AMD details potential Ryzen performance gains from AGESA 1.0.0.6

Remember that AMD AGESA update (version 1.0.0.6) that Jeff told y'all about back in May? It's finally finding its way to end users by way of motherboard firmware updates. Some gerbils have already gotten theirs, but if you have a Ryzen machine with some hot memory that hasn't been updated—and a few hours to spend tweaking the machine—you should probably go ahead and do that. AMD recently updated its gaming blog with some testing results from the new firmware, and at least with the configuration AMD's tester was using, there's some serious performance potential laying on the table versus stock settings.

Before we begin, it's worth noting that these are manufacturer-provided numbers, and they should all be taken with the appropriate grains of salt. That said, AMD had veteran overclocker Sami Mäkinen fiddle around with a Ryzen 7 1700 CPU mounted on an Asus Crosshair VI motherboard along with some fancy RAM. Unfortunately, the exact models of memory modules aren't specified in the post, but the testing was performed with both single- and dual-rank modules. The dual-rank modules maxed out at a data rate of 3200 MT/s, while Mäkinen achieved 3520 MT/s with the company's chosen single-rank modules.

The AGESA 1.0.0.6 update allows motherboard vendors to offer toggles in their setup utilities for a couple of new options: Bank Group Swap (BGS) and Gear Down Mode (GDM). The first option has an effect on how physical addresses are assigned to applications, while the latter option may improve memory compatibility by reducing the effective data rate on the command and address buses. AMD found that Bank Group Swap actually improves the performance of synthetic applications when it is enabled, while disabling it offers a minimal-but-measurable improvement in gaming performance.

The effects of the Gear Down Mode option appear to be a little more complicated. While it is enabled, setting the "command rate" option for memory timings is pointless; it will always effectively be "2T" (although the company notes it's not quite the same thing—1T versus 2T command rates may still be worth exploring on their own). Disabling GDM will allow memory overclockers to manually set the DRAM command rate to "1T" and realize the full benefits of doing so. Tweakers have known for a long time that a 1T command rate can have huge effects on memory performance, and the Ryzen platform is no exception. However, if your memory can't run a 1T command rate at its spec'd clock rate, or if you're overclocking and can't hit a high clock with a 1T command rate setting, you could just be making your system less stable by disabling GDM.

AMD tested the effects of most of these settings, as well as a few memory myths. The questions of whether auto tunings are "good enough," whether builders should prefer single-rank or dual-rank memory, and whether tweakers should prefer a high frequency or tight memory timings timings (at the expense of data rate) are also addressed in the post. Naturally, the answer to every one of these questions is "it depends." In general, for the best gaming performance, AMD says you want a good balance between a high transfer rate and tight memory timings, and that's going to take some hand-tuning.

The best performance AMD's overclocker found was using single-rank DIMMs at 3466 MT/s with CAS latency 14, hand-tuned sub-timings, GDM and BGS off, and a 1T command rate. Overall, comparing the worst case (DDR4-3200 CL14 with auto settings) to the best case above, Mr. Mäkinen garnered a 10% improvement in Rise of the Tomb Raider framerates, a 16% improvement in Hitman performance, and a 9% improvement in 3DMark Sky Diver scores. That's perhaps a small gain for some hours' work, but it's a nice upgrade for exactly zero dollars. Check out AMD's blog post for the full details on the performance gains potentially available in AGESA 1.0.0.6.

Comments closed
    • gerryg
    • 2 years ago

    Soo… based on this information what would be some of the best memory sticks to start tuning with? Is there a potential value pick that would tune well?

    • lordcheeto
    • 2 years ago

    The general rule of thumb stands – the lower the access time, the better.

    2000 x (CL / Speed) = access time (ns)

    • HisDivineOrder
    • 2 years ago

    I am glad AMD is giving us 8 cores, 16 thread CPU’s for < $300. That’s great. My biggest problem is I don’t want a motherboard that doesn’t have a good BIOS from the start and I wish you could just buy memory and have it work at the speeds it’s supposed to work at on other platforms.

