[SOLVED] VRAM clocks and its chip speed

PiffPuff

Commendable
Jan 8, 2021
71
1
1,535
Hi folks,

will highly appreciate your comments on the following:

  1. If my research was correct a) higher VRAM clocks provide faster task processing, however b) there's no linear dependency between Flops [#cores * clocks * 2] and FPS that GPU generates, besides c) VRAM bandwidth is one of the key factors too. With all those factors considered, how do these 2 scenarios differ in terms of the number of generated FPS: 1) "less clocks & very fast VRAM (HBM2)"; 2) "higher clocks & fast VRAM (GDDR5X)"? Assuming the # of cores and their architecture are the same.
  2. While GDDR5X is a standard by itself (80-112Gb/s) I've read that Samsung and Micron chips have the max. speed of 16Gb/s, while Hynix chips have 10-14Gb/s. Does this mean that among all the GDDR5X GPUs those Samsung/Micron-based are faster (taken that # of cores and clocks are equal)?
Thank you!
 
Solution
1. If my research was correct a) higher VRAM clocks provide faster task processing, however b) there's no linear dependency between Flops [#cores * clocks * 2] and FPS that GPU generates, besides c) VRAM bandwidth is one of the key factors too. With all those factors considered, how do these 2 scenarios differ in terms of the number of generated FPS: 1) "less clocks & very fast VRAM (HBM2)"; 2) "higher clocks & fast VRAM (GDDR5X)"? Assuming the # of cores and their architecture are the same.
What really matters is how much bandwidth, or how much data can be shoved to and from the GPU at once, the memory system has. You can increase bandwidth in one of two...
1. If my research was correct a) higher VRAM clocks provide faster task processing, however b) there's no linear dependency between Flops [#cores * clocks * 2] and FPS that GPU generates, besides c) VRAM bandwidth is one of the key factors too. With all those factors considered, how do these 2 scenarios differ in terms of the number of generated FPS: 1) "less clocks & very fast VRAM (HBM2)"; 2) "higher clocks & fast VRAM (GDDR5X)"? Assuming the # of cores and their architecture are the same.
What really matters is how much bandwidth, or how much data can be shoved to and from the GPU at once, the memory system has. You can increase bandwidth in one of two ways: increase the clock speed or increase how physical lines of data there are. GDDR based memory only has 32 lines for data transfer per chip, so they must be clocked really fast. For HBM it depends, but for the sake of information it can be 2048 or 4096 lines for data transfer, so it doesn't have to be clocked as fast to achieve the same bandwidth.

The equation for bandwidth is simple: [transfer rate (or "clock speed", but that term is kind of tricky when talking about RAM)] x [number of data lines] = bits per second. You can then divide by 8 to get bytes per second if you like that figure more.

To answer the question, it doesn't matter if they have the same bandwidth.

2. While GDDR5X is a standard by itself (80-112Gb/s) I've read that Samsung and Micron chips have the max. speed of 16Gb/s, while Hynix chips have 10-14Gb/s. Does this mean that among all the GDDR5X GPUs those Samsung/Micron-based are faster (taken that # of cores and clocks are equal)?
Thank you!
Probably, but it doesn't matter in the end. The memory video cards come with are going to be the best, after considering cost and performance, the manufacturer thinks will work for the GPU. So it doesn't matter if you have Samsung or Hynix in a low end card because the low end card won't take advantage of faster memory like a higher end card can. To put it in another way, if a GPU can only crunch on data at a rate of say 100 GB/s in total, there's no point in giving VRAM that can achieve 150 GB/s

It might matter if you're overclocking VRAM, but even then, you can only clock the VRAM so far before performance flatlines or falls off due to error accumulation.
 
  • Like
Reactions: dotas1 and PiffPuff
Solution

PiffPuff

Commendable
Jan 8, 2021
71
1
1,535
@hotaru.hino Thank you so much for such a detailed and technically substantiated reply. May I ask you to comment a bit on the "bandwidth" part, so that it becomes all clear to me:

  1. Do GPU manufacturers advertise the bandwidth at all? I know they market Flops quite actively, but based on the formula you've written bandwidth [transfer rate or clock speed * # of data lines] and Flops [# of cores * clock speed * 2] are 2 different things.
  2. Regarding the formula itself:
    1. is it the VRAM bandwidth that you called "transfer rate"? E.g. 80-112Gb/s for GDDR5X or 128Gb/s for HBM, etc. If yes, should I use the minimum (80Gb/s for GDDR5X) or the average/mean (96Gb/s for GDDR5X) figure?
    2. clock speed: the VRAM clocks or the core clocks? Taken that HBM has 64 (!) times more data lines than GDDR (2048 vs. 32) I can't imagine x64 boost in any clocks to make your bandwidth equation equal for GDDR vs. HBM.
Actually I guess the examples of calculations will explain a lot, smth. like:
  • same # of cores, different clocks - HBM vs. GDDR [to see the influence of clocks only]
  • GDDR with more cores vs. HBM with less cores, different clocks [to see how the core # influences the clocks & memory type battle]
If you kindly find time for these example too, I will be very grateful. The topic is slightly complicated for a novice PC builder, but your explanations are easy to read/understand, so I'm sure I'll manage it with your kindest help. Thank you!
 
