The Reason Why NVIDIA's GeForce RTX 3080 GPU Uses 19 Gbps GDDR6X Memory and not Faster Variants

[AleksandarK]

When NVIDIA announced its next-generation GeForce RTX 3080 and 3090 Ampere GPUs, it specified that the memory found in the new GPUs will be Micron's GDDR6X variant with 19 Gbps speed. However, being that there are faster GDDR6X modules already available in a 21 Gbps variant, everyone was left wondering why NVIDIA didn't just use the faster memory from Micron. That is exactly what Igor's Lab, a technology website, has been wondering as well. They have decided to conduct testing with an infrared camera that measures the heat produced. To check out the full testing setup and how they tested everything, you can go here and read it, including watching the video embedded.

Micron chips like GDDR5, GDDR5X, and GDDR6 are rated for the maximum junction temperature (TJ Max) of 100 degrees Celsius. It is recommended that these chips should run anywhere from 0C to 95C for the best results. However, when it comes to the new GDDR6X modules found in the new graphics cards, they are not yet any official specifications available to the public. Igor's Lab estimates that they can reach 120C before they become damaged, meaning that TJ Max should be 110C or 105C. When measuring the temperature of GDDR6X modules, Igor found out that the hottest chip ran at 104C, meaning that the chips are running pretty close to the TJ Max they are (supposedly) specified. It is NVIDIA's PCB design decisions that are leading up to this, as the hottest chips are running next to voltage regulators, which can get pretty hot on their own.



The takeaway here is that the heat produced from the card is quite huge with a TGP of 320 W, meaning that the cooling is quite a problem which NVIDIA managed to keep under control, however, the design decisions resulted in some possible performance loss. Instead of NVIDIA using Micron's faster 21 Gbps chips, they are forced to use the 19 Gbps variants. It is not like there would be supply issues as NVIDIA is Micron's only GDDR6X customer, however, the real reason is heat management. The speedup from a slower 19 Gbps chip to faster 21 Gbps chips is plus 10%, meaning that the heat output would follow the percentage, resulting in even greater TGP of the card. The conclusion drawn here is that the current chips are limited by temperature only, as the card is pretty hot, and that is why NVIDIA doesn't use the faster GDDR6X variant.

View at TechPowerUp Main Site

 

[fynxer]

This is where the partners come in to play with their custom cooling solutions.

With new cooling innovation that cools both VRM and memory properly enables them to use 21Gbit GDDR6X on partner cards.

What I heard there is no restrictions for partners to use the faster 21Gbit memory.

ALSO a tactic of nvidias have been to milk the the first wave of restless buyers with cut down and restricted versions of their GFX cards.

Then later release a "Super" version with 21Gbit GDDR6X or a 3080Ti with 352-bit memory bus with 22GB and 21Gbit GDDR6X.

Notice that the first wave of 3080 lacks any 20GB versions and I really don't see any reason for this other than to milk the first ware of buyers. 3080 with 20GB is needed for TRUE 4K future gaming, some 4K games already MAX out 10GB so to be able to fully enjoy future 4K games with 3080 you need 20GB.

 

[Vayra86]

More compromises on the memory side.

I'll see myself out now

 

[Vya Domus]

Like I said, form over function.

 

[noname00]

More compromises on the memory side.

I'll see myself out now

So you care more about the specs than the actual performance and price/performance of the card?

 

[Vya Domus]

So you care more about the specs than the actual performance and price/performance of the card?

You think faster memory would just be a number on a spec sheet ?

 

[Deleted member 24505]

Seems like ideally the 3080 needs water cooling

 

[Vayra86]

So you care more about the specs than the actual performance and price/performance of the card?

Nah, I just don't like my memory slow-cooked so that it can die quickly. The cost isn't just performance here.

 

[Parn]

At 320W+ TDP I'm a bit surprised that nvidia hasn't released a water cooled version of the FE. I believe AMD did that for Vega (or was it Fury?).

 

[Hossein Almet]

...and price. A 21 Gbps chip is certainly more expensive than the 19 Gbps chip.

