4 ms·

Holding out hope for PCIe5 and DDR5 too...

by tus88 7y ago

nobodyshere 7y ago
I wouldn't count on it. There aren't even that many pci-e 4 devices on the market.
- mattashii 7y ago
  Intel is planning for PCIe5 in 2021 (skipping on PCIe4), so AMD having gen 5 support in that generation would not be too unexpected. Though with the seperation of IO/compute dies I would not be suprised if PCIe5 would only appear in EPYC/Threadripper components. I expect that the upcoming Ryzen socket refresh (AM5, which is expected to come with Zen 4/Ryzen 5xxx) to have support for (or even require) DDR5, as by then DDR5 will already be available for a year.
- wtallis 7y ago
  In the server market where the extra bandwidth is actually useful, the transition from PCIe 3 to PCIe 4 is pretty far along. Switches, NICs, SSDs and FPGAs supporting PCIe 4 are all either available or announced, and server-class GPUs should also show up next year. Most of those products have good reason to continue on to PCIe 5 as soon as suitable hosts are on the way.
  - floatboth 7y ago
    Server-class GPUs kinda showed up first, or at least at the same time: Radeon Instinct MI50 was the first PCIe 4 GPU ever. (The consumer Radeon VII is the same chip)
    - wtallis 7y ago
      Ah, I forgot about that one. Possibly because they didn't enable PCIe 4 functionality on the Radeon VII.
- NullPrefix 7y ago
  Plenty of gen 4 SSDs on the market
- bryanlarsen 7y ago
  There was less than two years between the release of the two standards, so two years between the release of supporting CPUs wouldn't be surprising.
shmerl 7y ago
I'm more interested in HBM memory becoming cheaper for GPUs.
- philjohn 7y ago
  That's what HBM 3 is meant to accomplish.
- hajile 7y ago
  I want HBM for CPUs. It's faster, lower power, and lower latency. The only potential downside is being harder to upgrade, but I think that could also be handled with some work.
  - arcanus 7y ago
    HBM is not lower latency than traditional DDR. It is significantly higher bandwidth, but the latency can be higher. On GPUs this is not a problem, because they are quite latency tolerant. CPUs might have more trouble hiding the latency impact.
    - hajile 7y ago
      You are correct about HBM2 (latencies between 12 and 16ns vs 8-12ns latencies in typical DDR4 systems). My understanding was that HBM3 reduces those latencies and HBM4 potentially reduces them even further (in addition to lower latency due to closer proximity). AMD seems considers HBM on-chip to be an upcoming reality. https://www.guru3d.com/news-story/amd-working-on-vertical-(3d)-stacking-of-dram-onto-processors,2.html https://www.guru3d.com/news-story/amd-working-on-vertical-(3...
einpoklum 7y ago
Ok, then - expect to wait at least 5 years, maybe more. PCIe 4 was standardized in 2011 and it's still not that widely adopted. So let's say 9 years from standardization to wide availability. PCIe 5 was standardized in 2017... 2017+9 = 2025. Good luck.

Holding out hope for PCIe5 and DDR5 too...

by tus88 7y ago

Holding out hope for PCIe5 and DDR5 too...

nobodyshere 7y ago

I wouldn't count on it. There aren't even that many pci-e 4 devices on the market.

mattashii 7y ago
Intel is planning for PCIe5 in 2021 (skipping on PCIe4), so AMD having gen 5 support in that generation would not be too unexpected. Though with the seperation of IO/compute dies I would not be suprised if PCIe5 would only appear in EPYC/Threadripper components. I expect that the upcoming Ryzen socket refresh (AM5, which is expected to come with Zen 4/Ryzen 5xxx) to have support for (or even require) DDR5, as by then DDR5 will already be available for a year.
wtallis 7y ago
In the server market where the extra bandwidth is actually useful, the transition from PCIe 3 to PCIe 4 is pretty far along. Switches, NICs, SSDs and FPGAs supporting PCIe 4 are all either available or announced, and server-class GPUs should also show up next year. Most of those products have good reason to continue on to PCIe 5 as soon as suitable hosts are on the way.
- floatboth 7y ago
  Server-class GPUs kinda showed up first, or at least at the same time: Radeon Instinct MI50 was the first PCIe 4 GPU ever. (The consumer Radeon VII is the same chip)
  - wtallis 7y ago
    Ah, I forgot about that one. Possibly because they didn't enable PCIe 4 functionality on the Radeon VII.
NullPrefix 7y ago
Plenty of gen 4 SSDs on the market
bryanlarsen 7y ago
There was less than two years between the release of the two standards, so two years between the release of supporting CPUs wouldn't be surprising.

shmerl 7y ago

I'm more interested in HBM memory becoming cheaper for GPUs.

philjohn 7y ago
That's what HBM 3 is meant to accomplish.
hajile 7y ago
I want HBM for CPUs. It's faster, lower power, and lower latency. The only potential downside is being harder to upgrade, but I think that could also be handled with some work.
- arcanus 7y ago
  HBM is not lower latency than traditional DDR. It is significantly higher bandwidth, but the latency can be higher. On GPUs this is not a problem, because they are quite latency tolerant. CPUs might have more trouble hiding the latency impact.
  - hajile 7y ago
    You are correct about HBM2 (latencies between 12 and 16ns vs 8-12ns latencies in typical DDR4 systems). My understanding was that HBM3 reduces those latencies and HBM4 potentially reduces them even further (in addition to lower latency due to closer proximity). AMD seems considers HBM on-chip to be an upcoming reality. https://www.guru3d.com/news-story/amd-working-on-vertical-(3d)-stacking-of-dram-onto-processors,2.html https://www.guru3d.com/news-story/amd-working-on-vertical-(3...

einpoklum 7y ago

Ok, then - expect to wait at least 5 years, maybe more. PCIe 4 was standardized in 2011 and it's still not that widely adopted. So let's say 9 years from standardization to wide availability. PCIe 5 was standardized in 2017... 2017+9 = 2025. Good luck.