Do GPU manufacturers advertise the bandwidth at all? I know they market Flops quite actively, but based on the formula you've written bandwidth [transfer rate or clock speed * # of data lines] and Flops [# of cores * clock speed * 2] are 2 different things.
Yes, they'll advertise pretty much all the important bits about the memory.

is it the VRAM bandwidth that you called "transfer rate"? Eg. 80-112Gb/s for GDDR5X or 128Gb/s for HBM, etc. If yes, should I use the minimum (80Gb/s for GDDR5X) or the average/mean (96Gb/s for GDDR5X) figure?
Yes, VRAM bandwidth is the transfer rate. The transfer rate is dictated by where the RAM is being used and what the manufacturer set the clock speed at. For example, the GeForce GTX 750 Ti is clocked effectively at 5GHz, whereas the GTX 1070 Ti is effectively clocked at 8GHz even though they both use GDDR5. The manufacturer will advertise what clock speed the RAM is running at. Though note, if it's not expressed in GHz, it's expressed in Gbps.

clock speed: the VRAM clocks or the core clocks? Taken that HBM has 64 (!) times more data lines than GDDR (2048 vs. 32) I can't imagine x64 boost in any clocks to make your bandwidth equation equal for GDDR vs. HBM.
VRAM always runs at its own speed. Clock speed has nothing to do with it.

Actually I guess the examples of calculations will explain a lot, smth. like:
  • same # of cores, different clocks - HBM vs. GDDR [to see the influence of clocks only]
  • GDDR with more cores vs. HBM with less cores, different clocks [to see how the core # influences the clocks & memory type battle]
There aren't many actual examples of video cards with the same GPU, but different VRAM setup. The actual VRAM type doesn't really matter (at least as far as I know), just the bandwidth that it can provide. So to that end, here's a review of the GeForce GT 1030 DDR4 version, which includes the GDDR5 version. The gist of it is, at least in this example, reducing the VRAM bandwidth by 2/3rds cuts the performance in half.

As for the second example, we have the GTX 1070 and GTX 1070 Ti having the same VRAM bandwidth, but the 1070 Ti has more GPU cores and performs better. We also have the GTX 1080 and GTX 1080 Ti with a similar story.

Video cards are engineered to meet certain goals. In some cases, it's to get the maximum performance for minimal cost to manufacture with the components available. In others, it's simply to meet a market segment the manufacturer thought was needed to be met. So while yes, VRAM bandwidth is important, it becomes less so after a certain point and you really can't use the specs sheet to determine how this will affect performance. Especially since actual performance is subject to a massive amount of variables that are beyond our control.
 
  • Like
Reactions: dotas1

PiffPuff

Commendable
Jan 8, 2021
71
1
1,535
For example, the GeForce GTX 750 Ti is clocked effectively at 5GHz, whereas the GTX 1070 Ti is effectively clocked at 8GHz even though they both use GDDR5.
The manufacturer will advertise what clock speed the RAM is running at. Though note, if it's not expressed in GHz, it's expressed in Gbps.

OK, you're right - these calculations make small sense, because anyway both the bandwidth and the flops equations result in a certain final figure, which I can use for comparison (within the same GPU architecture of course).

Finally, I have to check for 2 figures: flops (how fast the GPU performs calculations) and bandwidth/transfer rate (how much data it can push to the system at a time). Though the manufacturer may express bandwidth/transfer rate in GHz, makes sense to convert it to Gbps (to have the very final figure). I hope I summarized your explanations correctly. :giggle:
 

PiffPuff

Commendable
Jan 8, 2021
71
1
1,535
@dotas1 Your question puzzles me. I need some general knowledge to be able to compare. Why?

I'm a novice in PC building, but at a certain stage I understood that some technical details which @hotaru.hino kindly highlighted (which I value a LOT) will be extra for my goals and I formulated the simplified approach (based on how I understood hotaru.hino) as follows:

I have to check for 2 figures: flops (how fast the GPU performs calculations) and bandwidth/transfer rate (how much data it can push to the system at a time). Though the manufacturer may express bandwidth/transfer rate in GHz, makes sense to convert it to Gbps (to have the very final figure).

If I got it right, I will be more than happy. Some things are just more complicated than others - forums are to share knowledge and to learn, right?
 
I was asking because if it was just for knowledge, is understandable and commendable.

If you wish to compare two GPUs then you don't need to see those numbers at all. Just look at benchmarks you need, games you play, apps you use, how each GPU performs. For example, you want to compare a 1070 and a 1080ti when playing battlefield 5.
View: https://youtu.be/rv_aTLqiK34


You will search youtube as above and you will get an idea how much better it performs. Of course you will have to look at the hardware that was used, drivers, resolution etc.

You can also get a very good idea of what to expect performance-wise from the GPU hierarchy here:
https://www.tomshardware.com/reviews/gpu-hierarchy,4388.html

As I mentioned, you will always need to make sure you are comparing not just the GPUs but the rest of the hardware to be as close performance-wise as possible.
 
  • Like
Reactions: hotaru.hino
Finally, I have to check for 2 figures: flops (how fast the GPU performs calculations) and bandwidth/transfer rate (how much data it can push to the system at a time). Though the manufacturer may express bandwidth/transfer rate in GHz, makes sense to convert it to Gbps (to have the very final figure). I hope I summarized your explanations correctly. :giggle:
Getting performance metrics isn't that easy still. For example, the Radeon Vega 64 is rated for 11518 GFLOPS and 483.8 GB/sec of VRAM bandwidth. Compare this to the GeForce RTX 2070 at 8920.32 GFLOPS and 448 GB/sec of VRAM bandwidth. You'd think the Vega 64 would be a no brainer in terms of who's the higher performer. Except, at least near the time of release, it was barely competitive: https://www.anandtech.com/bench/product/2512?vs=2513

The only way to compare video cards, or even hardware in general, is to actually benchmark it and compare the results. You can always make guesses, but it's not so cut and dry in practice.
 
  • Like
Reactions: dotas1