Skip to main content

Our Tests Show Not All Ryzen 3000 Cores Are Created Equal

Thoughts

The new Ryzen 3000-series chips come with a mix of faster and slower cores, some of which cannot reach the rated single-core turbo speeds, which is a notable change from what we normally expect from desktop processors. The semiconductor industry is experiencing new challenges as engineers grapple with the fundamental issues of scaling down to smaller process nodes. In some cases, we can expect to see clock regressions as transistors shrink further. We could also see more variation in die yields, among many other factors. It appears that AMD is circumventing some of those issues by optimizing its binning process rather than allowing it to be reduced to the lowest common turbo bin. This tactic isn't entirely unheard of, we've already seen this trend emerge with some mobile processors, but it appears to now be making its way to the desktop PC.

You can view this in several ways. Some will view this as smart utilization of the silicon that AMD has to work with. AMD has apparently made some tradeoffs to deliver chips with a smaller geometry to market faster than its competitor, and it has used clever engineering at both the silicon and software level to pull off the feat. The more the software knows about the hardware underneath, the better.

But this does open up another line of questions. How does AMD decide what the minimum bar for a 'slow' core is? Logic dictates that would be the base frequency of the chip, but that would theoretically mean some cores on a 3600X would only operate at 3.8 GHz.

How much variation can we expect in the cores of chips purchased at retail? How would the binning impact performance in systems where the software optimizations aren't installed, or if the scheduler doesn't profile all workloads correctly even when it is installed? That could lead to erratic performance characteristics as workloads fall into slower cores. In our testing, we did observe that in some cases, but we caution that these results come from a single sample. It's hard to make definitive declarations with a sample size of one, especially given some of the small frequency variations we observed between cores. But AMD did confirm some of our findings.

We asked AMD if some cores are faster, and if the minimum requirement for a core is to reach the base frequency, to which AMD responded:

There are faster cores, as noted in Ryzen Master. All AMD processors are tested to ensure boost clocks and performance across various workloads meet the product definition.

Technically, AMD's only specified boost clock applies to a single-threaded workload, which you could argue means AMD only has to deliver a single core capable of delivering the maximum frequency. But, if there are several slower cores that can only reach the base frequency, that would surely impact performance in various multi-threaded workloads. We hope that AMD provides more clarity in the days to come.

Intel says that all of its cores are capable of reaching the Turbo Boost 2.0 specification outlined on its spec sheet. The company has complemented that base level of turbo performance with its Turbo Boost Max 3.0 feature on its HEDT chips, which, much like AMD's implementation, targets workloads at faster cores. Theoretically, if Intel accepted similar trade-offs in its binning process, would 10nm processors already be shipping? Or, has Intel already decided to make these adjustments on future products? We aren't sure of the answer to either of those questions, but it is food for thought.

This shift in AMD's policy could foreshadow broader changes for the industry in the future. AMD's binning could also improve as the 7nm process matures, but that will be hard to determine with no public-facing specifications for slower cores.

Our testing revealed several interesting facets, but we have much more to learn. For instance, the Ryzen 5 3600X comes with two cores disabled, which might be due to the normal defects that occur during manufacturing, but perhaps some of the cores didn't meet a performance specification that is unknown to us. It will be interesting to profile chips with all eight cores active to see how they measure up. Not to mention the 12-core Ryzen 9 3900X. We'll certainly be busy at our test benches.

Modern processors are certainly a marvel, but they are also mind-bendingly complex. Core parking, power gating, thermals, power delivery, and power density, among many other variables, can all impact performance and testing. Measurement tools can, too. We've done our best to ensure a stable test environment (thanks to Ian Cutress at AnandTech for a sanity check), but this is, in some respects, uncharted territory. You could experience different results than ours.

In the end, performance boils down to the results you see during your daily use. We recently gave the Ryzen 5 3600X an Editor's Choice award with testing based off this very processor, which we purchased at retail. There is no doubt these chips are capable performers, but there is much to learn about this new definition of what we can expect from the individual cores in our systems, even though it actually isn't defined at all.

Image Credits: Tom's Hardware, AMD


MORE: Best CPUs


MORE: Intel & AMD Processor Hierarchy


MORE: All CPUs Content

  • rhysiam
    Really interesting article Paul and team, thanks!

    IMHO this is a case of clever engineering mixed that's been muddied by dubious (at best!) marketing. Intel's approach of specifying both Turbo Boost 2 (which all cores can meet) and Turbo Boost 3 (the highest possible clock speed on any single core) is a more transparent approach and one that AMD should adopt.

