Intel’s Binary Optimization Tool tested and explained — how the iBOT translation delivers up to 18% faster gaming performance, 8% on average

Intel Binary Optimization Tool
(Image credit: Intel)
23

Intel has something interesting going on with its new Binary Optimization Tool. Otherwise known as iBOT, it’s a translation layer along the lines of Microsoft Prism or Apple Rosetta. However, instead of translating instructions from one ISA (instruction set architecture) to another, it’s optimizing x86 applications to run more efficiently on Intel CPUs. As Intel puts it, it translates “other x86” to “Intel x86.”

It’s one of the marquee features on Intel’s new Core Ultra 200S Plus chips, the Core Ultra 7 270K Plus and Core Ultra 5 250K Plus. Although both show performance improvements strictly through silicon, at least part of their gaming boost comes courtesy of iBOT. It’s only enabled in 12 games right now, but Intel says iBOT is part of its long-term roadmap and that it’ll be a feature in Intel chips on desktop and mobile going forward. I wanted to see just how much of a boost iBOT represented, so I tested the feature in 10 of the 12 games it’s currently supported in, and I sat down with Intel’s Robert Hallock to better understand how the feature works and what it can offer.

Latest Videos FromTom's Hardware
TOPICS
Jake Roach
Senior Analyst, CPUs

Jake Roach is the Senior CPU Analyst at Tom’s Hardware, writing reviews, news, and features about the latest consumer and workstation processors.

  • Gururu
    It doesn't matter if it doesn't work on old processors. If Panther and Nova can utilize this, it will go a long way through that generation.
    Reply
  • cknobman
    So you only get the benefits if Intel takes the time to create the optimizations for a particular app/library?
    Reply
  • usertests
    Borderlands 3's 1% lows jumped from 73.67 to 120.64. That's nearly a 64% increase and you could play locked to 120 FPS forever if the GPU can keep up.

    A couple games had a tiny regression in 1% lows but it's not significant, probably a consequence of the CPU doing more work.
    Reply
  • TerryLaze
    cknobman said:
    So you only get the benefits if Intel takes the time to create the optimizations for a particular app/library?
    Making drastic changes on such a low level can have serious repercussions like crashes or corrupted data, so intel only allows it on things they tested long enough to be relatively sure that it's ok.
    Hopefully they integrate it into their compiler so that devs can skip the software part and compile directly for the new intel CPUs.
    Reply
  • bit_user
    The article said:
    Inside Arrow Lake Refresh CPUs, and Intel CPUs moving forward, there are hardware performance counters that can catch things like cache misses, branch mispredictions, spinlocks, and hardware interrupts.
    Such hardware performance counters certainly aren't new. Tools like VTune have used these for ages. It would be interesting to know which, if any, are new to Arrow Lake (Refresh). I'm skeptical it has much of anything that its immediate predecessors lacked.

    The article said:
    You could fix those issues by going back to source code, but Intel would then have to court the developer, convince them to spend time and money optimizing for its specific architecture, and then release a new binary that’s optimized for one specific architecture and not another.
    Yeah, you don't want to have a bunch of different code paths or builds that are specific to a bunch of slightly different microarchitectures, especially ones that basically all have the same ISA extenions.

    On Linux, there are some distros (like Gentoo, famously), where you compile all packages from source, at the time of installation. That way, you can specify -march=native and get a build that's optimized for your exact CPU. Languages that are Just-In-Time compiled, like Java and C#, do effectively the same thing.

    The article said:
    Then, Intel bundles up those optimizations as a profile, ships it out, and you’re on your way.
    Huh? Why does Intel need to create a profile? I figured iBOT would automatically profile your app, as you run it. The point of hardware PGO is that you can do this without much overhead.

    The article said:
    “Did the cache miss happen because that data wasn’t tagged with priority, and it just got evicted? OK, that is a place where binary optimization could certainly help,” Intel’s Robert Hallock says. “We could just tag it and say, ‘hey, stay local, don’t evict me.’ Pretty powerful capability.”
    LOL, I'm pretty sure x86 doesn't have a (non-privileged) instruction to do cache locking. Instead, he should've talked about iBOT issuing prefetch instructions.

    He shouldn't BS like this. It undermines the credibility of what else he says about it.
    Reply
  • bit_user
    cknobman said:
    So you only get the benefits if Intel takes the time to create the optimizations for a particular app/library?
    Sounds like it, but I don't really see why.

    TerryLaze said:
    Making drastic changes on such a low level can have serious repercussions like crashes or corrupted data,
    Only if the app has bugs like race conditions or use of uninitialized variables. I guess that's reason enough to use it with caution, but such issues could affect anyone running the same app via emulation.

    TerryLaze said:
    Hopefully they integrate it into their compiler so that devs can skip the software part and compile directly for the new intel CPUs.
    Compilers already have support for compiling for new CPUs or even detecting what CPU they're running on and compiling for that.

    What they could do is perhaps bundling the iBOT infrastructure into something you might directly link into your app.
    Reply
  • palladin9479
    cknobman said:
    So you only get the benefits if Intel takes the time to create the optimizations for a particular app/library?

    It seems to be a whitelist. Messing with binary code can have unforseen consequences, so they likely whitelist what they have tested to work.
    Reply
  • bit_user
    palladin9479 said:
    It seems to be a whitelist. Messing with binary code can have unforseen consequences, so they likely whitelist what they have tested to work.
    Yeah, kinda. But, like I said, it's not doing stuff that JIT-based emulators aren't already doing.

    It's not like in the bad old days of self-modifying code - you can't do that on modern CPUs! So, the risks are pretty much just limited to exposing latent bugs that the old instruction sequencing wasn't triggering.
    Reply
  • bit_user
    usertests said:
    Borderlands 3's 1% lows jumped from 73.67 to 120.64. That's nearly a 64% increase and you could play locked to 120 FPS forever if the GPU can keep up.
    It's possible there's some pathologically bad case that was impacting its lows that iBOT fixed up. There are faux pas, like split locks, that will absolutely kill performance, if you do them. Or, maybe it just somehow dealt with some cache thrashing.
    Reply
  • thestryker
    bit_user said:
    Such hardware performance counters certainly aren't new. Tools like VTune have used these for ages. It would be interesting to know which, if any, are new to Arrow Lake (Refresh). I'm skeptical it has much of anything that its immediate predecessors lacked.
    Hallock basically intimated some were added to ARL-S R which is why the backwards compatibility for ARL-S isn't there.
    bit_user said:
    Huh? Why does Intel need to create a profile? I figured iBOT would automatically profile your app, as you run it. The point of hardware PGO is that you can do this without much overhead.
    It requires DTT to function and while it's been added to the APO software it doesn't support all the games. They've also specifically blocked multiplayer titles due to anti-cheat concerns (APO is fine on multiplayer titles). They're obviously worried about unexpected behavior of some sort.

    An interesting note: IBOT PTL support seems random. I thought it was just not supporting anything other than the base Xe3 config, but it's missing the 366H and 336H too.

    https://www.intel.com/content/www/us/en/support/articles/000102604/processors.html
    Reply