SK Hynix GDDR7 Memory Overclocked to 34 Gbps on NVIDIA GeForce RTX 5070 Ti

[AleksandarK]

As a GPU supply chain gets regularly updated, NVIDIA quietly added a new memory partner to its GeForce RTX 50 series lineup, bringing SK Hynix on board alongside Samsung for GDDR7 modules. Early mentions suggested that SK Hynix memory would appear on standard RTX 5070 cards. Still, the first sightings were on RTX 5070 Ti models, especially in China, where these cards have arrived ahead of other regions. Users on Chiphell and Baidu, verifying their GPUs with GPU-Z, discovered that the SK Hynix chips, which officially run at 28 Gbps, can be safely overclocked to 34 Gbps. This shows that switching to a second supplier does not hurt performance or overclocking headroom, so enthusiasts can expect the same headroom they've enjoyed with Samsung-sourced modules.

That said, some owners have encountered an obstacle when cross-flashing BIOS files with SKUs that use Samsung memory. Flashing an RTX 5070 Ti BIOS from a Samsung-equipped card onto one built with SK Hynix memory sometimes prevents the GPU from booting. However, dual-BIOS designs let users switch back to factory firmware and restore normal operation without too much fuss. As VideoCardz pointed out, TechPowerUp's BIOS database indicates that current firmware versions from various board partners already include support for GDDR7 modules from Samsung, SK Hynix, and even Micron. This suggests that these flashing issues aren't simply a matter of unrecognized memory. It could be related to board-specific power settings or other configuration quirks, with every GPU maker designing their boards differently, yielding possible errors if the firmware is swapped. Since SK Hynix-based cards run fine out of the box, most users won't have any reason to flash different firmware.



View at TechPowerUp Main Site | Source

 

[bonehead123]

try it, buy it, fry it.

spurn it, churn it, burn it.....

Light 'em up if you got 'em, simulate if ya don't...

Shall I continue ?

hahahahahahaha....

 

[Sammoonryong]

I would still prefer Samsung memory over Micron and Micron over Hynix but that's just my personal opinion.

 

[bgx]

did that help performance?

with GDDR, many time, you have a face value stable overclocking, but it hurts the performance because it uses all the powerful error correction which however have a penalty to pay....

No idea about the normal 3Dmark score ...

Also, didnt we report the hynix chip to be running hot at 28gbps already?
34gbps would be concerning on the long run.

 

[N/A]

Wasn't this just the driver imposed limit. 36gbps on the 5080.

 

[Crackong]

Doesn't really matter when Nvidia kept 'Fixed' the same black screen issue over and over and over and over and over again in 6 months.

 

[Tomorrow]

Wasn't this just the driver imposed limit. 36gbps on the 5080.

Indeed the current ones with 28Gbps have a 34Gbps soft limit and 5080 as the only SKU with 30Gbps default speed has a 36Gbps soft limit.

 

[FreedomEclipse]

try it, buy it, fry it.

spurn it, churn it, burn it.....

Light 'em up if you got 'em, simulate if ya don't...

Shall I continue ?

hahahahahahaha....

Feeling pressure?

 

[x4it3n]

I would still prefer Samsung memory over Micron and Micron over Hynix but that's just my personal opinion.

I don't think it really matters. My 4090 SUPRIM LIQUID X has some Micron GDDR6X chips that are stable @ 25Gbps (+19%)... And my 2x32GB (64GB) DDR5 of SK Hynix are perfectly stable @ 6200MHz 30-38-38-38-30-68
They're all pretty similar nowadays, it all depends of silicon lottery imo.

 

[W1zzard]

How is that news? I've reached 34 Gbps on every single RTX 5070 Ti that I've reviewed, in my launch day reviews, in February

 

[Legacy-ZA]

How is that news?

It's news because in the past, people would try and reach the same performance with their OCs as others which had initially received Samsung memory, they assumed, rightfully, they too had Samsung memory, but later to find out that AIBs and nGreedia did the old switcharoo to save a buck and used Micron or Hynix instead, *cough* so consumers ended-up damaging their GPUs. Totally not me btw

I've reached 34 Gbps on every single RTX 5070 Ti that I've reviewed, in my launch day reviews, in February

I can overclock my RTX5070Ti to 34Gbp/s but the data throughput remains the same as when I am on 32Gbp/s, I tested with Memtest_Vulkan-v0.5.0 and receive 795GB/s on both accounts (+2000/+3000), I also stressed the VRAM in OCCT to check for stability issues and errors (A.I says OCCT can detect errors, *shrugs* but I dunno for sure, since it's ECC).

As for benchmarks I only get around 80 points more in SpeedWay. I am just going to assume the reason I can't get any higher performance with the extra +1000, is because of the 256-bit bus/GPU itself. Or maybe it's just my card. Most people don't push their cards beyond +2000 as they don't know how to unlock the extra +1000, so unfortunately, there is very limited data available to me to compare my results.

All I know is, the drivers are a mess, even 4 months after launch I find so many display issues in-game, stuttering, black screens etc etc, it being so bad, you might just think your GPU is busted until you use Cyberpunk as a baseline because it receives fixes often, all this and they are already prepping Super variants of the 5080/5070, it's beyond disgraceful.

 

[W1zzard]

I can

Yeah, nobody said that it will turn into actual performance, but this is normal.

for stability issues and errors

Out of all the GDDR7 cards that I've tested, maybe 50, or more, only a single one couldn't run at the clock limit, I also haven't noticed any performance loss on any of these cards, so not a stability/ECC thing

 

[Tomorrow]

The soft limit could be product segmentation (gotta leave some perf for inevitable Super series) or temperature reasons.
Or both.

 

[wolf]

So far it seems like changing memory chip provider is a nothing burger. Let's see long term if they fail any quicker, but I wouldn't anticipate they would or that we even get meaningful data around that.

 

[Legacy-ZA]

Yeah, nobody said that it will turn into actual performance, but this is normal.

Interesting, I wonder though, if perhaps this could be related to some peoples stutters? I mean, if it has to "correct" something, data needs to be re-sent and verified. Kind of like if you have packet-loss on your internet connection.

Out of all the GDDR7 cards that I've tested, maybe 50, or more, only a single one couldn't run at the clock limit, I also haven't noticed any performance loss on any of these cards, so not a stability/ECC thing

Out of curiosity, that "one card", what was the maximum memory clock you could reach on it?

 

[W1zzard]

Out of curiosity, that "one card", what was the maximum memory clock you could reach on it?

Like 20 MHz below the limit

 

[bonehead123]

Feeling pressure?

???????

What pressure, this was just me being sarcastic,