8800 GT Vs. HD 3870

[Hardwire]

nVidia GeForce 8800 GT
ATi Radeon HD 3870


Ever since I got here I have been pounded with Intel and nVidia is better... so I decided to build an AMD machine with a Radeon card. (Yes, I would in fact cut off my nose to spite my face.) And after reviewing all the specs I have to ask. What the hell?!

Manufacturer: nVidia
Series: GeForce 8
GPU: G92
Release Date: 2007-10-29
Interface: PCI-E 2.0 x16
Core Clock: 600 MHz
Shader Clock: 1500 MHz
Memory Clock: 900 MHz (1800 DDR)
Memory Bandwidth: 57.6 GB/sec
FLOPS: 168000 GFLOPS
Pixel Fill Rate: 9600 MPixels/sec
Texture Fill Rate: 33600 MTexels/sec

Manufacturer: ATi
Series: Radeon HD 3k
GPU: RV670
Release Date: 2007-11-15
Interface: PCI-E 2.0 x16
Core Clock: 775 MHz
Shader Clock: 775 MHz
Memory Clock: 1125 MHz (2250 DDR)
Memory Bandwidth: 72 GB/sec
FLOPS: 248000 GFLOPS
Pixel Fill Rate: 12400 MPixels/sec
Texture Fill Rate: 12400 MTexels/sec


I've highlighted who wins in individual specifics. So my question is, is texture fill rate and shader clock speed really so importaint?

I mean seriously, can someone tell me why the 8800 GT gets so much hype when for $10-30 less you can get a card that craps on it in so many different ways?

I'm not trolling or anything, I just seriously want to know whats so special about the 8800 GT.

[jlreich]

First have a look at this thread (http://www.pcguide.com/vb/showthread.php?t=39536). It about two and half years old but still a very good read.

Since that thread even more has changed. ATI have totally changed the way some of it's architecture works. The term " pixel pipelines" no longer apply in the way they used to when it comes to either manufacturer.

As I said before the two different manufacturers are very different on how their cards work. You can't judge them side by side by the numbers. Apple and oranges, apples and oranges....

In the end the 8800GT overall beats out the the 3870 in benchmarks. Don't get me wrong the 3870 is a good card, but the 8800GT does do better in most cases.

I wish I could give you the in depth analysis like Saphalline can, but I don't have dual channel brain or blue blood and pointy ears like he does. :p

[Heartborne]

In EVERY benchmark on the web the 8800 GT outperforms the 3870 for gaming. The 3870 really shines in the HD video area... so if you're watching blu-ray movies on your computer, the radeon will pull you through there. But if you want the best gaming card for your money, the 8800 GT is the hands-down winner. Google "8800 gt benchmarks" and "HD 3870 benchmarks" and you'll see what I mean.

[Variable]

Drivers will continue to help the 3800's give better performance. It outscores the GT in hardware specs but in real world it is not able to use it all right now.

If you have the right motherboard you can add multiple 3870's. Mine supports up to 4. So as the 3870X2 gets cheap, I will add one. Nice thing about the new AMD architecture is the platform is made to be upgradeable for gaming. One 3870 looks great running on my machine. In a year the drivers will be much better and I can add more cards as they get cheap and I need the extra speed. I never fall for benchmarks anyway, who cares if the game gets 240 or 280 FPS. It is a moot point. It is like saying my car will go 180mph and yours will go 200 mph but you never drive outside your little neighborhood. Some people just like to have a super machine sitting in their driveway. They never actually race it.

[Heartborne]

Good points, variable... but we're not talking the difference between 240 and 280 fps with a mid-level card like this. We're talking the difference between under 40 and over 60 fps, unplayable vs. seamless, in many cases.

On that note, by the time the ATI drivers catch up to the hardware, the hardware will be obsolete.

On that note, the architecture of the cards is so dramatically different that the technology on the ATI card means something entirely different from that on the Nvidia card.

When you break it all down the simple fact is that the AMD/ATI platform writes checks it can't cash. Why wait for ATI to make drivers that utilize their video cards properly when Nvidia's drivers do it now? That's like asking me to wait a week before I hire you to paint my fence when my neighbor says he'll do it this afternoon, have it finished sooner and charge me less for it. It just doesn't make sense.

Of course we can argue this forever; if you've made an investment in a radeon card the last thing you want to do is admit you'd have been better off with an nvidia card. But you would have.

[Variable]

meh, you know what they say about opinions. I think most benchmarks are useless for normal players because of the other components factor but, here is one comparing the 3870 and 8800 GT, among others, in several games - one of which I play, COD 4. I think you'll notice that where the 8800 GT performs better it is between 3-10 FPS. Neither shine all that much in these benchmarks.

Hardware geeks and their opinions abound but the average gamer uses the same rig for several years and updates piece meal. Fewer than 60FPS games become choppy. Neither the 3870 or the 8800 GT can play DX 10 Crysis with all the bells and whistles on - so given that, you are looking at average performance for the dollar over time. Taking in to account the upgrade over rebuild mentality of many gamers, ATI's decision to cater to that seems like a smart idea. You cannot put 4 8800Gt's in one machine can you? Can you mix an 8800GT with a 9800GX2 in SLI?

