviously in purely shader test like perlin noise, more shaders equals more performance and as Wizzard pointed out that perlin noise seems to be a very good benchmark as long as you only compare the same architecture.

Accordingly in a fillrate test pretty much only the texturing power will be shown.

In real world things change though, you need a balance. For instance I consider F@H a very good real world shader compute benchmark. As good as it can get in real world testing at least, and there the difference is small. (in real applications you will never be able to bypass other parts of the architecture.)

You could think that the more the better, and so it is as long as the are no drawbacks to that aproach. In my personal view the drawbacks in this case are enough (power consumption, heat, price) that I consider those extra shaders worthless. Not only on gaming where the difference between the cards (both HD4830s and HD4850) can be explained by the pure clock speed diference, but F@H shows almost no difference in performance (clock for clock) even when compared to HD4850 or HD4870, and IMHO that's unforgetable for an architecture that was supposed to be efficient. They should have just forget about the marketing division and release RV770 as they first devised it: with 640 SPs. But that's only my opinion.

[MrMilli]

Damage from Tech Report has this to say:

I checked with AMD. They told me that they believe this utility is reading the card's registers wrong and that there is no performance difference between cards that show 560 and 640 SPs in the utility. They told me they're following up with their engineering folks, the utility author, and the publication to get this sorted out. We'll keep an eye on it, but it appears there's no real issue with the cards and their performance here, at least right now.

[MrMilli]

PCGH had these issues too:
http://www.pcgameshardware.com/aid,6...viewed/?page=5

[infrared]

A bit strange AMD denying it. And accusing gpuz of being wrong.

[jbunch07]

Blasphemy! GPU-Z is never wrong!
Plus the performance is noticeable.

[btarunr]

No, GPU-Z isn't reading anything wrong. The shader defficiency clearly reflected in the scores. If AMD cannot open their eyes to this simple fact, I'm more amused than concerned.

[eidairaman1]

Originally Posted by alexp999

Wow, thats not good. I could see no mention of it in any other reviews. Maybe the AMD sample was an old one or a dud, but all the MF's now have the true 640SP version?

If its advertised as 640SP and then one you get is 560SP, you could take it back right?

yes you can, you can state that the board you bought was the wrong part number.