4 ms·

They Ryzen chips have a significant bump in PCIE bandwidth compared to Intel chips of this generation. Without knowing specifically which Atom to pull up in In

by mey 7y ago

They Ryzen chips have a significant bump in PCIE bandwidth compared to Intel chips of this generation. Without knowing specifically which Atom to pull up in Intel Ark, I can't compare but I expect that anything in this lineup would smoke an Atom chip for raw IO capacity.

tpolzer 7y ago
You'd compare them to Xeon D (which are more comparable in terms of power/intended usage), so e.g. for top of the line 16 core models, the EPYC Embedded 3451 vs. the Xeon D-2183IT, with 64 vs. 32 PCIe 3.0 lanes.
- rbanffy 7y ago
  Isn't that a lot of IO capacity for a home server? BTW, I checked Xeon D motherboards and got scared with the prices.
  - devonkim 7y ago
    There's 2 generations of Xeon D's - the 15xx series from circa 2015 and the 21xx series from about a year ago. Their target market is essentially for low powered but somewhat powerful compute farms. Their pricing is part of the Intel mark-up for supporting RDIMMs on top of other datacenter-centric features such as remote management BOMs, SFF-8088 connector for onboard LSI HBAs, and even SFP+ support. Those features are quite reasonable for the $600 range when including an efficient low-TDP CPU. My home server pegs my I/O when I run some tests to push a ton of bytes through Kafka, Postgres, Rook, etc. on top of a ZFS zpool. I do K8S at home along with some deep learning + big data-y toy projects because it's still cheaper than firing the equivalent up in any cloud provider after about 6 months of operation for myself, and at least I can repurpose my toys.
    - rbanffy 7y ago
      > I run some tests to push a ton of bytes through Kafka, Postgres, Rook, etc. on top of a ZFS zpool. Step aside, Xeon D. This is a job for EPYC. Or POWER9.
      - devonkim 7y ago
        There's always a time/cost trade-off to be made in any engineering endeavor and I'm willing to accept some longer processing times for hobby usage until it becomes more serious I should be investing a lot more into it. I have a 3900x I use mostly for video encoding, K8S, local dev, and games if I need more power. And honestly, if I want to run stuff super fast in a hurry and am willing to pay for EPYC and Threadripper kinds of hardware, I'm running it all into Terraform and Packer and shipping it off to a cloud to crunch it for a whole $30 in an hour or two. It's just that along the way I'd spend thousands screwing up on silly problems.

by mey 7y ago

tpolzer 7y ago

You'd compare them to Xeon D (which are more comparable in terms of power/intended usage), so e.g. for top of the line 16 core models, the EPYC Embedded 3451 vs. the Xeon D-2183IT, with 64 vs. 32 PCIe 3.0 lanes.

rbanffy 7y ago
Isn't that a lot of IO capacity for a home server? BTW, I checked Xeon D motherboards and got scared with the prices.
- devonkim 7y ago
  There's 2 generations of Xeon D's - the 15xx series from circa 2015 and the 21xx series from about a year ago. Their target market is essentially for low powered but somewhat powerful compute farms. Their pricing is part of the Intel mark-up for supporting RDIMMs on top of other datacenter-centric features such as remote management BOMs, SFF-8088 connector for onboard LSI HBAs, and even SFP+ support. Those features are quite reasonable for the $600 range when including an efficient low-TDP CPU. My home server pegs my I/O when I run some tests to push a ton of bytes through Kafka, Postgres, Rook, etc. on top of a ZFS zpool. I do K8S at home along with some deep learning + big data-y toy projects because it's still cheaper than firing the equivalent up in any cloud provider after about 6 months of operation for myself, and at least I can repurpose my toys.
  - rbanffy 7y ago
    > I run some tests to push a ton of bytes through Kafka, Postgres, Rook, etc. on top of a ZFS zpool. Step aside, Xeon D. This is a job for EPYC. Or POWER9.
    - devonkim 7y ago
      There's always a time/cost trade-off to be made in any engineering endeavor and I'm willing to accept some longer processing times for hobby usage until it becomes more serious I should be investing a lot more into it. I have a 3900x I use mostly for video encoding, K8S, local dev, and games if I need more power. And honestly, if I want to run stuff super fast in a hurry and am willing to pay for EPYC and Threadripper kinds of hardware, I'm running it all into Terraform and Packer and shipping it off to a cloud to crunch it for a whole $30 in an hour or two. It's just that along the way I'd spend thousands screwing up on silly problems.