People who don't buy hardware a lot and just go by benchmarks get confused. My machine plays COD 4 much better than Tom’s benchies. So either the drivers are better or my other components are better. This means for me, the benchmark is a guideline only. And that is my point, anyone looking at buying hardware should take benchmarks with a grain of salt. It does not matter what they score are 1920X1200 if your monitor only supports 1280 X 1024 for instance.


http://www.tomshardware.com/reviews/amd-radeon-hd-3800,1726-17.html

[Heartborne]

Granted, Variable, the difference is small... but it is still a difference. Why not get up to 20% better performance for the same price? In many games you might get away with turning on AA or AF without losing any performance with the 8800 gt whereas you'll lose a few frames on the Radeon. The radeon is not a BAD card, it's just not as good.
Quite matter of factly, better is better. You can argue it all you want, but the fact is the same... 63 fps is better than 54 fps, even if the player doesn't notice.
I do have only one major qualm about the ATI cards, which is that in some benchmark tests they didn't run certain games AT ALL. That's a big problem right there.
So, in any case, if you don't mind losing a portion of performance because you just prefer one over the other then that's a personal choice. I go with the most performance for my money.

[George Hallam]

That benchmark is under different condition though, they find certain hardcore scene that will put the graphics cards to there limits, as you may know Variable when someone throws a smoke nade and you end up shooting through it you frames drop dramatically (i hate smoke nades i cant use my sniper ^^) so if they did the whole benchmark with 25people in the map with smoke everywhere and bullets flying obviously it will be slower.. Im quite sure if their benchmark was in closed hallways mostly looking at the ground and sky then people would be disappointed to find they dont really get 200Fps.

benchmarks do help to show the outlining performance of a card, so you can see what you will be getting and able to compare it against other cards.

off the point now Variable do you have Xfire? or something i would like to have a game against you ;)

P.S - a birdie told me that someone is popping in to make this thread alot more lengthy and in depth :p

Edit: (just seen hearbones thread) I go with the most performance for my money. so why you buy the 9800GTX? :p

i agree with you to a point there, but if we were all like that people wouldn't buy the 9800GX2 (SLI two 8800Gt's for less money) or the 9800GTX (just a beefy 8800Gts). Like at this moment im saving for a 9800GX2 (or the 9900GTX when it comes out) not because it have good bang for buck just because it is good. from going from the 8800Gt to a 8800GTX you may only get 5-10FPS difference but that MATTERS when you are only getting 30 in the first place. The top end cards last longer and you get much more enjoyment. How much of a disappointment would it be someones gets a 8800Gt gets it home and put a game on full high and only get 20FPS when they could of spent an extra $100 (saved $20 a week) and bought a 9800GTX or saved a little longer and SLI'ed?

[Variable]

Well, the GT IS slightly better than the 3870. But "better is better" is subjective - I'm not saying that the 3870 is better than the 8800GT but that the video card you purchase is not, and should not, be based on benchmarks.

You have to have a motherboard and power supply that supports the video card. So your choice of other parts dictates what video card you get. If you have an AMD machine with a motherboard that supports quad crossfire... that the 8800 GT is slightly better is a meaningless and the price was not equal when I bought mine, not even close. Nvidia had to cut prices because the AMD cards were killing them on price/performance metrics.

Many gamers do not rush out and buy the best card, they buy the best card they can afford and what their hardware supports. If you did have to buy a new rig that would be upgradeable and future proof for several years ATI is offering that now. Given that fact, the 8800 GT is not a better buy than a 3870. It is a better card but, it is not a better deal.

[Rick]

Granted, Variable, the difference is small... but it is still a difference. Why not get up to 20% better performance for the same price? .


I'll give you one more MAJOR point to consider

This was true with the 8800's above that I can't say from personal experience

Nvidia driver support and updates for ANYTHING other than games is way behind or NON existent

Video playback in windows is way to slow using the 8800's
The 8800 gts 512 specifically

In Every case upgrading the drivers reduced the performance of video playback

The only cure was to step back to the original drivers and Block the vista update drivers


if your gaming it's cool.
If your Not a gamer or only an occasional gamer and want Vista to perform as a working computer
then Nvidia is NOT the way to go

remember also that some of us who ended up with these cards paid the full price when the were selling for $300 to $500 and UP

BTW.
They still haven't fixed the drivers for US Non gamers

[Heartborne]

All very good points, Rick.
Actually, to answer George's question, I upgraded to the 9800 gtx from the 8800 gts 320... rather than the 8800 gts 512, which would have actually cost me more.

If I could trade in my 9800 gtx for 2 8800 gt's I'd do it in a heartbeat.

[Variable]

benchmarks do help to show the outlining performance of a card, so you can see what you will be getting and able to compare it against other cards.

off the point now Variable do you have Xfire? or something i would like to have a game against you


I agree they are a baseline for comparison because they should be using the same hardware and only subbing out the video cards.

I don't have X fire. Havn't seen the need for it (yet). When the 3870X2 gets cheap I will buy one. I played MoH for about 6 years. I will probably play COD 4 and it's variants from now on and for a like amount of time. I don't have time to game that much. I havn't bit off on Vista yet, I use XP pro 64bit.

[saphalline]

George Hallam emailed me (he's the only one who does!) and we talked about this thread. I thought it was interesting but at first I didn't want to get involved because of my time constraints. But GH suggested a deal: I post once a week. That's it. Saves me the burden of having to answer every hardware question under the sun while also not voiding my presence here altogether. I hadn't really thought about that before, but it seemed like a good idea. So here I am! Is everyone ready for the answer? Huh? Any takers?...


First, I'll address a few of the comments thus far:

Ever since I got here I have been pounded with Intel and nVidia is better...You're missing one very important aspect: Intel and NVidia are better right now! I've never seen this piece of advice given here on the forums without that little caveat. There have indeed been times when AMD is better, or ATI, or a myriad of other companies that are now defunct. But as of right now, May of 2008, Intel and NVidia are better. They consistently give the best performance per unit of currency spent for most custom PC uses. Some of the implications are in fact changing, but if you're talking about a gaming system, this advice is 100% solid right now!

is texture fill rate and shader clock speed really so importaint?YES!!! Haven't you seen the benchmarks?? If the 8800 GT pulls ahead of the HD 3870 more often than not, then it stands to reason that its areas of better specs mean the most overall (not that correlation implies causality but it happens to match this time ;)). The only thing I have to ask in relation to this specific question is: Are you looking at the complete picture? There are several specs related to each GPU that are not listed but are imperative to answering the question. Thus it further stands to reason that these heretofore unknown specs are integral to crowning a champ!

