Given the limited memory on GPUs, I'm wondering why there are no socketed GDDR5 memory modules so that you can install more RAM.

The main challenge seems to be maintaining signal integrity since GDDR5 runs 2.5x faster than the fastest DDR4 (4 GHz vs. 1.6GHz, double pumped). Unlike DDR4, it needs to train links to find the phase offset between bits (to account for different PCB trace lengths) so that they can be sampled together. Also, they need to use CRC.

XDR 2 runs even faster at 800 MHz, x16 pumped! but uses differential signaling which means a lower bandwidth per pin than GDDR5.

So will putting GDDR5 onto DIMMs greatly increase wire length, signal attenuation, and negate the speed advantage?

Or is it lack of interest? There would hardly be a market for just GDDR5 upgrades for GPUs. CPU makers would need to support it too. AMD already has for their APU in the PS4.

  • 1
    \$\begingroup\$ GDDR5X has been approved for use. It is twice as fast as GDDR5 and has a 190 pin socket. GDDR5 has a 170 pin socket. Those who make motherboards/plugins are reluctant to use a chip that is constantly being upgraded and is not legacy compatible. \$\endgroup\$ – Sparky256 Jun 1 '16 at 22:07
  1. GPUs are often already on a plug in module. These modules often have limited room and/or component height allowances. So DIMMs are typically not really feasible from a mechanical standpoint.
  2. DIMMs place a rigor on fixed pin assignment and order. This can add significantly to the overall trace length when the routing a memory bus through a DIMM connector.
  3. Each connector in a PCB interconnection scheme represents at least some kind of impedance discontinuity to signals that route through the connector. This can induce reflections which impact signal integrity and jitter.
| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.