10 Comments

Back to Article

• nandnandnand - Wednesday, January 25, 2023 - link

  https://hothardware.com/news/innosilicon-lpddr5x-d...
  
  Here's a demo of LPDDR5X-10000. This 'T' name is marketing BS.

• lmcd - Wednesday, January 25, 2023 - link

  The official product name cannot use X designation. The standard is named and done, and unlike consumer RAM, overclocking is not an option.

• nandnandnand - Wednesday, January 25, 2023 - link

  JEDEC has published two updates to the DDR5 spec. I'm sure they can do the same with LPDDR5X and raise the maximum allowed speeds.

• meacupla - Wednesday, January 25, 2023 - link

  does it really work at 9600Mbps, or is it 4800Mbps x2?
  because if it's the latter, isn't that just quad channel?

• Ryan Smith - Wednesday, January 25, 2023 - link

  9600Mbps would be the per pin rate. Memory speeds are measured per pin, not per package.

• meacupla - Wednesday, January 25, 2023 - link

  Okay, in that case, this stuff looks like a win for soldered on RAM.

• brucethemoose - Wednesday, January 25, 2023 - link

  So that's 384 bits for the equivalent bandwidth of the RTX 4070 TI's 192-bit, 504GB/s bus.
  
  A 1024-bit bus like the dual-die M1 Max would be 1.2TB/s, higher than 1TB/s on the 4090 (but still less than an A100).
  
  I hope some wide, low power stuff comes out of this.

• schujj07 - Thursday, January 26, 2023 - link

  This will be used in laptops and phones where you usually have 128bit and 64bit buses.

• brucethemoose - Thursday, January 26, 2023 - link

  Yeah thats what I figured, but some exotic, wide designs (like the M1/M2 Max) would be cool.

• mode_13h - Wednesday, March 15, 2023 - link

  Nvidia's Grace is using in-package LPDDR5X, similar to Apple's M-series. The main difference is that Grace is a server-oriented CPU, and Nvidia was willing to forego memory upgradability mainly for the power-savings of putting it in-package. Current configuration is 480 GB per CPU.
  
  https://developer.nvidia.com/blog/nvidia-grace-cpu...