I don't have dual channel brain or blue blood and pointy ears like he does.Hey hey, now! No name calling! :p

Drivers will continue to help the 3800's give better performance.Indeed! An excellent point considering the first-looks at the HD 3870 X2, which exposed CrossFire driver issues! This is one of the downsides to ATI's multi-GPU solution that AMD bought - it's very software-dependent! The upside to this is that unlike SLI, CrossFire is hardware independent, meaning that ANY 3D application can take advantage of the scalable performance offered, not just games that are "coded" to take advantage of it (*cough* HyperThreading! *cough*). The problem with CrossFire for the past few years, however, is that the driver support never really caught up to what SLI can do in hardware, which hurt its reputation (perhaps even irreparably). Now that AMD has more or less consolidated the efforts, CrossFire performance is improving by leaps and bounds lately, as evidenced by the HD 3870 X2's performance showing against the mighty 9800 GX2!

but we're not talking the difference between 240 and 280 fps with a mid-level card like this. We're talking the difference between under 40 and over 60 fps, unplayable vs. seamless, in many cases.Another excellent point highlighting the reason why ATI had been playing catch-up since the days of the X1800! The buy-out didn't seem to help matters, especially with AMD inheriting the awe-inspiring failure that was R600! However, it's important to note that ATI was once top-dog, with NVidia in the catch-up position - like say, oh I don't know... the Radeon 9700 Pro vs the GeForce 5800 Ultra, perhaps? :rolleyes: The 5800's subpar performance, enormous power appetite, and faulty stock cooler certainly didn't place NVidia in the best light back then, did it!? But the main point here is that we're running into games these days that are either playable with a certain vid card, or they aren't. It's that simple. There's no grey area here as far as gamers are concerned. well, not the small percentage that pay $150+ USD just for their vid card! All you have to realize is that Crysis can make ANY computer cry and you have your answer!

anyone looking at buying hardware should take benchmarks with a grain of salt.A truism if ever there was one! :p And one that should come as no surprise to anyone! If the price of a gallon of crude oil is anything to go by, you'd think the world economy was collapsing!! :eek: Can any one metric really give us the straight answer all by its lonesome? I think not! But you have to realize that benchmarks are not entirely synthetic. Most game-based benchmarks are in fact pre-recorded input series that someone actually "plays", so that when other systems use that benchmark they're all doing the same thing in-game. The only point you can argue there is whether or not the pre-recorded gameplay matches your own style! ;) Benchmarks pulled from games are a lot more reliable than something like 3DMark, which is entirely synthetic.

I do have only one major qualm about the ATI cards, which is that in some benchmark tests they didn't run certain games AT ALL.That's largely unfair from what I've seen. It's one thing when the performance drops so low that the game engine crashes (which is a hit against the game itself, IMO) but it's another thing when the reviewers just plain can't get a game to run on a certain hardware/software configuration. Good reviewers will own up to this. Bad ones let it slide, causing collateral implicative damage down the line. Remember: Just because they couldn't get it to run doesn't mean you can't!!

You have to have a motherboard and power supply that supports the video card. So your choice of other parts dictates what video card you get. If you have an AMD machine with a motherboard that supports quad crossfire...Exactly! Every computer system is a unique platform. You have to design using a top-down approach or you'll end up with a hack-job! I've seen plenty of newbies pair up an SLI mobo with an ATI/AMD vid card without even knowing what they've done! If you don't plan to go multi-GPU, then it works out. But if you had planned for CrossFire, guess what!? You're screwed!


There are many other good points that both Hartborne and Variable have made, but it all comes down to perspective. On the one hand I have to agree that performance reigns supreme and this is what separates a good vid card from a great one. On the other hand, you have to keep an eye on the big picture in terms of what make sense given the limitations of mere mortals. A quad-CrossFire mobo necessitates the buying of an HD 3000 series vid card, despite the superior price/performance ratio of the 8800 GT (or GTS 512MB for that matter). You also have to take into consideration that most of today's A-list titles are simply unplayable at 1680 x 1050 with a sub-$200 USD vid card! Which means that most of us will be playing COD 4 and Bioshock at 1280 x 1024 or even 1024 x 768, much less a monster game like Crysis! At this point, it becomes an issue of visual extras. Ie, if you aren't using AA and AF at a minimum of 4x each, you're doing yourself a disservice!! AA and AF are what really separate the HD 3870 and 8800 GT (I'll explain this below).

When it really comes down to it, the best of the best for the price is the 8800 GT, hands down. But if you want scalability... quad-CrossFire with a bunch of HD 3000 series vid cards gets my vote! It's a gamble, but I think it will pay off considering the DX10 debacle. We haven't seen a true SM 4.x title yet, a problem that continues to plague PC gamers everywhere! (Don't even get me started on this one!!! :mad:) The slow, or nearly non-existent, adoption of DX10 means that a bunch of HD 3000 series in CrossFire will continue to pay off in terms of performance without becoming obsolete any time soon. If you look at the history of DX10 parts, you'll see very little progress over the past 18 months, a period of time that has historically given us amazing leaps in the graphics arena!! The relatively ancient 8800 GTX is still one of the top-dogs to this day, despite its late-2006 release!! WTF!? And the GeForce 9's, which are a re-hashed version of the 8's? Again, WTF!? And what about AMD, which has merely caught up with NVidia in that same time period? IMO, the hardware is far ahead of the supporting games available, thus the stagnancy. Heck, by the time we see a fully SM 4.0 title, we'll all be on SM 5.x hardware! This is too slow for my tastes...

