sorry if this doesn't belong here but the fanboy war (nVidia vs. ATI) is raging a major selling point of ATI is that it has 256 bit memory bandwidth, opposed to 128 bit on the geForce FX. I was wondering if anybody would care to elaborate on this. I would think that design would be more important in determining performance from the standpoint of your everyday gamer, rather than a number.
-
We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.
128 bit memory vs 256 bit
- Thread starter llkaimq
- Start date
the "bitness" as far as memory pathways determines how wide the channel (or channels) from the GPU core to memory is. A "256-bit" memory pathway is just that, 256-bits wide and allows 256-bits of data to travel down the pathway at once as opposed to 128-bits at once. This provides for theoretically twice the memory throughput. It may not always reach the theoretical peak but it can come close. And with as much data as the GPU needs to send and receive, it's pretty important.
In that case, peak theoretical bandwidth would be the same as having a data path that is double the width. However, since the operating frequency is much faster, you can usually reduce latencies incurred when turning the bus around for reads or writes, because of the greater clock granularity.
Thus, 128 bit DDR SDRAM banks operating at 1000MHz will be faster than 256 bit DDR SDRAM banks at 500MHz. The difference betwen the two is very small though, and will only become apparent where high RW loads are placed on the interface. It is here that bus turn around latencies will show the 1000MHz clock frequency's advantages, assuming latencies are the same number of clock cycles in the two interfaces.
Thus, 128 bit DDR SDRAM banks operating at 1000MHz will be faster than 256 bit DDR SDRAM banks at 500MHz. The difference betwen the two is very small though, and will only become apparent where high RW loads are placed on the interface. It is here that bus turn around latencies will show the 1000MHz clock frequency's advantages, assuming latencies are the same number of clock cycles in the two interfaces.
128bit@1000MHz vs 256bit@500MHz
same peak theoretical speeds.
if you have a continous stream of packets that are 80bits, then the 128bit will be faster since 80bits fits in 128bit, and the 128bit is working at a higher frequency. if you have a continous stream of packets that's 200 bits, there will be a question as to which is faster. sure the 128bit is still running at a higher frequency, but 200bits does not fit into 128bit. the management of the 200bit packet within the 128bit pipeline will determine the winner. obviously you need 2 cycles to manipulate the 200bit packet in the 128bit pipeline, but all you need is one cycle in the 256bit. of course if you consider a 300bit packet... you will need 3 cycles for the 128bit, and 2 cycles for the 256bit pipeline.
you have to remember that there are latencies involved in both. and latency will determine who the real winner is. and if anything any wasted cycles will be more profound in the slower frequency despite pipeline width.
same peak theoretical speeds.
if you have a continous stream of packets that are 80bits, then the 128bit will be faster since 80bits fits in 128bit, and the 128bit is working at a higher frequency. if you have a continous stream of packets that's 200 bits, there will be a question as to which is faster. sure the 128bit is still running at a higher frequency, but 200bits does not fit into 128bit. the management of the 200bit packet within the 128bit pipeline will determine the winner. obviously you need 2 cycles to manipulate the 200bit packet in the 128bit pipeline, but all you need is one cycle in the 256bit. of course if you consider a 300bit packet... you will need 3 cycles for the 128bit, and 2 cycles for the 256bit pipeline.
you have to remember that there are latencies involved in both. and latency will determine who the real winner is. and if anything any wasted cycles will be more profound in the slower frequency despite pipeline width.
also remembeer, the human eye cant see beyond 75 fps refresh rate. anything above that
is invisible to the eye.
tv, for ex, is 30 fps interleaved, the equivalent of 60 fps
non-interleaved. so if both vid cards are refreshing above
75 fps, there will be no visible difference.
is invisible to the eye.
tv, for ex, is 30 fps interleaved, the equivalent of 60 fps
non-interleaved. so if both vid cards are refreshing above
75 fps, there will be no visible difference.
You have some things mixed up I think. Refresh rate and framerate are different. I can perceive flicker over 75Hz (not at 85Hz), but I dont' think I can perceive anything extra with a framerate of over 60. Anyway, the point of fast cards is to be fast at really high settings, so you get 60 fps at 4xFSAA, 1600x1200, 32 bit color, highest settings and so on.
i consider the original post to be off topic, but i will interject anyway... it is true that 60 fps or 75 fps will allow for smooth frames. however, it is very stupid to assume that an average fps will extend to a min fps. even with an average fps of 60 or 75, there will be extreme scenes where the fps at some discrete time will drop to 30 or less. there is no way to guarantee a fps of at least 60, since there will be some scenes where it will dip below that. of course, you assume you can run quake 2 at the lowest settings, and lowest resolution, and have at least 60 fps with a gf4 4600 or a rad 9700.
