GDDR5 vs GDDR3

[PC_User]

Just curious but what's the real-world difference between the two?

I came across this article which shows that the gains are minimal given the speed differences between the two. What's the verdict?

 

[risqu3]

its not about the speed, its about the cost savings for the bus width, something that would require 512bit bus for gddr3 would only require a 256bit bus, thus reducing costs and price, and memory chips will eventually become cheaper then adding more bus width to a video card (right now id estimate its about the same cost wise)

 

[defaultluser]

First of all: do not believe the "clock speeds" for GDDR5 in the article. The baseline clock speed is the same as GDDR3; GDDR5 simply quad-pumps the data clock while GDDR3 dual-pumps the data clock. When they claimed to run the DDR5 memory at 993 MHz, they were really running it at half that speed.

GDDR3 = 2 data per clock.
GDDR5 = 4 data per clock (with additional latency overhead over GDDR3).

What does GDDR5 mean for the 4870? It means that they can offer you a higher-performing card than the 4850 without going to a fatter bus. The fact is, the 4850 is really pushing what GDDR3 on 256-bit bus can manage, and although GDDR5 on that same bus is overkill for the 4870, it does unlock some impressive performance increases.

The key to GDDR5 is enabling the NEXT generation of graphics cards without the need to go to even fatter buses, and also reducing the memory power consumption vs GDDR3. It also allows the creation of cheap 128-bit enthusiast cards (e.g. 4770).

 

[grimreeferx09]

rumor has it that the nvidia dx11 card will have a 512bit bus with gddr5 for 300GB/s of bandwidth.

 

[PC_User]

Thanks for the explanations. The reason is because I was comparing the GTX 285 and Radeon 4890 and was wondering if the GDDR5 of the 4890 places it at a big advantage over the 285.

 

[Sojuuk]

i was under the impression that fatter buses were harder to manufacture and needed more layers and probably a more complicated chip interface than a simpler one using GDDR5 but hey if im wrong: im wrong.

 

[defaultluser]

i was under the impression that fatter buses were harder to manufacture and needed more layers and probably a more complicated chip interface than a simpler one using GDDR5 but hey if im wrong: im wrong.

No, you're right, that's why the 4770 has a 128-bit bus and can sell for just $100. Now that GDDR5 is mainstream and cheap, they can make this happen.

 

[Creig]

Thanks for the explanations. The reason is because I was comparing the GTX 285 and Radeon 4890 and was wondering if the GDDR5 of the 4890 places it at a big advantage over the 285.

No it doesn't. The GTX 285 uses slower GDDR3 compared to the faster GDDR5 on the 4890, but it uses a wider 512-bit memory interface compared to the narrower 256-bit interface on the 4890. To give an approximate comparison, GDDR5+256-bit roughly equals GDDR3+512-bit. The GDDR3 memory modules are slightly cheaper than GDDR5, but the 512-bit PCB is more expensive to produce than the 256-bit.

They're really just differing ways to achieve the same goal.

 

[Trimlock]

rumor has it that the nvidia dx11 card will have a 512bit bus with gddr5 for 300GB/s of bandwidth.

That is absolutely retarded and would drive costs up on a card for a company having a hard time with cash at the moment.

Going with GDDR5 is a smarter choice then developing a wider bus which would undoubtedly bring yields down, board prices up and ensure longer development times just to keep using a slightly cheaper GDDR3 module. If you need an example, take a look at the 3870, that thing had all sorts of problems with it, can we say alot of these were attributed to them trying to shove a 512bit bus on there? maybe.

 

[PC_User]

No it doesn't. The GTX 285 uses slower GDDR3 compared to the faster GDDR5 on the 4890, but it uses a wider 512-bit memory interface compared to the narrower 256-bit interface on the 4890. To give an approximate comparison, GDDR5+256-bit roughly equals GDDR3+512-bit. The GDDR3 memory modules are slightly cheaper than GDDR5, but the 512-bit PCB is more expensive to produce than the 256-bit.

They're really just differing ways to achieve the same goal.

Thanks. Very informative and helpful.

