About GPU memory bus.

[Winterlight]

I'm not tech expert but I'm really not understand why a lot people spread a bit misleading information about GPU memory bus ?
Especially than talk about 4070 Ti and it's 192 bit bus.  Why so many people say that memory bus affect GPU performance ? Sure it affect but is not that simple as it look. Memory bus it self don't have impact on GPU performance only memory bandwidth actually can affect performance that is depends from bus and memory type, frequency. That is really misleading say that 192 bit is bottlenck in 4070 Ti even that GPU have 192 bit bus it still have slightly higher memory bandwight compred too 3070 Ti.

Some example

4070 Ti - 192 bit - 504 GB/s

3070 Ti - 256 bit - 448 GB/s

4080 - 256 bit - 716 GB/s

7900 XT- 320 bit 800 GB/s 

That show that GPU in lower bus sometime still have higher memory bandwidth. Especially upcoming GDRR7 will be massive improvent we will  probably see than 128 bit bus give nbetter menmory basndwight than odl GPU with GDDR6X with 256 bit for example.

[Blqckqut]

3 minutes ago, Winterlight said:

I'm not tech expert but I'm really not understand why a lot people spread a bit misleading information about GPU memory bus ?
Especially than talk about 4070 Ti and it's 192 bit bus.  Why so many people say that memory bus affect GPU performance ? Sure it affect but is not that simple as it look. Memory bus it self don't have impact on GPU performance only memory bandwidth actually can affect performance that is depends from bus and memory type, frequency. That is really misleading say that 192 bit is bottlenck in 4070 Ti even that GPU have 192 bit bus it still have slightly higher memory bandwight compred too 3070 Ti.

Some example

4070 Ti - 192 bit - 504 GB/s

3070 Ti - 256 bit - 448 GB/s

4080 - 256 bit - 716 GB/s

7900 XT- 320 bit 800 GB/s 

 

honestly i dont even think nvidia knows what they are doing

 

[Agall]

7 minutes ago, Winterlight said:

I'm not tech expert but I'm really not understand why a lot people spread a bit misleading information about GPU memory bus ?
Especially than talk about 4070 Ti and it's 192 bit bus.  Why so many people say that memory bus affect GPU performance ? Sure it affect but is not that simple as it look. Memory bus it self don't have impact on GPU performance only memory bandwidth actually can affect performance that is depends from bus and memory type, frequency. That is really misleading say that 192 bit is bottlenck in 4070 Ti even that GPU have 192 bit bus it still have slightly higher memory bandwight compred too 3070 Ti.

Some example

4070 Ti - 192 bit - 504 GB/s

3070 Ti - 256 bit - 448 GB/s

4080 -256 bit -716 GB/s

Memory bandwidth can matter especially at 4K. If you look at the 7900 XTX and RTX 4090's performance at 4K, it gaps the latter disproportionately for a reason, especially the RTX 4090. Even with a light memory OC, an RTX 4090 is getting over 1TB/sec memory bandwidth with its 384 bit bus.

 

Another thing to remember with the RTX 4090's 4K performance gap is that the RTX 4090 is so far the only proportionally binned GPU of the 4000 series, where the RTX 4080 is proportional to last generation's RTX 3070ti, and the RTX 4070ti being somewhere inbetween the RTX 3060 12GB and RTX 3060ti. Nvidia is basically upselling their GPU binning scheme for everything but the RTX 4090, being a slightly chopped AD102 quite equivalent to the slightly chopped GA102 that Nvidia used for the RTX 3090. A proprotionally binned RTX 3000 series to RTX 4000 series GPU should reach ~60% performance increase, not the 20-40% we're seeing with Nvidia's upsold binning scheme.

 

If you're not trying to run 4K at high framerates, I wouldn't worry about memory bandwidth.

[Middcore]

5 minutes ago, Winterlight said:

Why so many people say that memory bus affect GPU performance ?

Because it does. As you yourself immediately admit:

 

5 minutes ago, Winterlight said:

Sure it affect

 

The rest of your post is just arguing that the memory bus in one specific card isn't as much of a handicap as some people think it is. 

[Winterlight]

1 minute ago, Blqckqut said:

honestly i dont even think nvidia knows what they are doing

 

I think most people just spread misleading information. It show that 192 bit bus have better bandwight than old gen 256 bit.

[YoungBlade]

I think the issue is that people know that bus width can make a difference, and they see that the 4070Ti has a lower number than previous generations of cards, and they are already mad about how poor value the card is, so they jump and shout and point at that lower number to justify their anger. Which is unnecessary, they just need to point at the price/performance ratio.

 

Bus width as a variable can make a difference, but as you point out, if the memory speed is different, then you're missing an important factor that impact total bandwidth. Ultimately, that total bandwidth is what makes the really difference.

[WereCat]

Because at higher resolutions the extra cache can't compensate for the memory bandwidth well enough so memory bus matters.

 

But only apples to apples comparison where only the memory bus is the difference (VRAM amount aside, does not really matter in those benchmarks as both 8GB and 12GB are enough) are the RTX 306012 GB vs RTX 3060 8GB.