    As you suggest, AMD likely have binning requirements on the "worst" cores, but these should be included in the spec table for the CPU. We should know what they are. On Z390 motherboards you can flick an "MCE" switch and it just works (provided you can cool it), because Intel have a minimum frequency that each individual core must be able to hit. That's the way it should be.

    Intel have work to do around transparency too: their stupidly low base clocks and lack of transparency around dual + quad core boost frequencies is similarly misleading at best. But on the issue of this article, AMD should be more open. AMD's response quoted in this article is totally inadequate. Hopefully you gain some traction and they decide to give you more than "blah blah blah meet the product definition".
    Reply
  • AlistairAB
    Wow this is fantastic work. I'm back to loving Tom's again. Remember Paul Alcorn ;)
    Reply
  • TJ Hooker
    Something to keep in mind is that some Intel CPUs may fail to hit their max turbo speeds as well, albeit for different reasons. Intel is more specific about what the max turbo is for a given number of cores loaded, and some CPUs will only hit their max speed on a single core. The prevailing theory seems to be that with a typical system there will always be stuff going on in the background, such that there's never truly only one core being loaded, and therefore you may never actually see the single core turbo speed.

    I know my 6700k won't hit 4.2 GHz at stock settings running benchmarks like P95 or Cinebench single threaded.
    Reply
  • AlexTSG
    Interesting reading this just a few days after the Silicon Lottery binning results.

    Given that AMD wants to use as much of the silicon they produce as possible, I wonder if we'll see 6 and 8 core processors that utilize two processor chiplets, where only 3 or 4 of the cores are active.

    I'm planning on building a new Ryzen 3000 system towards the end of the year, so I'm hoping that another 6 months sees some improvements made to the 7nm process, and possibly new processor steppings.
    Reply
  • rhysiam
    TJ Hooker said:
    Something to keep in mind is that some Intel CPUs fail to hit their max turbo speeds as well, albeit for different reasons. Intel is more specific about what the max turbo is for a given number of cores loaded, and some CPUs will only hit their max speed on a single core. The prevailing theory seems to be that with a typical system there will always be stuff going on in the background, such that there's never truly only one core being loaded, and therefore you may never actually see the single core turbo speed.

    I know my 6700k won't hit 4.2 GHz at stock settings running benchmarks like P95 or Cinebench single threaded.
    That's interesting, I wasn't aware there were regular issues on Intel's side too.

    To be fair though, presumably you can just flick the MCE switch (or equivalent) in your BIOS and, assuming you have a capable motherboard and cooler, you should hit 4.2Ghz on all cores? In other words, all four cores should be capable of sustaining 4.2Ghz at ambient-coolable voltages. That's simply not the case with the new Ryzen processors.

    I like where AMD are going with Ryzen and I'm saving my pennies for a 3900X, but I'm glad these boost scenarios are being explored and I hope they are more transparent with future releases.
    Reply
  • UncleElias
    admin said:
    Not to mention the 12-core Ryzen 9 3950X. We'll certainly be busy at our test benches.

    Typo?
    I'm looking forward to your testing of the 12 core 3900X and 16 core To be released 3950X.
    Reply
  • mdd1963
    "not all cores can reach the advertised boost frequency'?

    That's a pretty big understatement...

    It seems lucky for many users if even the best single core can hit only 50-100 MHz below it....for 1/4 second or so...and many seemingly are lucky to achieve even 100 MHz below it...

    All core? Most definitely, subtract 200-300 MHz...maybe 400 MHz...

    I'd not be surprised to eventually see some 'class-action' activity involving attorneys, frankly...; I don't think AMD's little 'you need to understand' sponsored lectures really help much.
    Reply
  • redgarl
    Okay, the CPU have been released for a mere 3 weeks. How do we know it is not related to BIOSes? You lost all credibility with Just Buy It!
    Reply
  • greenreaper
    AlexTSG said:
    Gen that AMD wants to use as much of the silicon they produce as possible, I wonder if we'll see 6 and 8 core processors that utilize two processor chiplets, where only 3 or 4 of the cores are active.
    It's more likely we'll see those put into Athlon as single chiplets, although what you envisage could be suitable for low-end EPYC designed for storage servers. Ryzen isn't the high-end brand, but it's also not the low-end.

    Of course, they could do Athlon as a no-graphics derivative of the Zen 2 APU instead, which may be a monolithic integrated design, sans chiplets, as in Zen 1.
    Reply
  • 13thmonkey
    Will this improve with time? Will next year's chips behave differently, given how close to the edge they are perhaps they will. Intel have sufficient headroom that you may not notice any beneficial drift, other than in the most extreme of max OC's
    Reply