[saphalline]

Now, let's all take a trip inside a GPU to see WHY the 8800 GT is better than the HD 3870. (Remember to set your Hearthstone to the main forums page!)

8800 GT

G92, 65nm, 754 M trannies
112 SP's
56 effective TMU's
16/128 max theoretical pixel output per clock
16 ROP's (32-bits each)
600 MHz core
1500 MHz SP's
256-bit RAM bus width
900 MHz GDDR3
57.6 GB/s RAM bandwidth
10500 Mverts/s (processing rate)
42000 Mpixels/s (processing rate)
33600 Mtexels/s (applied rate)
9600 Mpixels/s (fillrate)
168 TFLOPS
DX10, SM 4.0

HD 3870

RV670, 55 nm, 666 M trannies
64 SP's (1+4 per cluster, 320 total)
16 effective TMU's
16/32 max theoretical pixel output per clock
16 ROP's (32-bits each)
775 MHz core
775 MHz SP's
256-bit RAM bus width
512-bit Ringbus width
1125 MHz GDDR4
72.0 GB/s RAM bandwidth
15500 Mverts/s (processing rate)
62000 Mpixels/s (processing rate)
12400 Mtexels/s (applied rate)
12400 Mpixels/s (fillrate)
248/49.6 TFLOPS
DX10.1, SM 4.1

So... notice anything? Obviously there's a lot to chew on here, so let's start... well, heck! We can start anywhere!

The stream processors are the obvious starting point because the counting contention affects unit-to-unit ratios across the board. The 8800 GT has 112 SP's. The HD 3870 has 64 + 256 SP's, or 320 if you take it at face value. Why the confusion? Well, because the SP's in the R600 variants come in clusters of 5. In each cluster, only one SP is a "full" SP. Ie, only one in each cluster of 5 can do the fall gammut of computation. The other 4 are limited to scalar and add-shift ops, or "simple math". This means that the total number of FLOPS for the HD 3870 could be as high as 248 billion, but it could also be as low as 49.6 billion. It depends on the type of math you're doing. Transcendental or multiply-shifts are more intensive and can only be run on 64 of the 320 SP's. To be fair, the number of FLOPS is really only applicable to GP-GPU solutions, but the distinction is important to consider when comparing benchmarks. Sometimes a game is going to pull out a complex shader routine that can run on only 64 of the 320 SP's. This doesn't mean that the other 256 go idle, but it does mean that in such cases, the 112 "full" SP's of the 8800 GT are going to pummel the HD 3870. This accounts for much of the variation in performance between the two, specifically the minimum frame rate of any game. The other implication here involves the access level of SP's to other rendering resources within the GPU, such as the fact that only the 64 "full" SP's have access to the ROP's or Ringbus. It's the little things that add up!

The next obvious stop in the differences would be the TMU's, or rather the effective TMU's. Because of the decoupling of the modern rendering pipeline (at the hardware level) the TMU's are no longer separate entities. As such, they are more or less built into the SP's. Thus the availability of them is much less important than the access or ratio of them. With groups of SP's being tied to clusters of TMU's, the importance of the DX10 pipeline comes into play when thinking of data as a TLB structure. In the case of these two GPU's, the number TMU's is the same (as is the access in DX10 mode) but the number of SP's is not. Moreover, the number of SP's is in contention within one of the GPU's! Thus the 8800 GT's ratio of 2:1 vs the HD 3870's ratios of 4:1/20:1 helps it to overcome the ratio of clock speed between its SP's and core. This is arguably more efficient in some cases, less in others. NVidia took a gamble on fillrate in favor of texturing power. Why? Because games based on older engines still use more textures than shaders! The flip side of this is the reduced reliance on simple shader routine IPC in favor of access power in DX10 mode. This means that NVidia's design is better for older games and future games, but not necessarily the interim games (such as COD 4!).

After reviewing those two aspects, the nature of GPU-wide efficiency now takes a turn towards clock speed. Because of the two approaches to this, a delta has occurred unlike any other. ATI's approach (inherited by AMD) is one of maximum usage. By making the SP's the same clock speed as the rest of the core, ATI banked on keeping the SP's busy as much as possible. Considering their design framework of 1+4 SP clusters, this was a wise choice. NVidia on the other hand banked on reduced bottlenecks due to complexity by jacking up the clock speed of the SP's relative to the core. This reduces their core clock speed potential by quite a bit and also creates more potential for idleness, but it maximizes the minimum frame rate of games by not choking as much on sudden added complexity within a scene. This makes the HD 3870 more efficient by a long shot while not necessarily having as much raw muscle potential - this is inline with ATI's and NVidia's respective approaches, historically speaking.

So how does clock speed tie in the other two in a concrete manner? Access and availability! The HD 3870's 64 SP's that actually have access to the other rendering resources in DX10 mode hurts performance much less on a GPU level because the SP's run at the same clock speed as the rest of the core. This is most important when comparing the number of ROP's (which is the same between the two GPU's) to the clock speed and thus availability relative to the SP's. This means that the access and availability of the ROP's must be weighted by the extra 175 MHz speed that these units receive relative to the 8800 GT core clock speed. The extra 30% boost in speed softens the blow of the larger access ratio, coupled with the fact that there is no clock speed disparity. When everyone is working at the same speed, collaboration is easier! Thus the increased efficiency in the HD 3870 rears its head once again. Meanwhile, the 8800 GT must balance instruction retire rate with fillrate at the front-end because there is a potential that the SP's could get more work done than can be rasterized, creating issuance failures per committed port on the front-end. In practice, the SP's will rarely get stuck with completed unretired instructions in modern and future games, but major hiccups can occur when this does happen (such as older non-shader or shader-lite game engines). Again, the slower core clock speed of the 8800 GT plays a part here.

