Nvidia back at their memory game - RTX 3080 20GB found and benchmarked

[leadeater]

On 1/30/2021 at 5:39 AM, Kisai said:

The reason certain numbers show up has to do with channels (eg "bits") of memory that can be accessed. The 3080 is listed as 320-bit with 10GB GDDR6X, The 3090 is 384-bit with 24GB GDDR6X, the 3070 is 256-bit with 8GB. 3060 Ti is 256 bit with 8GB GDDR6, and 3060 is 12GB GDDR6 with 192 bit.

 

The less wide the bits are, the narrower that bus is, so that effectively means to access larger amounts of memory, more latency will be induced the more of that memory is utilized. It's like how your PC might operate in single channel or dual channel mode. It doesn't result in the memory speed being cut in half, but it does noticeably impact the performance when the memory is fully utilized.

I'm not sure you've explained that the best. GDDR modules tend to come in two commonly used and mass produced sizes, a wider bus means a greater number of memory modules. A graphics card could be 128 bit or 256 bit and have the same amount of VRAM, one of them would have twice the number of memory modules of half the size. The one with more memory modules would have far greater memory bandwidth while only a very tiny reduction in  memory latency, something which isn't much of a concern to GPUs and the workloads they run.

 

Each GDDR6 module is 32 bits, you can just divide the bus width by 32 to get the number of memory modules, then divide the memory capacity by the number of modules to get the size of module being used. That is if you want to know these two things.

[Kisai]

6 hours ago, leadeater said:

I'm not sure you've explained that the best. GDDR modules tend to come in two commonly used and mass produced sizes, a wider bus means a greater number of memory modules. A graphics card could be 128 bit or 256 bit and have the same amount of VRAM, one of them would have twice the number of memory modules of half the size. The one with more memory modules would have far greater memory bandwidth while only a very tiny reduction in  memory latency, something which isn't much of a concern to GPUs and the workloads they run.

 

Each GDDR6 module is 32 bits, you can just divide the bus width by 32 to get the number of memory modules, then divide the memory capacity by the number of modules to get the size of module being used. That is if you want to know these two things.

Most benchmarks I've seen, be it CPU or GPU, the improvement with wider channels is never double, it's always more like 10-20%. With GPU's often software demands a specific level of bandwidth, so you would really want to avoid the low-end parts.

https://knowledge.autodesk.com/support/civil-3d/learn-explore/caas/sfdcarticles/sfdcarticles/System-requirements-for-Autodesk-Civil-3D-2020.html

Quote

Display Card

Minimum: 1 GB GPU with 29 GB/s Bandwidth and DirectX 11 compliant Recommended: 4 GB GPU with 106 GB/s Bandwidth and DirectX 11 compliant

Even though that says "minimum" and "Recommended" no software, not since 1986 has ever been honest about requirements. Usually the "minimum" is "this is what the software requires, and anything less we didn't test on" and Recommended being "This is the lowest-end thing we tested"

 

As I have a variety of systems at the office, that "recommended" spec is exactly, or near to GP107 (Quadro P500/Geforce GTX 1050), which you might go "oh, well a GTX 1050 is sufficient for AutoCAD", but then if you actually try to use the software, no, it certainly isn't unless you want to hear the GPU fan sound like a jet engine the entire time. The 1050 with a 96bit GDDR5 bus is about 84GB/sec, the 1050/1050Ti with a 128bit bus is 112GB/sec. All the CAD systems have a minimum of a P1000, which have 96GB/sec on a 128bit bus, falling slightly below the Recommended. All of these mentioned are GP107. The actual CAD users get laptops with GP106 or desktops with GP104, which aren't configured with such low memory bandwidth. However even these high end laptops, sound like a jet engine after about 2 minutes running AutoCAD. Desktops don't break a sweat however.

 

Neural Net stuff also demands higher GPU memory bandwidth. So my feeling here is that a 20GB 3060 may really be intended for something like this, but the lack of bandwidth doesn't impair it as much (even the bandwidth difference between a 2060 and a 2080 isn't that much, 336GB/sec vs 448GB/sec) compared to the sub x60 parts. The 3060Ti is 448/GB sec which puts it at the same bandwidth as the 2080 on GDDR6.  The 3080 is 760GB/sec, and the 3090 is 936GB/sec on GDDR6X.

 

Honestly, it actually kinda feels like nVidia didn't intend to release a 3080 10GB at all and was forced to due to the lack of GDDR6X chips. The same strange reason why we're seeing 3060Ti's with 20GB, is probably to soak up the GDDR6 inventory that was for previous 20xx parts.

 

[leadeater]

4 hours ago, Kisai said:

Most benchmarks I've seen, be it CPU or GPU, the improvement with wider channels is never double, it's always more like 10-20%. With GPU's often software demands a specific level of bandwidth, so you would really want to avoid the low-end parts.