    As a result, I basically want AMD to do a soft reboot of the launch at some later date, possibly when they release a new stepping of CPU, where all the motherboards have the latest and greatest BIOS’s that include proper memory support and new naming so that I don’t have to wonder if I’m going to get a motherboard from launch that isn’t ready.

    Consequently, I’ll probably be waiting until Zen 2 just to be sure.

      • ronch
      • 2 years ago

      Waiting for the smoke to clear too. No rush to upgrade so I’m just sitting back with some popcorn.

      • ptsant
      • 2 years ago

      Just buy Ryzen-validated memory and you should be fine. If you are willing to tweak, almost any Samsung B-die, single-rank memory can achieve more than 3200, usually 3400. There are very few benefits to going above 3400 with Ryzen.

      The fact that RAM speeds is coupled with the fabric speed internally means that gains from 2133 -> 3000/3200 are major, but also means that exotic memory speeds (4000+) are not achievable without BCLK overclocking (decoupling ryzen clock from system clock). Usually not worth it.

      So, in my opinion you should strive for 2800-3200 and this is something that can readily be achieved with validated RAM or careful tweaking.

      • DoomGuy64
      • 2 years ago

      [quote<]I wish you could just buy memory and have it work[/quote<] What a pebkac statement. Anyone who's done any reading would know that it already works exactly like you are asking after the bios update, and this doesn't mean boards with older bioses are defective either. The older bios boards will boot just fine, and you update it afterwards to max out your timings. So hard. Even Intel requires some basic bios setting changes to enable XMP and other performance enhancements, so it's not like there is any board that magically does it all for you. Asus had 3200 working before this update, btw. The new bios merely allows SR speeds higher than 3200, DR to hit 3200, and gives access to extra settings. Anyone who can't handle a bios update that happens automatically with most motherboards automatic update tools (Asus EZ Update), and enabling XMP mode shouldn't be even posting in a PC building site. Just buy a prebuilt if you're that incompetent, because I can't see how you can even install ram in the slots with that mentality.

    • ozzuneoj
    • 2 years ago

    When I see things about fast memory finally making a performance difference, I just think about how lame it is that DDR4 prices are through the roof. If 16GB of good DDR4 was the same price that 16GB of good DDR3 used to be, Ryzen with a bunch of super fast RAM would be much more attractive. As it is, the high memory cost just eats away at the price advantage of Ryzen, and makes upgrading at all (Intel or AMD) a lot less attractive for the budget conscious builders out there.

    I bought 16GB of DDR3-2133 for $60 two years ago. Now, 16GB of DDR4-3200 (fast but not top tier by any means) is $125-$150.

    I just priced up a quick and dirty Ryzen build and it’d be $436 for an R5 1600, an Asrock B350 board and 16GB of G.Skill 3200. If the RAM pricing was more like 2015 pricing (taking into account a similar grade of memory), that build could be under $370.

    • albundy
    • 2 years ago

    i haven’t seen much in the same regard. 3200mhz hynix ram is still NOT fully supported. reddit forums are filled with people unable to get past 2677mhz stable with cas 15/16/17/18 hynix sticks. 1.0.0.6 was supposed to fix that but it seems that it didn’t. many of the motherboards clearly advertise support for 3200mhz speed as well as ram makers. if my memory serves me right, the first release of ddr4 ram at this speed was 2 years ago. what say you, amd?

      • ultima_trev
      • 2 years ago

      Maybe it’s the people building these systems not knowing what they’re doing as opposed to AGESA updates being borked? Maybe they didn’t even bother flashing their BIOS to the latest version?

      I had zero effort getting my cheap @$$ ADATA DDR4-2400 17-17-17-39 kit running at DDR4-2666 14-14-14-32 after I updated my relatively cheap @$$ GA AB350 Gaming Vanilla to the most recent BIOS update.