Another major area of concern related to clock speed is the front-end. The front-end of a GPU contains the decoder, much like a CPU's front-end. However, even more important for rendering is the thread dispatcher! This is the part of the front-end that is responsible for doling out instructions to the SP's, and it is imperative that this is as efficient and effective as possible. Fortunately for rendering, this job is much easier in GPU's than your standard x86 junk, meaning that pipeline flushes are the rare exception in GPU's (despite the increased tendency towards CISC/VLIW with increasingly programmable GPU's). But here we see once again the increased efficiency of the HD 3870. By having a level playing field of clock speeds, and a higher core clock than the 8800 GT, the thread dispatcher is faster and has an easier job. However, the 8800 GT benefits from being able to execute complex instructions far faster. And with its higher-clock SP's, this means a higher minimum frame rate across the board and a higher maximum frame rate under a complex load. That last one is two-fold since the HD 3870 suffers from the 1+4 dilemma of only being able to offer 64 complex SP's at 775 MHz, vs the 112 complex SP's at 1500 MHz of the 8800 GT. In other words, the better front-end leverage of the HD 3870 may be more efficient, but the 8800 GT still has an edge in terms of execution power. Thus even if pipeline bubbles and idleness and flushes rear their ugly heads, the 8800 GT is in a position to recover faster. So in practice, are you going to bank on 30% more core clocks or 93% more shader clocks?

[saphalline]

The ROP's also deserve some limelight here because of the pixel output rate differences. The sheer number of them makes sense from a clock speed perspective, but why is it that ATI eschewed depth and texturing power in favor of fillrate? This is confusing on a number of levels, not the least of which concern the Ringbus and AA implementation! With a 512-bit Ringbus emulating multi-porting topologies, why not up the ante here?? At first glance, you might say that AMD didn't do their GPU revision correctly! 55 nm process vs 65 nm, with less transistors to boot?? WTF!? As I see it, ATI was already late to the game with a DX10 part, and it sucked from the beginning! AMD merely cleaned it up. So the lack of ROP's (and TMU's for that matter) are a limitation of R600 variants in general. It's not like you can wave your magic wand and get 1 FP128 pixel per clock just like that! Not when you start at FP32, anyway. AMD took ATI's design and revamped it to compete, and that's good enough for me! In any case, the Ringbus will be utilized a lot more in future designs, I'm sure. For now, though, the relatively puny ROP's of the HD 3870 severely hinder its efforts to compete better in older games, and the FP64/FP128 performance hits seem out of line with its SM 4.1 spec. Not that there's a formula for this, mind you, by why include a tesselator and not work on pixel depth output?? Even if NVidia's design doesn't necessarily gamble correctly on future code, I gotta give the nod to NVidia for better balance, at the very least. Better texture performance, better pixel processing muscle, and better pixel depth output. Performance reigns supreme!!

As to the AA implementation of ATI's original design, there's no contest here. ATI screwed up from the beginning, and no amount of efficiency can save it here. I've ranted on this before, and I'll continue to rant until it's fixed! Who the heck came up with the idea to hand AA duties to the SP's instead of the RBE's!? WTF!? If you didn't understand this, don't worry about it. Suffice to say that the HD 3870 suffers terribly (and needlessly, I might add!) from high taps of anti-aliasing. If you've ever wondered why turning on AA on an ATI/AMD DX10 part makes your frame rate tank, this is why! Supreme win for NVidia here based on the fact that they didn't screw up! Moving on...

The RAM in general. So much to say, so few chars left! AMD wins here, but not for the same reasons you might expect. The HD 3870 wins here because it uses GDDR4, which is able to scale to higher clock speeds, has reduced latencies, runs cooler, and uses less energy than GDDR3. The fact that it's connected at 256-bits to a Ringbus that operates internally at 512-bits doesn't hurt, either - but considering the Ringbus is underutilized, this alone cannot grant a win to the HD 3870. The RAM bandwidth differences are of little concern.

Read that last part again...

That's right, people! The extra 25% RAM bandwidth enjoyed by the HD 3870 means about as much as a coin on the ground! It's nice to have but it won't solve your problems! In this case, there actually is no problem. Neither of these vid cards suffer from RAM bandwidth limitations. In fact, the primary reason for truncating the 320/384-bit RAM bus width of the older 8800's to the newer ones is because the extra RAM and bandwidth cost more but didn't improve performance! Vid cards currently do not make good use of RAM bandwidth over 60 GB/s. Once you reach that, it's better to reduce latencies (such as GDDR4 vs GDDR3), invest in PCIe 2.0 over 1.x (all new vid cards), or tack on more RAM! Think about it: why is it that the 8800 GTS 320 MB and 8800 GT 256 MB can't hack it for high-end gaming? Is it because they don't have enough RAM to begin with? Bingo! Minimum for modern gaming is 512 MB of graphics RAM at 50-60 GB/s with a PCIe 2.0 interface. Done. Pack your bags.

As far as gaming performance goes, I think that just about covers it. Any questions? (Keep in mind I won't be back for a week! :p)

[George Hallam]

Wow very informative, i had my coffee and wide eyes :p

well i know i have learned something today ;)

[Variable]

I've read as much prior to buying a 3870 during a 12 month on and off again reasearch project on a new machine. How depressing was the 2600 series...

I went ahead and got a better grade of motherboard, memory, and PSU - fully anticipating a CPU upgrade an additional video card in the future. To be honest, I got tired of waiting to build a new machine. I wanted something that would be kind to upgrades and dole out good HDTV. Nvidia doesn't offer quad card support and the new 3800's do well with multiple monitors. Until I bought a new 24 inch LCD I used two and sometimes three.