That's because memory bandwidth is not directly related to performance, it never has been. Memory bandwidth is only ever about if you have enough for the compute resources. With enough memory bandwidth you get 100% performance, with more you get 100% performance, with not enough you get something below 100%.

 

4 hours ago, Kisai said:

Honestly, it actually kinda feels like nVidia didn't intend to release a 3080 10GB at all and was forced to due to the lack of GDDR6X chips. The same strange reason why we're seeing 3060Ti's with 20GB, is probably to soak up the GDDR6 inventory that was for previous 20xx parts.

Doubt it, Nvidia has never favored high capacity VRAM on Geforce cards ever and has always been criticized for it. Nvidia only ever puts the more expensive memory modules on Titans and Quadros. For all we know there is an abundance of the expensive high capacity GDDR6 modules but that doesn't mean Nvidia would be willing to use them on Geforce cards. 

[X-System]

Now, RTX 3080 Ti will maybe have 12GB G6X...

 

https://videocardz.com/newz/nvidia-geforce-rtx-3080-ti-with-12gb-memory-to-launch-in-april

 

12GB or 20GB ? Wait and see when it will release...

[Dedayog]

On 1/29/2021 at 11:06 AM, germgoatz said:

why cant we just get a 12gb 3080 because 10gb is not enough but 20gb is way too much and a waste of money for most gamers

10GB isn't enough?  When do you break this?  In real use, not allocation.

 

Btw, if 10 isn't enough, I'd want more than 12.  16GB would be a nice step up, not 12.

 

I think Nvidia, while being asshats, know exactly how to get into our wallets.  20GB VRAM?  I bet some 3080 owners are going to sell them and rebuy this.  Strong work, Nvidia.

[germgoatz]

7 minutes ago, Dedayog said:

10GB isn't enough?  When do you break this?  In real use, not allocation.

 

Btw, if 10 isn't enough, I'd want more than 12.  16GB would be a nice step up, not 12.

 

I think Nvidia, while being asshats, know exactly how to get into our wallets.  20GB VRAM?  I bet some 3080 owners are going to sell them and rebuy this.  Strong work, Nvidia.

ive seen many benchmarks for 1440p and 4k where all the ram is being used, thats one of the reasons why the rx 6000 series has more ram than its nvidia counterparts

but i do agree that 16gb would be better

[HenrySalayne]

1 hour ago, Dedayog said:

16GB would be a nice step up, not 12.

1 hour ago, germgoatz said:

16gb would be better

You want even slower transfer rates because they have to step down the interface width to 256 bits? Or something like a 970 with some parts of the memory being slower?

If they want to use the same GA102 for the 3080 TI, they basically have to use either 12 GB with a 384 bit interface, or 20 GB with a 320 bit interface and memory modules on both sides of the card. I would prefer 12 GB. Cooling the modules on the back (like 3090s) seems to be a lot harder. And 12 GB should be enough for now.

 

But it doesn't matter anyway. With the current rise of cryptocurrency we are not going to see any cards in the wild any time soon.

 

[leadeater]

1 hour ago, germgoatz said:

ive seen many benchmarks for 1440p and 4k where all the ram is being used, thats one of the reasons why the rx 6000 series has more ram than its nvidia counterparts

but i do agree that 16gb would be better

That's just ram allocation, it can be next to impossible to tell how much vram is actually in use. You typically have to go by secondary factors like frame time spikes and large loses in 1% and 0.1% lows.

 

Nvidia always been very good at putting just enough vram on their cards to do the intended job, and for the RTX 3080 that includes 4K. The gaming community has been down this path before, multiple times in the past in fact, and it typically turns out that the lesser vram compared to AMD only becomes a factor that gets talked about around 2 years after the release of the GPUs, right around the time Nvidia is about to release a new generation or refresh 

[TVwazhere]

On 1/29/2021 at 11:04 AM, williamcll said:

A man from Bilibili managed to buy himself a 3080 with 20GB of memory from an anonymous person with bitcoins, he benchmarked it:

3080 20GB is in the middle, note doubled VRAM size

TBH, to me the 3080 20GB looks more like a cut down 3090 than a beefed 3080 (yes I know a 3080 is essentially a cut down 3090) but given they both have the same CUDA core count, maybe it should be like the 3090 Base or 3090 SE or 3090 sub 3080ti super rev 2.1a

[CTR640]

2 hours ago, TVwazhere said:

TBH, to me the 3080 20GB looks more like a cut down 3090 than a beefed 3080 (yes I know a 3080 is essentially a cut down 3090) but given they both have the same CUDA core count, maybe it should be like the 3090 Base or 3090 SE or 3090 sub 3080ti super rev 2.1a

Or 3090 Gimped.