Announcement

Collapse
No announcement yet.

Benchmarking The Updated Intel CPU Microcode For SRBDS / CrossTalk Mitigation

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Benchmarking The Updated Intel CPU Microcode For SRBDS / CrossTalk Mitigation

    Phoronix: Benchmarking The Updated Intel CPU Microcode For SRBDS / CrossTalk Mitigation

    Following yesterday's disclosure of CrossTalk / SRBDS after a nearly two year embargo period for this Special Register Buffer Data Sampling vulnerability, I have been running benchmarks on multiple systems for the past nearly 24 hours. Here are some preliminary data points for both synthetic and real-world workloads on various Intel CPUs before/after mitigating SRBDS with the updated Intel microcode.

    http://www.phoronix.com/vr.php?view=29268

  • #2
    Bulldozer/Piledriver architecture and performance was revolutionary after all!!!

    Comment


    • #3
      prolificily
      Dyslexia...?

      Comment


      • #4
        Anyone even still running an Intel CBU, Central Bug Unit? ;-)

        Comment


        • #5
          Yikes.... just when I thought things couldn't keep getting worse.
          Kinda funny how AMD screwed up their rdrand instruction and basically decided to fix it by turning it off, but after this mitigation, you might as well not have it enabled on Intel either.

          Anybody know of how Windows has held up with all these mitigations? It's not so much that I care about how much Windows' performance has been penalized, but rather, if Windows isn't properly secured, it might have a misleading performance advantage over Linux.

          Originally posted by siyia View Post
          Bulldozer/Piledriver architecture and performance was revolutionary after all!!!
          But seriously... I'm legitimately curious how much the performance margins for FX have closed up as a result of all these mitigations. I'm sure it's still an inferior architecture but at this rate it can't be too far behind.

          Comment


          • #6
            While most workloads are not heavily impacted, and for the ones that are it is possible to generate randomness in other ways (cryptographically secure PRNGs were used long before RDRAND existed), the big problem with this mitigation is that it locks the memory bus, thus impacting other cores in a similar way to split locks. I would like to see some numbers of how much other workloads are impacted by a rdrand workload running concurrently on another core.

            Depending how bad the performance hit is, this could be used for DoS attacks in public clouds, and it is likely hard to track down if intermittent, as it wouldn't leave any traces in logs etc.

            Comment


            • #7
              But seriously... I'm legitimately curious how much the performance margins for FX have closed up as a result of all these mitigations. I'm sure it's still an inferior architecture but at this rate it can't be too far behind.
              My guess is that Piledriver is 10% behind the offerings of equivalent Intel products with all mitigations enabled, no real data though.

              Comment


              • #8
                Well for my good old Haswell i am running, the "bugs" info in /proc/cpuinfo now has two lines in a standard console, as it is > 80 characters long lol..
                Haswell/Broadwell is the architecture that suffers the most IMHO..

                Intel probably needs another R&D team to develop their architecture for 2023+ as the current team is hung up with fixing and managing all those speculative execution microcode updates?! / doing silicon level fixes in the current "refreshes".

                My guess is that they will just have tripple the ports per core and a very wide vector unit which takes massive amounts of chip area to beat AMD at IPC / regain the lead in their next architecture.. By this time probably most data center level software is optimized to scale on 100des of cores / massive memory bandwidth and single thread performance is only relevant for gamers / old software.. In the Exascale datacenters ARM Neoverse will take over IMHO at least that´s what we are seing currently. Maybe we will see another completly custom architecture from one of the big 4 big cloud providers?

                Comment


                • #9
                  I definitely want to see the effect on a non-affected benchmark with that rdrand benchmark (denial of service tool?) running at the same time and locking the memory bus.
                  Both on single-socket bare-metal, single-socket VM, and multi-socket.

                  Comment


                  • #10
                    While I don't like Intel (at least their hardware department) for their insecure CPUs, but this is just too much. So many companies relies on it, so many powerful servers with hundreds of CPUs and tera bytes of memory, so much investment in Linux from Intel. I hope they'll will stop this blizzard somehow. The good thing is their monopoly will probably end.

                    Comment

                    Working...
                    X