Insider Info Alleges SK hynix Preparing HBM3E Samples for NVIDIA

[T0@st]

Industry insiders in South Korea have informed news publications that NVIDIA has requested that SK hynix submit samples of next-generation high bandwidth memory (HBM) for evaluation purposes—according to Business Korea's article, workers were preparing an initial batch of HBM3E prototypes for shipment this week. SK hynix has an existing relationship with NVIDIA—it fended off tough competition last year and has since produced (current gen) HBM3 DRAM for the H100 "Hopper" Tensor Core GPU.

The memory manufacturer is hoping to maintain its position as the HBM market leader with fifth generation products in the pipeline—vice president Park Myung-soo revealed back in April that: "we are preparing 8 Gbps HBM3E product samples for the second half of this year and are preparing for mass production in the first half of next year." A new partnership with NVIDIA could help SK hynix widen the gulf between it and and its nearest competitor - Samsung - in the field of HBM production.



View at TechPowerUp Main Site | Source

 

[Legacy-ZA]

So, RTX5080 for a mere $10 000?

 

[TumbleGeorge]

Rumor, HBM3E will be ~2.5 times faster than HBM3. I didn't believe but?

 

[R0H1T]

Based on what?

 

[Denver]

Rumor, HBM3E will be ~2.5 times faster than HBM3. I didn't believe but?

25% faster than latest commercially available HBM3E

If I'm not mistaken there was a Samsung slide showing plans to reach that speed only in 2025-2026...

 

[TumbleGeorge]

Based on wcckek.
8GT vs 3.2GT transfer rate.

 

[R0H1T]

WTF tech could be wrong ~

SK Hynix Preps HBM3E Memory: A 25% Speed Boost Over HBM3

HBM3 to get a boost before HBM4 arrives.

www.tomshardware.com

SK Hynix has the fastest HBM on the market & Samsung's already at 5GB/s so they should be higher.

 

[Denver]

Based on wcckek.
8GT vs 3.2GT transfer rate.

"Assembly aside, the performance specifications for SK hynix's 24GB HBM3 stacks are identical to their existing 16GB stacks. That means a maximum data transfer speed of 6.4Gbps/pin running over a 1024-bit interface, providing a total bandwidth of 819.2 GB/s per stack."

Wccftech's comparison should be between the first (worst) version of HBM3.

 

[R0H1T]

The leap also sounds too much, given it's still HBM3 (based) right now.

 

[Space Lynx]

i better get that second job sooner rather than later if I want to play the future of games.

 

[Flyordie]

Whats the pricing these days for HBM2 Aquabolt etc? I know GDDR6 has fallen in price to the point its cheap as dirt but, my gosh.. cards have gotten to the size of ITX motherboards now. I'd just like a card as clean and compact as the V64XTX is. Make a high end GPU, throw 2 8G stacks on there for 16GB of VRAM, add an AIO and sell it for $1,000 or so.

 

[Denver]

Whats the pricing these days for HBM2 Aquabolt etc? I know GDDR6 has fallen in price to the point its cheap as dirt but, my gosh.. cards have gotten to the size of ITX motherboards now. I'd just like a card as clean and compact as the V64XTX is. Make a high end GPU, throw 2 8G stacks on there for 16GB of VRAM, add an AIO and sell it for $1,000 or so.

Expensive.

"Samsung & SK Hynix GPU DRAM Prices Shoot Up: HBM3 5x More Expensive As Demand Grows For NVIDIA GPUs In ChatGPT"

Samsung & SK Hynix GPU DRAM Prices Shoot Up: HBM3 5x More Expensive As Demand Grows For NVIDIA GPUs In ChatGPT

The increasing demand for NVIDIA GPUs in ChatGPT has caused prices of DRAM including HBM from Samsung & SK Hynix to go up.

wccftech.com

 

[bonehead123]

Supply & demand...Supply & demand...

'nuff said

 

[mashie]

Don't expect any of these to end up in consumer GPU.

These are simply for the £200k+ 8 GPU servers used for AI workloads.

 

[NeuralNexus]

Why are most of the comments speaking as if Nvidia will be sourcing this memory for consumer products? They are clearing going to be using this for Enterprise use exclusively, like they have been doing for quite some time already. Unless you are in the market for a $10k GPU. The discussion around pricing and theoretical speed is irrelevant.

 

[TumbleGeorge]

Why are most of the comments speaking as if Nvidia will be sourcing this memory for consumer products? They are clearing going to be using this for Enterprise use exclusively, like they have been doing for quite some time already. Unless you are in the market for a $10k GPU. The discussion around pricing and theoretical speed is irrelevant.

It's obviously not irrelevant since you've posted here.