 

[Trimlock]

I just re-read the article, the writer goes into imprecable detail and points out something extremely obvious...

At the same frequencies GDDR5 is slower, well yea, each release has higher latencies, why would it be faster? there is a reason why they ramp up frequencies on these to make up for those latencies.

Good article though, explains alot about DDR, his testing doesn't really show much of the benefits of GDDR5 but it was a fun read.

 

[Jon55]

Awesome thread. I had always wondered if GDDR5 really made the difference, but I now I understand bigger bus/smaller bus impacts the final benchmarks.

 

[kllrnohj]

At the same frequencies GDDR5 is slower, well yea, each release has higher latencies, why would it be faster? there is a reason why they ramp up frequencies on these to make up for those latencies.

At the same frequency GDDR5 is roughly twice as fast as GDDR3. Unless the author really meant at *half* the GDDR3 frequency then GDDR5 is slower (which would be a comparison of *effective* clock speeds) - but that is a rather stupid comparison

 

[Roland00]

That is absolutely retarded and would drive costs up on a card for a company having a hard time with cash at the moment.

Going with GDDR5 is a smarter choice then developing a wider bus which would undoubtedly bring yields down, board prices up and ensure longer development times just to keep using a slightly cheaper GDDR3 module. If you need an example, take a look at the 3870, that thing had all sorts of problems with it, can we say alot of these were attributed to them trying to shove a 512bit bus on there? maybe.

I think you mean the 2900 Pro and the 2900 XT. The 3870 didn't have a 512 bit bus, ATI/AMD was just starting to smart up. But regardless you are completely right on substance.

There is a reason why with the 275 Nvidia went with a 448 bit buses (as well as using 2 448 bit buses with the 295.) It keeps costs down allowing them them to make more profit with each product as well as shipping greater volume of parts.

It is also the same reason ATI went with a 128 bit bus with the 4770.

 

[PC_User]

Found this on another forum:

GDDR5 is better due to the fact the manufacture of said video card doesn't have to spend an expensive amount of money on chips and PCB layouts to get the same throughput from memory.

Its a trade off that the company has to be willing to take. In this case ATI got in early, and went GDDR5 w/ 256bit bus, while nVidia went in with GDDR3 say 512bit bus. While GDDR3 is more known out there and can get some good clock rates out of it, GDDR5 is where it will be at.

Currently though GDDR3 at least on some of the newer nVidia cards can push upwards of 2,800Mhz on a 512Bit Bus, where as ATI's can push 4,000Mhz on a 256bit bus. So in the end nVidia is winning with the GDDR3 because it can still push more information though there ram while being slower it has more highway lanes to process information. If to be equal in this case now totally depending on latencies on the ram chips (we'll take that out of this example) the GDDR3 w/ 512bit bus would have to run 2,000Mhz (easy by GDDR3 standards). Or ATI would have to run GDDR5 w/ 256bit bus at 5,600Mhz (no used on GPU's currently) to match performance.

In the end GDDR5 is better, cheaper for the manufacture to use, and in the end of the day better for you with less power being used to power the chips to get said speed and at a cheaper cost (GDDR5 is supposedly cheaper to manufacture).

 

[Trimlock]

I think you mean the 2900 Pro and the 2900 XT. The 3870 didn't have a 512 bit bus, ATI/AMD was just starting to smart up. But regardless you are completely right on substance.

There is a reason why with the 275 Nvidia went with a 448 bit buses (as well as using 2 448 bit buses with the 295.) It keeps costs down allowing them them to make more profit with each product as well as shipping greater volume of parts.

It is also the same reason ATI went with a 128 bit bus with the 4770.

actually you are right, I ment the 2900 series, which is what the 3870 was derived from, a very good attempt from ATI to bring a bust into the buyable market.

At the same frequency GDDR5 is roughly twice as fast as GDDR3. Unless the author really meant at *half* the GDDR3 frequency then GDDR5 is slower (which would be a comparison of *effective* clock speeds) - but that is a rather stupid comparison

hmm... yes I think you are right with that, he did say effective 1800mhz, he didn't say if he was counting the 4bits or what.