None of the cards out early in the year were playing DX10 games all that well. So rather than wait until that elusive next big card, I am betting on multi GPU scaling to keep my machine viable for at least 4 years. I'm pulling for AMD, we need competition to keep prices down. I've been building my machines for about 15 years now. I'm over the latest and greatest fad.

[Heartborne]

One really good thing about ATI is the DX 10.1 support. There's apparently quite a bit of controversy about this because Assassin's Creed looks better with an ATI card due to improvements to the Xbox 360 video emulation included in DX 10.1.

What is really sad is that, according to a recent article in PC Magazine, AMD's market share is way down... they're losing a lot of business. If AMD and ATI don't step up their game(s) soon they're looking at a good deal of trouble.

While I always argue that the Nvidia cards offer better performance, there is some bias there due to personal preference because I've been using Nvidia since my first ever video upgrade on my old Dimension L866R... a geforce 4 pci card. I'm used to the driver, the interface and I'm familiar and comfortable with it.

[jlreich]

there is some bias there due to personal preference
No, we haven't noticed at all. :p :D I was starting to think you work for either Intel or nVidia. Or were fired by AMD. :D

[Heartborne]

Yes, given all that I've decided to go ahead and play down the negative comments about ATI/AMD. They're a fine company that is having a hard time right now and I hope they pull themselves out of the rut that they're in and really blow us away with their next line of products.

After all, nvidia is being pretty lame with the whole "let's stick a few more chips on the 8 series and call it the 9 series" fiasco.

[jlreich]

The stream processors are the obvious starting point because the counting contention affects unit-to-unit ratios across the board. The 8800 GT has 112 SP's. The HD 3870 has 64 + 256 SP's, or 320 if you take it at face value. Why the confusion? Well, because the SP's in the R600 variants come in clusters of 5. In each cluster, only one SP is a "full" SP. Ie, only one in each cluster of 5 can do the fall gammut of computation. The other 4 are limited to scalar and add-shift ops, or "simple math". This means that the total number of FLOPS for the HD 3870 could be as high as 248 billion, but it could also be as low as 49.6 billion. It depends on the type of math you're doing. Transcendental or multiply-shifts are more intensive and can only be run on 64 of the 320 SP's. To be fair, the number of FLOPS is really only applicable to GP-GPU solutions, but the distinction is important to consider when comparing benchmarks. Sometimes a game is going to pull out a complex shader routine that can run on only 64 of the 320 SP's. This doesn't mean that the other 256 go idle, but it does mean that in such cases, the 112 "full" SP's of the 8800 GT are going to pummel the HD 3870. This accounts for much of the variation in performance between the two, specifically the minimum frame rate of any game. The other implication here involves the access level of SP's to other rendering resources within the GPU, such as the fact that only the 64 "full" SP's have access to the ROP's or Ringbus. It's the little things that add up!
If my understanding of SP's is correct, in that they are simplified processors (relatively speaking) that basically take all the data from the other units and pump it out very efficiently, then nVidia went the route of raw horsepower that can just pump out the data, where ATI went the route of more complicated but can really shine with certain things that can take advantage of their architecture? So when we see higher scores for ATI it is because that particular piece of code was able to better utilize the clusters and ring bus versus raw horsepower? Basically thinking of an overall TLB structure nVidia just went with more raw horsepower? {As I read further that seems to be what you are saying}

meaning that pipeline flushes are the rare exception in GPU's (despite the increased tendency towards CISC/VLIW with increasingly programmable GPU's)
In other words, the better front-end leverage of the HD 3870 may be more efficient, but the 8800 GT still has an edge in terms of execution power. Thus even if pipeline bubbles and idleness and flushes rear their ugly heads, the 8800 GT is in a position to recover faster. So in practice, are you going to bank on 30% more core clocks or 93% more shader clocks?
This can't hold out forever though I wouldn't think. Raw horsepower will eventually falter in the face of increasing tendencies towards CISC/VLIW. This makes me think of the pre-Core architecture days when Intel was so focused on raw power/speed and large pipelines, then they finally figured out it wasn't going to keep working anymore. Then again if the hardware can keep up with the power and heat issues raw power and speed does solve most problems. :p

The fact that it's connected at 256-bits to a Ringbus that operates internally at 512-bits doesn't hurt, either - but considering the Ringbus is underutilized, this alone cannot grant a win to the HD 3870. The RAM bandwidth differences are of little concern.
I've always thought the Ringbus was a good/interesting design from the beginning, but I think it is as you mentioned ahead of the current software and not of near as much use as it could be.

In fact, the primary reason for truncating the 320/384-bit RAM bus width of the older 8800's to the newer ones is because the extra RAM and bandwidth cost more but didn't improve performance! Vid cards currently do not make good use of RAM bandwidth over 60 GB/s.
This is a great point. I was very concerned about this when the GT came out with only 256bit. But soon realized that's all that is needed. The rest is just extra baggage.

[saphalline]

None of the cards out early in the year were playing DX10 games all that well.Actually, we don't have any DX10 games yet. All current games that claim to be DX10 are mostly DX9c games with a few DX10 rendering calls thrown in for fun. Crysis, for instance, only uses DX10 for soft shadowing, which is next to impossible to discern in motion. (This is largely why uninformed gamers the world over are exclamining that Vista + DX10 is a useless combo. When 99% of the game's rendering calls are emulated, what do you expect!?)