 

[noname00]

You think faster memory would just be a number on a spec sheet ?

I'm sure 10% faster memory would bring better performance, probably around 5%, but that does not invalidate the fact the card is really fast, and worth the money.

What I'm saying is the extra performance that could have been is not a reason not to buy a something.

Edit: probably another reason for the slower memory is the difference to GDDR6 that is in the RTX3070, the performance difference would be too high, and the performance of the RTX3080 would get too close to the performance of RTX3090.

 

[delshay]

This was a big concern for me too how they put the memory chips right next to the inductors which will run very hot. They can use the faster memory chips, but they have to fit higher quality thermal pads. So it's not an issue or a problem or an excuse for not using the faster variant.

 

[Vayra86]

I'm sure 10% faster memory would bring better performance, probably around 5%, but that does not invalidate the fact the card is really fast, and worth the money.

What I'm saying is the extra performance that could have been is not a reason not to buy a something.

Next time say what you mean then. Assumption is the mother of all F*ups.

I've already clarified as well

 

[Xuper]

Did nvidia expose Memory temp ? like AMD navi cards ?

 

[Vya Domus]

Did nvidia expose Memory temp ? like AMD navi cards ?

It was measured externally, which means the chips run even hotter in reality now that I think about it.

 

[chaosmassive]

let's see if Nvidia allows AIB to put G6X memory since they have much better cooling solution than Founders Edition.
if not, then this reason is bs and Nvidia just want to gimp RTX 3080 with memory bandwidth for segmentation purposes.

 

[BoboOOZ]

These cards are not memory bandwidth starved anyway, preliminary memory overclock tests show very little performance gains. The extra performance is not a real problem, however, existing temps already seem to be a problem.

I remember some 5700XT having to sell for very low prices because they ran their memory too hot (Asus tuf and MSI evoke gen 1), and the temps were slightly under 100°C.

 

[fynxer]

These cards are not memory bandwidth starved anyway, preliminary memory overclock tests show very little performance gains. The extra performance is not a real problem, however, existing temps already seem to be a problem.

I remember some 5700XT having to sell for very low prices because they ran their memory too hot (Asus tuf and MSI evoke gen 1), and the temps were slightly under 100°C.

Reviewer said that overclocking the GPU yielded very little performance gain, being able to overclock the memory by +800-1000Mhz saved the day or else there would be almost no overclocking increase at all. So what I see is that if 21Gbit memory can add another +5-10% performance it is well worth it.

 

[Calmmo]

let's see if Nvidia allows AIB to put G6X memory since they have much better cooling solution than Founders Edition.
if not, then this reason is bs and Nvidia just want to gimp RTX 3080 with memory bandwidth for segmentation purposes.

All AIB cards have the same 6x memory.

 

[nguyen]

Ah just screw the FE, let the AIBs do their job .
Although this time around the FE's pricing and thermal performance has forced AIB to lower the price on their premium models.
So it's still a bonus.

 

[Caring1]

Memory should have been mounted on the back of the card and cooled separately.

 

[BoboOOZ]

Reviewer said that overclocking the GPU yielded very little performance gain, being able to overclock the memory by +800-1000Mhz saved the day or else there would be almost no overclocking increase at all. So what I see is that if 21Gbit memory can add another +5-10% performance it is well worth it.

That's because the GPU OC was only 70MHz, due to the power envelope limit. Applying an 800MHz overclock for 5% performance gain doesn't seem reasonable to me, the speed of the memory is not the main bottleneck here, it's the GPU frequency.

 

[londiste]

This is worse for 3090 where VRAM chips are on the backside of the card. At least at the front they are connected to the big heatsink.

At 320W+ TDP I'm a bit surprised that nvidia hasn't released a water cooled version of the FE. I believe AMD did that for Vega (or was it Fury?).

Since RAM and VRM seem to be the problem here, a water cooled version with a fullcover block, please

 

[chaosmassive]

All AIB cards have the same 6x memory.

err I mean higher clocked G6X one, sorry for missing out the information.

 

[Outback Bronze]

Nvidia should give me a rebate for putting blocks on their cards : )