    • synthtel2
    • 2 years ago

    [quote<]Tweakers have known for a long time that a 1T command rate can have huge effects on memory performance[/quote<] Then why, in TR's Ryzen 5 reviews where Jeff said matching memory settings was a priority (and the resulting memory settings are otherwise well-matched), do Sky/Kaby systems run at 2T while everything else runs at 1T? (I know this isn't the right place to ask, I just never got a response last time.)

      • chuckula
      • 2 years ago

      Because Intel illegally bribes TR to double the T value!

      T means throughput right?
      So 2T is twice the throughput as 1T right?

        • Redocbew
        • 2 years ago

        Good punchline, but your timing is off.

      • DrDominodog51
      • 2 years ago

      Explanation here to be rewritten.

        • synthtel2
        • 2 years ago

        Those reviews have Ryzen running 1T and Sky/Kaby running 2T, not the other way around.

          • RAGEPRO
          • 2 years ago

          This doesn’t really answer your question but I can sidestep it somewhat by pointing out that prior to AGESA 1.0.0.6 Geardown Mode was enabled by default (as per DDR4 spec cf. JEDEC) so it was 2T anyway regardless of setting.

            • synthtel2
            • 2 years ago

            I wasn’t too concerned about the review results to start with, but that does make that even less of an issue. The question is more about the process of selecting this stuff, though. There are a lot of mundane potential causes, and I’m just curious which one it was.

          • DrDominodog51
          • 2 years ago

          Yeah. I might have messed up somewhere in that explanation. I’ll rewrite it when I’m thinking straight.

            • synthtel2
            • 2 years ago

            No worries. It looked like a pretty solid answer, just to a slightly different question.

          • DrDominodog51
          • 2 years ago

          I think it was just a typo.

          If the memory was capable of running those timings on a 1T command rate, the 7700k’s memory controller could certainly run it at 1T.

          Also, what Zak said above.

            • synthtel2
            • 2 years ago

            A typo seems likely, but if that’s what it was, it would be nice if Jeff said so.

    • odizzido
    • 2 years ago

    It’s nice to see the continuing improvement of zen. I know they keep happening so it might make the article out of date rather quickly, but I would be interested in seeing ryzen performance revisited.

    • ptsant
    • 2 years ago

    My experience with AGESA 1.0.0.6 has been exceptionally positive. The platform feels polished for the first time.

    I had bought the worst-case memory before Ryzen launch (it was on sale, 32GB for $199, which is very cheap for EU, now sells at $299). These are 2x16GB, dual rank with Hynix chips sold as 3000C15. However, dual rank according to AMD are guaranteed to run at 2133 (which it did, automatically) but not much more.

    With AGESA 1.0.0.4 I got 2400 without much tuning and 2667 with some tweaking, but it wasn’t exactly stable. One of the problems is that the platform cannot set the DDR voltage before, you know… booting, so the system could manage 2667 on a hot reboot (ie restart) but would fail 50% of the time from cold boot (switch on).

    With AGESA 1.0.0.6 I got 2667 easily and, using a bunch of cryptic options, I managed 2800 really stable, without any sort of overvolt or overclock. I could easily push 3000+ with overvolt and BCLK overclocking, but didn’t want to do that. The end result is 2800CL15 at stock (XMP) 1.35V and 2T.

    The difference, using the very simple CPU-Z bench, was >5%. According to the same, this puts me at a couple percent faster (2-3%?) than a reference 1700X.

    Anyway, the current Asus Crosshair BIOS has gotten to a stage that I consider “polished”. Good job, better late than never.

    Edit: Helpful blog post from AMD explains some of the stuff and how to tune it:
    [url<]https://community.amd.com/community/gaming/blog/2017/05/25/community-update-4-lets-talk-dram[/url<]

      • Blytz
      • 2 years ago

      I agree. I ran a few different bios revision including a couple betas on my crosshair and with my vengeance lpx 3200 I could get it to boot at 2400 regulatly, but soon as I punched it up to 3200 it would run fine, reboot fine, but default bakc to 2133 once it cycled a cold boot.