Exactly. You can turn off vsync to let more frames be rendered, but if you are at, say, 60hz and render 180fps, then most likely each third of the screen would be showing a different frame, but each pixel on the screen still only changes 60 times per second.
I think Mday answered the original question above, but I think he was calling your post the "original post" off topic.
i answered the 128 vs 256 bit question to the best of my ability without being bullsh!tish about it.
there is no real answer since the frequencies are different. given set A, there is an answer to the question, but given set B, the answer may very well be different. however, i will say that in all practicallity, the 128bit will win by shear frequency concerns, where lost cycles (no matter how they are lost) have much less of an impact.
if you want to discuss a pure bit width discussion, 128 bit at 100MHz may not be much slower, if at all, than 129bit at 100MHz. but that's neither here nor there ;-)
--
as for off topic, i refered to the post where refresh rates and frames per second were first mentioned. "i" consider the topic to be about pure memory pipelines... width\freq.
there is no real answer since the frequencies are different. given set A, there is an answer to the question, but given set B, the answer may very well be different. however, i will say that in all practicallity, the 128bit will win by shear frequency concerns, where lost cycles (no matter how they are lost) have much less of an impact.
if you want to discuss a pure bit width discussion, 128 bit at 100MHz may not be much slower, if at all, than 129bit at 100MHz. but that's neither here nor there ;-)
--
as for off topic, i refered to the post where refresh rates and frames per second were first mentioned. "i" consider the topic to be about pure memory pipelines... width\freq.
zephyrprime
Diamond Member
256bit is better. Some people here have said that "128bit@1000MHz vs 256bit@500MHz same peak theoretical speeds." which is true but misleading. Neither ATI or NVidia make memory chips. They use whatever the memory manufacturers like samsung, hynix, infineon, or micron can manage to make. So it's not going to be 128bit@1000MHz vs 256bit@500MHz on the R350 or NV30. It's going to be 128bit@1000MHz vs 256bit@1000MHz or something kinda like that. Of course, even with half the bandwidth, the NV30 could still make up the difference by being faster in other areas but it is going to be quite difficult.
SDRAM modules are 64-bits wide. To feed a 128-bit or 256-bit data channel, they have to be interleaved in order to provide the bandwidth. Whether one uses 2 SDRAM banks and clocks them at 1 GHz or 4 SDRAM banks and clocks them at 500MHz is really irrelevant. And avoiding extremely wide data channels will allow you to achieve higher clockrates due to less data skew and strict timings.
kylebisme
Diamond Member
true, too big of a data bus and you are looseing preformace instead of gaining it. however, i don't think 256bit is too big in anybody's mind, espesialy with anisotropy and anti-alising. as for the the fanboy arguments, take them with a grain of salt and wait for real reviews. one little feature does not make any product a winner; while it is true that the larger memory bus on the 9700 is advantages, to say how that will alow it to preform against a differnt design with a smaller bus is pure speculation.
BFG10K
Lifer
As has been stated before, a 256 bit memory path generally allows for double the effective bandwidth over 128 bit memory path assuming that the clock speeds are identical. In one clock cycle double the data can be transfered on a 256 bit bus than can be transferred on a 128 bit bus. Of course in practice actual memory bandwidth on video cards can't simply be calculated based on memory specs alone because HSR algorithms, crossbar controllers and caches all increase the practical memory bandwidth available.
Based on pure specifications alone it looks like the GeForce FX will beat the 9700 Pro in fillrate limited situations but lose in memory bandwidth limited situations. Of course we don't know how much colour compression will help nVidia in non-FSAA situations, nor do we know if LMA IV exists which improves on the previous versions so really anything could happen when the final benchmarks are run.
Obviously though 120 FPS on a 120 Hz monitor is the best option of the three.
Based on pure specifications alone it looks like the GeForce FX will beat the 9700 Pro in fillrate limited situations but lose in memory bandwidth limited situations. Of course we don't know how much colour compression will help nVidia in non-FSAA situations, nor do we know if LMA IV exists which improves on the previous versions so really anything could happen when the final benchmarks are run.
That comment is wrong on multiple levels. Even if you implicitly implied that refresh rate loosely equals framerate (which I don't think you did) it's still wrong.