In terms of future DX10 gaming, as in real games that are released to store shelves with 100% DX10 rendering calls, what we have now won't cut it. Crysis is already a system-killer despite being DX10 lite. What does that tell us? Perhaps that the performance necessary to move into full DX10 territory has not yet materialized? Even with DX10's much improved optimizations over its previous ilk, current hardware across the board is not ready for either DX10 or OpenGL 3.0. It's not. There's not enough power, not enough precision. Not for satisfying the insatiable frame rate lusts of high-end gamers, anyway. No one on the planet can wave a magic wand and give us significantly better visuals with a next-gen rendering API AND better performance over previous games! Not gonna happen!!

So when we see higher scores for ATI it is because that particular piece of code was able to better utilize the clusters and ring bus versus raw horsepower? Basically thinking of an overall TLB structure nVidia just went with more raw horsepower?Basically, yes. AMD's GPU's are designed to waste fewer resources. If that means falling behind in performance sometimes, then so be it.

Also keep in mind that each SP may be modeled after a simple single-ported vector execution unit, but the changes are significant enough to warrant a large academic departure. This isn't just semantics! While entirely RISC-based and dependent upon specific data input types, each SP (complex or not) can do so much more than just "slide and spit"! There's a lot of power locked away inside each SP, believe me...

This can't hold out forever though I wouldn't think. Raw horsepower will eventually falter in the face of increasing tendencies towards CISC/VLIW. This makes me think of the pre-Core architecture days when Intel was so focused on raw power/speed and large pipelines, then they finally figured out it wasn't going to keep working anymore.Well, interesting you should say that! What trend do you see with modern vid cards? Bigger, faster, hotter - right? The amount of power required to fuel a modern gaming system is outrageous! Many hundreds of Watts? When did that happen!? Take a look at the "lean" 8800 GT and tell me why we need an expansion card that guzzles almost 150W? Now tell me why we consider this to be "green"?

The performance we're getting per Watt is certainly higher than technology from yester-year, but at the same time we're also in the age of energy-guzzling computers! Remember when 75W was more than enough to power your shiny new 486-class system? What about the golden age of sub-$1000 gaming computers that required "only" 300W? Now we can't recommend less than 500W for any new system build and that's just the starting point! Is it NetBurst all over again? The CPU industry cleaned up. It may be time for the graphics industry to do the same thing.

I've been using Nvidia since my first ever video upgrade on my old Dimension L866R... a geforce 4 pci card.Woah! :eek: Yer a young'un! :p

Be careful about biases. They can lead you down the wrong path if you don't watch your step. When it comes to PC gaming, you have to err on the side of caution - most of us have been into computers since loooong before shaders! Heck, the really old farts around here pre-date the concept of out-of-order execution! ;) I myself first cut my teeth on games back in the days of arcades and the Atari 2600. Then it was NES, Sega, SNES, Sega Genesis, early text-based PC games, Gameboy, Game Gear, N64, Win95 gaming... blah blah blah. Over 20 years of pixels and vectors. It never gets old! :cool:


nvidia is being pretty lame with the whole "let's stick a few more chips on the 8 series and call it the 9 series" fiasco.They've been doing this for years...

Anyone remember the TNT2 Pro vs the TNT2 M64? How about the infamous GeForce4 MX series? NVidia has done this crime before, but even in the face of the 9's vs the 8's, I still hold the GF4 MX in the lowest esteem! That was a dirty trick for post-2000 releases! :mad:


One really good thing about ATI is the DX 10.1 support. There's apparently quite a bit of controversy about this because Assassin's Creed looks better with an ATI card due to improvements to the Xbox 360 video emulation included in DX 10.1.Ahahahahaaaa! Oh man, that sh*t is just too funny!! After reading this little snippet, I had to do some research on the topic. What I found was interesting...


1) - First of all, there is no image quality difference in Assassin's Creed with the extra DX 10.1 support. Or at least there shouldn't be. The game's use of MSAA acceleration is incorrect, and only results/resulted in image quality differences because the calculations were "wrong". Rumours abound, especially considering NVidia's stake in the matter, but the fact remains that the shortcut in MSAA implementations offered by DX 10.1 only affect performance and not image quality.

Moreover, the forms of MSAA used in the game engine are suitably archaic by PC gaming standards. This is to be expected from a port. :rolleyes: Even at "maximum quality", or whatever ambiguous settings the in-game menu lists, AA and AF are limited to just 4x each!! :eek: For AF, that's just a crime for a modern game! (And one of the reasons I won't play ports anymore.) 8x and 16x AF offer noticeably better image quality for outdoor games such as Assassin's Creed (nevermind that AF is completely broken using ATI/AMD parts! :mad:) and would have been a welcome addition for those of us with a high-end DX 10.x part! And to limit the AA to just 4x MSAA and not include some sort of fragment-based algorithm!? No wonder the visuals don't look next-gen!! Heck, at that point, they probably run depth-of-focus on the CPU instead of inside the unified shader units on both platforms! Stupid ports...


2) - Secondly, there is no "emulation" of the Xenos GPU included in DX 10.1. The reason is simple, and founded in set theory: Xenos was the original prototype unified shader design. Therefore, no emulation is necessary. All of the features offered by Xenos are included in DX 10.1 and DX 10. There are also more features included in DX 10/10.1 that aren't included in Xenos because it was a DX10 prototype. So anything that the X360 can render can also be rendered by any PC gamer with Vista and a DX10 part. Because of Xenos's sub-DX10 rating, it is technically only fully capable of DX 9.0c/SM 3.0 with unified shaders thrown in merely for performance advantages. Thus the cry from console fanatics that DX10 is not needed, both because they can't run it :p and it offers no visual enhancements (again, we haven't seen a true SM 4.x title yet, so they don't know what they're talking about). As it stands, no console is capable of SM 4.0 or higher, which leads into my third point...