 

 



 

 

This is avg over 12games but on some the games tested the 8GB with 128-bit buss is almost 30% slower.

[Winterlight]

2 minutes ago, Middcore said:

Because it does. As you yourself immediately admit:

 

 

The rest of your post is just arguing that the memory bus in one specific card isn't as much of a handicap as some people think it is. 

Higher is better but direcllty memory bus not affect preformance only memory bandwight that is not always mean lower bus less bandwight.

[Agall]

1 minute ago, YoungBlade said:

I think the issue is that people know that bus width can make a difference, and they see that the 4070Ti has a lower number than previous generations of cards, and they are already mad about how poor value the card is, so they jump and shout and point at that lower number to justify their anger.

 

Bus width as a variable can make a difference, but as you point out, if the memory speed is different, then you're missing an important factor that impact total bandwidth. Ultimately, that total bandwidth is what makes the really difference.

Well the RTX 4070ti is not a xx70ti class card with respect to last generation. The way I measure this is taking the RTX 3090 and RTX 4090 as respective maximum values since they're equivalently binned xx102 GPUs that are comparable intergenerationally. Take that as a proportion with the lesser cards of the generation, whether its die size and/or shader count, and you'll see why the RTX 4090 is the only good card of the generation so far.

[Winterlight]

Well I just want say that is missleading say that bus directly affect preformance higher bus affect total memory bandwight but is not all time mean that higher bus give you more bandwight than lower bus GPU.

It's good example 4070 Ti with lower bus still have higher total memory bandwight compred too 3070 Ti 256 bit bus.

[WereCat]

4 minutes ago, Winterlight said:

Higher is better but direcllty memory bus not affect preformance only memory bandwight that is not always mean lower bus less bandwight.

That's like saying that memory speed does not affect memory bandwidth if you have large enough bus

 

Also you can't compare different architectures because memory compression may work differently

[Winterlight]

3 minutes ago, WereCat said:

That's like saying that memory speed does not affect memory bandwidth if you have large enough bus

Well memory bus, type and frequency is affect total memory bandwight that is affect GPU preformacne, but out of context can't say that higher bus is better due example show that GPU with lower bus in some case can have higher bandwight.

[Agall]

4 minutes ago, Winterlight said:

Well I just want say that is missleading say that bus directly affect preformance higher bus affect total memory bandwight but is not all time mean that higher bus give you more bandwight than lower bus GPU.

It's good example 4070 Ti with lower bus still have higher total memory bandwight compred too 3070 Ti 256 bit bus.

Doesn't 4000 series use a higher bandwdith form of DDR6x than the 3000 series?

[WereCat]

Just now, Winterlight said:

Well memory bus, type and frequency is affect total memory bandwight that is affect GPU preformacne. But out of context can't say that higher bus is better due example show that GPU with lower bus in some case can have higher bandwight.

Your example with RTX 3070ti vs 4070ti can't be compared as well because RTX 4070ti has much bigger cache to compensate for the lower memory bandwidth.

 

Also check my 1st reply in the Spoiler tag.

[Agall]

13 minutes ago, YoungBlade said:

I think the issue is that people know that bus width can make a difference, and they see that the 4070Ti has a lower number than previous generations of cards, and they are already mad about how poor value the card is, so they jump and shout and point at that lower number to justify their anger. Which is unnecessary, they just need to point at the price/performance ratio.

 

Bus width as a variable can make a difference, but as you point out, if the memory speed is different, then you're missing an important factor that impact total bandwidth. Ultimately, that total bandwidth is what makes the really difference.

Operating based exclusively on price/performance is what gives Nvidia the ammunition to upsell their binning scheme, which is why we have a quite underwhelming AD104 GPU with a 192 bit bus in a xx70ti class card, beside the fact that its $800.

[Winterlight]

3 minutes ago, WereCat said:

Your example with RTX 3070ti vs 4070ti can't be compared as well because RTX 4070ti has much bigger cache to compensate for the lower memory bandwidth.

 

Also check my 1st reply in the Spoiler tag.

Actually 4070 Ti have higher total memory bandwidth with just 192 bit bus while 3070 Ti use 256 bit but still have lower total bandwidth. I'm not understand about what compensate you talk ?

[Agall]

3 minutes ago, WereCat said:

Your example with RTX 3070ti vs 4070ti can't be compared as well because RTX 4070ti has much bigger cache to compensate for the lower memory bandwidth.

 

Also check my 1st reply in the Spoiler tag.

Well the idea of comparing the RTX 3070ti to the RTX 4070ti in itself isn't reasonable. The RTX 3070ti has a much higher quality bin of silicon versus the 4070ti. I'd look to comparing the RTX 3070ti to the RTX 4080, and the RTX 3060 12GB to the RTX 4070ti to have a more comparably representation.

[YoungBlade]

13 minutes ago, Agall said:

Well the RTX 4070ti is not a xx70ti class card with respect to last generation. The way I measure this is taking the RTX 3090 and RTX 4090 as respective maximum values since they're equivalently binned xx102 GPUs that are comparable intergenerationally. Take that as a proportion with the lesser cards of the generation, whether its die size and/or shader count, and you'll see why the RTX 4090 is the only good card of the generation so far.

