5 ms·

I hope to god the big three cloud providers pick up EPYC in good quantities. High core counts, lots of PCIe lanes, competitive perf per watt, what’s not to like

I hope to god the big three cloud providers pick up EPYC in good quantities. High core counts, lots of PCIe lanes, competitive perf per watt, what’s not to like? We badly need some kind of viable competition for Intel.

noir_lord 8y ago
Given the pasting meltdown had on Intel performance and cloud providers pretty much have to run the patches, if Epyc doesn't sell well then Intel is pretty much unassailable. I'm not convinced they are Opteron showed they can be vulnerable and Intel's handling of meltdown and PR was terrible, I imagine lots of pissed off CTO's.
- gameswithgo 8y ago
  It might be an opportunity for a cloud startup to create a whole business around AMD, it may be too hard for the established players to switch at this point.
drewg123 8y ago
The problem for folks below that tier is that the ecosystem for AMD is terrible. On the software side, basic tools like numatop don't work on epyc. That's because the the equivalent of the Intel uncore counters that measure infinity fabric (AMD's new Hyper Transport, like Intel's QPI) are not available. Perhaps the big three have the time to do these tools themselves, and the NDAs in place to get access to the counters. Since even their single socket offering has up to 4 numa domains, getting access to these counters is pretty important. There is also a limited selection of motherboard offerings from just a few vendors. This won't impact some of the cloud providers (who do their own board design), but for the rest of us, it is a pain. We've had an epyc we've been testing. I'm its biggest fan, but the motherboard is terrible in an enterprise setting. Just this week I updated the BIOS on it to get the new AGESA. Unlike our Intel / Supermicro boards, the BIOS update wiped all the settings. Earlier BIOS updates have required a physical (at the wall) power cycle, rather than a remote ipmi chassis power cycle.
- pas 8y ago
  Why AMD is not publishing these counter docs/specs? Is the Intel equivalent public? (I assume, since numatop is open source, and uses perf counters from the standard Linux kernel - no binary blobs, right?)
- satai 8y ago
  numatop was not updated since 2015 AFAIK...
  - drewg123 8y ago
    It was actually updated in late 2017 to add POWER support.

noir_lord 8y ago

Given the pasting meltdown had on Intel performance and cloud providers pretty much have to run the patches, if Epyc doesn't sell well then Intel is pretty much unassailable. I'm not convinced they are Opteron showed they can be vulnerable and Intel's handling of meltdown and PR was terrible, I imagine lots of pissed off CTO's.

gameswithgo 8y ago
It might be an opportunity for a cloud startup to create a whole business around AMD, it may be too hard for the established players to switch at this point.

drewg123 8y ago

The problem for folks below that tier is that the ecosystem for AMD is terrible. On the software side, basic tools like numatop don't work on epyc. That's because the the equivalent of the Intel uncore counters that measure infinity fabric (AMD's new Hyper Transport, like Intel's QPI) are not available. Perhaps the big three have the time to do these tools themselves, and the NDAs in place to get access to the counters. Since even their single socket offering has up to 4 numa domains, getting access to these counters is pretty important. There is also a limited selection of motherboard offerings from just a few vendors. This won't impact some of the cloud providers (who do their own board design), but for the rest of us, it is a pain. We've had an epyc we've been testing. I'm its biggest fan, but the motherboard is terrible in an enterprise setting. Just this week I updated the BIOS on it to get the new AGESA. Unlike our Intel / Supermicro boards, the BIOS update wiped all the settings. Earlier BIOS updates have required a physical (at the wall) power cycle, rather than a remote ipmi chassis power cycle.

pas 8y ago
Why AMD is not publishing these counter docs/specs? Is the Intel equivalent public? (I assume, since numatop is open source, and uses perf counters from the standard Linux kernel - no binary blobs, right?)
satai 8y ago
numatop was not updated since 2015 AFAIK...
- drewg123 8y ago
  It was actually updated in late 2017 to add POWER support.