      Now run 1403 (1.0.0.6) it’s ‘xmp’ profile (I’ve forgotten the ASUS acronym – like DOCP – thanks MOSFET) picked it up at 3200 and boots at 3200 every time. Notable gains I’ve gotten in league of legends (I know it’s esport and low stress) at 1080p with everything running flat out, the main map of summoners rifts used to see pull 200-240fps (topping out at 250) at the same time frames I am now seeing spikes of just over 300.

      In game under load the spike has been equally 10-20% better and it feels smoother.

        • MOSFET
        • 2 years ago

        D.O.C.P. – Dynamic OverClock Profile

        I have an Asus Prime X370 Pro with an R5 1600 running as my home ESXi and vSphere host. With the ability to have 64GB in one system, I no longer need two hosts, and the power savings are pretty tremendous! I got the G.Skill Fortis 2400 64GB 16×4 kit. Even with all 4 DIMM slots populated, it runs XMP(DOCP) 2400 CL15 no problems. It passed memtest86+ v7 with Hammer tests producing 0 errors on the first BIOS, and there has not been once moment of trouble with ESXi. Since installation, I have upgraded BIOS twice, with the most recent containing AGESA 1.0.0.6 (I believe). Memory has stayed stable, and I even bumped the CPU up to 3.6 GHz all cores, leaving turbo on with the included Wraith cooler, although I don’t see any P-states above 3.6 available in esxtop. Unlike most systems, I didn’t install Windows first to benchmark, just went straight to ESXi 6.5. Figured I would just benchmark from Windows VMs. One other quick note from that build, SATA performance is the perhaps the best I’ve ever experienced in ESXi. The vmw_ahci module actually works like a real driver with X370!

    • Voldenuit
    • 2 years ago

    10% performance gain from memory is huge, especially in the age of high-refresh monitors. I’m assuming that’s at 1080p, and presumably not on Ultra settings.

    Nice to see Ryzen maturing nicely.

      • DPete27
      • 2 years ago

      How user-friendly is this to lay-persons though? Remember what you’re looking at is performance numbers from a professional overclocker.

        • derFunkenstein
        • 2 years ago

        It’s not super friendly, but if AMD-specific kits have a couple XMP profiles that will let users flip a switch and overclock their RAM, then this is great. It’s way better than it was at launch, although still not perfect.

        • Voldenuit
        • 2 years ago

        I suppose the ‘glass half empty’ viewpoint would be that Ryzen is unreasonably sensitive to memory timings for a modern CPU, exacerbated by the low L3 bandwidth and high latency between CCX clusters.

        But the architecture looks pretty promising as something that can challenge intel.

        • ptsant
        • 2 years ago

        I can tell you from experience tuning AGESA 1.0.0.6 that I had to tweak some stuff that I had never, ever before heard about, including stuff with scary names (Proc On-Die Termination resistance, for example).

        It was fun but this is, without any doubt, much, much more complex than punching frequency and CAS latency.

          • wingless
          • 2 years ago

          I remember how alien Intel terminology was when I got my Sandy Bridge 2600K way back then. We all got used to it, didn’t we? Now it’s time for us to shake it up a little and get used to an all new way of overclocking.

          It’s not scary, it’s exciting! We finally get to change it up a little and learn something new.

            • Voldenuit
            • 2 years ago

            Just wait until every mobo manufacturer has different names for the same settings.

            • derFunkenstein
            • 2 years ago

            They already do.

            • ptsant
            • 2 years ago

            I agree. I was a bit annoyed with the roughness of the Ryzen platform in the beginning but in the end tuning all that stuff was fun. I used to spend much more time doing that when I was a poor student and MP3 playback meant 30% CPU utilization.

Pin It on Pinterest

Share This