Why does it matter what the die size or bus width or shader count or any of those specs are?

 

The price to performance isn't there - that's the problem and that's why people are upset.

 

Would you really be mad if they made an RTX 4060 if it had a 32 bit bus and a 12mm² die if it only cost $250 and performed on par with the RTX 3080?

 

2 minutes ago, Agall said:

Operating based exclusively on price/performance is what gives Nvidia the ammunition to upsell their binning scheme, which is why we have a quite underwhelming AD104 GPU with a 192 bit bus in a xx70ti class card, beside the fact that its $800.

 How has demanding more performance put us in this spot? Where is the better price/performance we've been demanding?

[Agall]

2 minutes ago, Winterlight said:

Actually 4070 Ti have higher total memory bandwight with jsut 192 bit bus while 3070 Ti use 256 bit but stil lhave lower total bandwight. I;m not understand about what compensate you talk ?

The RTX 3070ti and RTX 4070ti aren't comparable GPUs because Nvidia is upselling the binning scheme for 4000 series with respect to 3000 series. 

 

Comparable 'close enough' bins of silicon based on proportional shader account adjusted by the RTX 3090/4090:

 

RTX 3090- 936.2 GB/s

RTX 4090- 1,008 GB/s

 

RTX 3070ti- 608.3 GB/s

RTX 4080- 716.8 GB/s

 

RTX 3060 12GB- 360.0 GB/s

RTX 4070ti- 504.2 GB/s

[Agall]

6 minutes ago, YoungBlade said:

Why does it matter what the die size or bus width or shader count or any of those specs are?

 

The price to performance isn't there - that's the problem and that's why people are upset.

 

Would you really be mad if they made an RTX 4060 if it had a 32 bit bus and a 12mm² die if it only cost $250 and performed on par with the RTX 3080?

 

 How has demanding more performance put us in this spot? Where is the better price/performance we've been demanding?

Its at $1600 minimum with the RTX 4090.

 

If they made that RTX 4060, the biggest problem with it is not the die itself, but the fact that they're calling it a xx60 card and selling it for more than last generation while upselling the bin.

 

4000 series problem isn't just the price hike, but also the upsold silicon. Nvidia is double dipping in the gouge.

[WereCat]

4 minutes ago, Winterlight said:

Actually 4070 Ti have higher total memory bandwidth with just 192 bit bus while 3070 Ti use 256 bit but still have lower total bandwidth. I'm not understand about what compensate you talk ?

4070ti has 504.2GB/s bandwidth

3070ti has 608.3GB/s bandwidth

 

4070ti has much bigger cache so it can compensate by loading more data into the cache instead of relying on memory because the cache is significantly faster and lower latency. Similar to how AMD RDNA2 architecture works with it's extra L3 cache except NVIDIA has bigger L2 cache.

Also they fall off on higher resolution where memory bandwidth is much bigger factor as the cache is not large enough to compensate due to the amount and size of  the data it needs to work with.

[Agall]

On 3/20/2023 at 10:42 AM, WereCat said:

4070ti has 504.2GB/s bandwidth

3070ti has 608.3GB/s bandwidth

 

4070ti has much bigger cache so it can compensate by loading more data into the cache instead of relying on memory because the cache is significantly faster and lower latency. Similar to how AMD RDNA2 architecture works with it's extra L3 cache except NVIDIA has bigger L2 cache.

Also they fall off on higher resolution where memory bandwidth is much bigger factor as the cache is not large enough to compensate due to the amount and size of  the data it needs to work with.

Its also a 5nm TSMC vs 8nm Samsung GPU, where that 5nm TSMC can hit 3GHz versus maybe 2200MHz on 8nm Samsung.

[Winterlight]

5 minutes ago, WereCat said:

4070ti has 504.2GB/s bandwidth

3070ti has 608.3GB/s bandwidth

 

4070ti has much bigger cache so it can compensate by loading more data into the cache instead of relying on memory because the cache is significantly faster and lower latency. Similar to how AMD RDNA2 architecture works with it's extra L3 cache except NVIDIA has bigger L2 cache.

Also they fall off on higher resolution where memory bandwidth is much bigger factor as the cache is not large enough to compensate due to the amount and size of  the data it needs to work with.

Sorry my bad 3070 Ti come with higher bandwight than just 3070.

But if rumors will be true 4070 will use same memory as 4070 Ti and than you can compere 4070 vs 3070 that show lower bus in some case still can be better total memory bandwight.

[WereCat]

Just now, Agall said:

Its also a 4nm TSMC vs 8nm Samsung GPU, where that 4nm TSMC can hit 3GHz versus maybe 2200MHz on 8nm Samsung.

Well yes, that's why I am saying this for the 3rd time... check my first reply where you can see direct comparison within the same architecture where the only difference is memory bandwidth and VRAM amount (which does not play the role as it's enough).

[porina]

Is this thread going to offer anything more on this argument that hasn't already been covered in thread linked: