Samsung Launches "Flashbolt" HBM2E Memory

[erek]

Beastly memory bandwidth!

"Samsung's Flashbolt provides a highly reliable data transfer speed of 3.2 gigabits per second (Gbps) by leveraging a proprietary optimized circuit design for signal transmission, while offering a memory bandwidth of 410 GB/s per stack. Samsung's HBM2E can also attain a transfer speed of 4.2 Gbps, the maximum tested data rate to date, enabling up to a 538 GB/s bandwidth per stack in certain future applications. This would represent a 1.75x enhancement over Aquabolt's 307 GB/s.

Samsung expects to begin volume production during the first half of this year. The company will continue providing its second-generation Aquabolt lineup while expanding its third-generation Flashbolt offering, and will further strengthen collaborations with ecosystem partners in next-generation systems as it accelerates the transition to HBM solutions throughout the premium memory market. "

https://www.techpowerup.com/263560/samsung-launches-3rd-generation-flashbolt-hbm2e-memory

 

[defaultluser]

I could see Nvidia using this for powering Compute Ampere.

 

[viivo]

AMD is already researching ways to implement this memory on GPUs that can't make use of its full potential while remaining prohibitively expensive.

 

[Dark12]

AMD is already researching ways to implement this memory on GPUs that can't make use of its full potential while remaining prohibitively expensive.

SHOTS FIRED

 

[Lakados]

I could see Nvidia using this for powering Compute Ampere.

I would like to see it on the new Tesla's. They are already expensive the cost difference at that bracket is marginal at best.

 

[ManofGod]

I see this in professional applications only.

 

[defaultluser]

I would like to see it on the new Tesla's. They are already expensive the cost difference at that bracket is marginal at best.

If the rumors of double throughput are to be beloeved, then they will need it. t

They're going to need at least three stacks to feed 17 tflop[s FP64!

v100 was pushing the limits of 4 stack HBM2.

 

[Jandor]

If the rumors of double throughput are to be beloeved, then they will need it. t

They're going to need at least three stacks to feed 17 tflop[s FP64!

v100 was pushing the limits of 4 stack HBM2.

The top HBM2 from that time used on V100 was very slow compared to even the previous HBM2 form Samsung. This is already around 3 times faster and could be pushed to 4 times faster.