3) - Thirdly, the attempts by the game programmers to use DX 10.x is laughable! The only reason 10.1 is used at all on ATI/AMD hardware is the accelerated MSAA, not the more unique abilities of the programmable tesselator! In all other respects, visual acuity seems to suffer greatly on ATI/AMD hardware as evidenced by the severe lack of decent aniso and texture resolution. For crying out loud, is there anything that is rendered with basic DX10 calls?? It doesn't look like it, as even something as basic as volumetric shadows is mysteriously missing, to say nothing of pixel-shader texture wraps! What you end up with is a game engine that makes a few DX 10.1 calls for MSAA, and the rest using plain old DX9c.

So... what's the point of playing this game on the PC again?? Yet another graphically-disappointing port. Gee, what a surprise! :rolleyes:

[Heartborne]

Well, for someone like me who refuses to shell out $450 for a game console and an additional $1000 for a tv worth playing it on,the point of playing such a game on the pc is to play it at all. My console gaming life began with the NES and ended with the PS2, the last game system that was made for "the working man" and not built as the leisure of "the elite" who have that kind of money to burn.

Anyway, I have seen some photos of games being run in DX10 next to DX9c, and I saw little if any difference at all. Tell me Saph, are there any real improvements in DX10? Are there any examples of a game that actually looks or runs significantly better in DX10? I resist the move to Vista for myriad reasons, and DX10 was one of the elements tipping the scales... between DX10 support and getting all of my 4 GB of RAM I as very close to just moving up to 64-bit vista... however, seeing that DX10 has yet to show me any actual improvement in video rendering and I've yet to break 2 GB of RAM usage, I don't see any reason to actually move to the well-documented headache that is Vista!

[saphalline]

Well, for someone like me who refuses to shell out $450 for a game console and an additional $1000 for a tv worth playing it on,the point of playing such a game on the pc is to play it at all.A valid point if you do not already own such a TV/display. Those who play consoles almost exclusively oftentimes have more than just one reason to own a TV/display of that magnitude! ;) Besides, the prices have been coming down for over a year now. Not that they're cheap yet, but affordable for most.

I have seen some photos of games being run in DX10 next to DX9c, and I saw little if any difference at all.The point is that games are not DX10 yet. That's the crux of the issue. (DX10 can run DX9, albeit in an emulator.) What you've seen so far is DX9 games being run on Vista w/DX10, but using its DX9 emulator. DX10 cannot magically enhance a game. If DX10 isn't used, then you won't get DX10 visuals.

I won't re-hash my DX10 thread here. (Not much point anyway since no one on the forums seemed to want to actually learn about Vista & DX10. They all seem content to hate M$ simply because they like to hate M$. :rolleyes:) So if you want the details, look there. If you want an example of what DX10 has to offer, this (http://www.futuremark.com/products/3dmarkvantage/) is all we have. Ironically enough, 3DMark Vantage is the only piece of DX10 software in existence. Games that claim to be DX10 only make a few SM 4.x rendering calls. Like I said before, this means that even so-called "DX10 games" currently consist of over 95% DX9c code.

Is it any wonder you aren't impressed yet!?

[Heartborne]

Ah, so as usual in the technology world, we're getting a product that isn't even supported properly yet but is released anyway. I'm guessing microshaft pushes DX10 as a reason to buy vista.
At this point, I've heard more about Vista being pain in the ass than anything else, so I don't think I'll move along to it until the trouble of having the OS is worth the benefit.

[saphalline]

I don't see any reason to actually move to the well-documented headache that is Vista!

[...]

I've heard more about Vista being pain in the ass than anything elseYour concerns are certainly warranted, especially when it comes to Vista and anything even remotely productive in nature. :p Older printers don't have updated drivers, the expensive Office 2007 seems to be the only office suite that even pretends to play nicely :rolleyes:, and a myriad of other issues related to sound/audio and the 32-bit/64-bit conundrum.

I cannot comment on many of these things because I do not use them. Not on my main gaming machine, anyway. And therein lies the issue. Vista works perfectly fine for me! Seriously! I don't know where the crashes and lock-ups and blue screens are coming from! I've seen only one BSOD in Vista the entire time I've used it (just passed the 1-year anniversary) and that was due to my X-Fi driver being 3 months out-of-date (completely my fault). I've seen one lock-up in Vista, but that was related to a game that doesn't seem to want to work, EVER! (I've tried WinXP and WINE and the darn thing just seems suicidal!! :mad:) I've seen zero crashes with Vista.

You might think this is impossible at first glance :D especially since I actually USE Vista and don't just let it sit there idle, but I assure you that Vista works wonderfully as a gaming OS, as long as you know what you're doing! (I can say these same things about WinXP in my experience with it as a gaming OS.) What I can logically conclude, then, is that Windows is good at gaming when properly set-up the first time. Other things? Not so much. Or at least mixing business with pleasure seems to take its toll on the poor thing. Not surprising since performance and stability seem to constantly be at odds with each other, depending on the paradigm. If you can separate the two, you'll be much better off. But if you can't spare your system for gaming only, then you'll have to heed the warnings of those who have gone before you. I am fortunate enough to have been building gaming systems for so many years that my previous core can be rebuilt for business while my new upgrades can be used exclusively for pleasure.

So far so good. But YMMV (your mileage may vary). ;)

[jlreich]

I haven't had any BSOD's in the few months I have been using Vista HP64. In fact I really haven't had any issues with it at all and I have hardly touched XP since I installed it.

The only things are a few older programs that are not compatible. But that is to be expected. I use Open Office on my machine and have had no issues. I installed Office XP (2002) on my wifes laptop with Vista HP32 and have had no issues, and she uses it extensively for school.

My wifes laptop (T7250, 4GB DDR2-667) does crash regularly (2-3 times a week or more). But it was set up by the OEM so I am not surprised. She will be finished with here current class shortly and I will wipe it, reinstall fresh, and set it up correctly and I am sure the problems will go away.