The comparison of TV/movies to 3D games is also flawed.
No, the floating point precision is FP16 (64 bits) and FP32 (128 bits). This has nothing to do with the memory architecture but rather with the precision of the core, data caches and rendering pipelines on the card.
Incorrect. There is a difference between say, 120 FPS and 60 FPS on a 60 Hz monitor. You can still see the effects of partially drawn frames and the most obvious gain is the improved mouse sensitivity and reponsiveness which you get from them.
Obviously though 120 FPS on a 120 Hz monitor is the best option of the three.
Shalmanese
Platinum Member
Hmm... Just to add something interesting to the discussion. Some of you might remember way back in the days of yore that the GF2MX was specced to be equipped with either a 128bit SDR path or a 64bit DDR path. Of course, this meant that the theoretical bandwidth was exactly the same and DDR cost a bit more $$$ so 99% of mfgs went for the 128 bit SDR RAM. However, Creative wen't for the 64bit DDR, ostensibially for performance reasons but, more cynically, for marketing to the stupid.
Now, the GF2MX was a memory bandwidth starved card if there ever was one, it basically increased performance perfectly lineraly with overclocking the memory and even an extreme core clock never gave it more that 1% performance increase. The perfromance increase the Creative card had over other GF2MX cards was 3 - 5% is best case scenarios. Now, the GF GX is arguably far less memory dependant that the GF2MX. Thus, I would estimate that a 1Ghz 128bit GF FX would maybe have an extra 2% performance increase over a 500Mhz 256bit model. Coupled with the fact that the ATI has significantly MORE than 500Mhz of memory bandwidth, I would say that internal compression is all that the GF FX has going for it.
Now, the GF2MX was a memory bandwidth starved card if there ever was one, it basically increased performance perfectly lineraly with overclocking the memory and even an extreme core clock never gave it more that 1% performance increase. The perfromance increase the Creative card had over other GF2MX cards was 3 - 5% is best case scenarios. Now, the GF GX is arguably far less memory dependant that the GF2MX. Thus, I would estimate that a 1Ghz 128bit GF FX would maybe have an extra 2% performance increase over a 500Mhz 256bit model. Coupled with the fact that the ATI has significantly MORE than 500Mhz of memory bandwidth, I would say that internal compression is all that the GF FX has going for it.
the memory on video cards has been 128bit for years, and now they are using 256bit. matrox started the 128bit with interleaving 2x64 bit, and matrox also started 256bit with 2x128bit. now video cards like the ati 9700 is fully 256bit. and yes, the system RAM is still 64bit, but the original post refers to video cards. if you take a look at a comparison between the 9500 and the 9700, you will notice that even with a higher clock speed, and a wider bus, the 9700 is not THAT much faster than the 9500. the reason 256bit memory will have any effect on speeds is that video textures are fairly large (unless compressed, which they actually are), and that's pretty much what is stored on video memory for the most part. i'd prefer 128bit at 2x the speed than 256bit, but that's NEVER going to happen.
With proper compression, 128bit @ 100MHz can be QUITE faster than 256bit @50MHz, much more so than just "2%" maybe 50% to 80%.
With proper compression, 128bit @ 100MHz can be QUITE faster than 256bit @50MHz, much more so than just "2%" maybe 50% to 80%.
This GF2MX analogy is incorrect.
The creative 64bit DDR card was slower than the 128bit SDRAM card by ~4% at comparable "theoretical" bandwith because of additional latency in the DDR system.
Now I don't know how the latency's between the NV30 and R300 implimentations compare, but considering that the two memory subsystems are completly different and that there are hundreds of factors which affect memory performance...
Let's wait for the reviews!
BFG10K
Lifer
The performance delta isn't as high as expected because you're not looking at high enough detail settings, because of ATi's efficient FSAA and anisotropic algorithms and because of the simple fact that most of today's games are CPU limited, even at 1600 x 1200 x 32.
Try running the UT2003 flyby benchmarks at 2048 x 1536 x 32 with 16x anisotropic, with the highest FSAA possible and combined with a 3.06 GHz P4 and the performance gap will be huge.
A wider memory bus helps in many more situations that just moving texels.
Interesting since GeforceFX should be out the same time Hammer is out so I guess will solve the CPU limited problem shortly. Even if hammer isnt out Barton will help things out considerable! I think that ATI R350 can be very comparable to the GeforceFX that is if it also uses some sort of DDR 2 and .13 micron process.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 24K
-
-
