HyperMemory

HyperMemory is essentially a memory management system. Put very simply, it's meant to extend the apparent memory capacity of a card by transferring data into the system RAM if needed. This is similar to the way Windows sometimes pages data in system RAM to the hard drive when more space is needed.

This system won't have much of an effect on higher memory cards. HyperMemory has already been implemented in earlier versions of the catalyst driver, but only for specially designated cards with less than 64MB of RAM. With Catalyst 5.7, ATI has included HyperMemory style memory management on 64MB and 128MB cards as well.

PCI Express motherboards will make best use of this system, but theoretically, you won't see as much of an improvement with HyperMemory on an AGP board due to its limited bandwidth. It definitely wouldn't hurt your framerates though, and in some cases, it will improve them, specifically on games with higher memory usage running 1600x1200 with AA/AF enabled.

In the past, there was some confusion about whether ATI's HyperMemory is hardware or software based. We suspected it to be mostly software even though there were a few products billed as HyperMemory cards. But with the latest catalyst release, we now find that HyperMemory is solely a software technology.

In the release notes for catalyst 5.7, ATI makes some pretty big claims about their latest driver. We were inclined to do some tests and see for ourselves what kinds of improvements the new drivers would have, and if they were really as dramatic as ATI said they would be. As the release notes also state, these performance gains only apply to 64MB and 128MB cards.

Here are some of the improvements that ATI claims are "most apparent" in the latest driver:
  • 5-15% gain in 3DMark05 on some products
  • 20-60% gain in Far Cry Regulator across multiple settings
  • 25-60% gain in Half-Life 2 at 1600x1200 4xAA 8xAF
  • 40-60% gain in UT2004 at 1600x1200 6xAA 16xAF
  • 50-100% improvement in Comanche4 at 1600x1200 6xAA 16xAF
  • 50-200% gain in certain ShaderMark tests at high resolution
As you'll see on the next page, we used this list to determine which games and settings to test. We didn't test 3dmark and ShaderMark because we generally don't care too much about performance increases in synthetic benchmarks. (It's not really a game, and no one is going to play it, so performance gains here aren't that important.) And we didn't test anything with 6xAA and 16xAF because 4xAA and 8xAF are our settings of choice for benchmarks. Besides, not as many people will be playing games with 6xAA and 16xAF at 1600x1200 on a 128MB or lower card, but if that's your thing, then go for it.

Index X600 128
POST A COMMENT

23 Comments

View All Comments

  • nserra - Tuesday, August 16, 2005 - link

    Why no one see if Ati claims are true?

    Well Why not test with the 6X AA....

    Maybe enabling AF will not make the same difference as not enabling it....

    Also I don’t understand the deal with always 4X AA.

    I play almost all games with no AA, or 2X AA with Temporal AA enabled and 4X AF or 8X. Some times I lower Mipmap Detail Level option to Quality because almost all games give 0% image quality difference. I do that with DOOM3 and got almost 5% performance increase.
    Reply
  • Jeff7181 - Sunday, August 14, 2005 - link

    ... memory optimizations help most when the memory is the limit. Nice to get some free performance... too bad you have to have a crappy card with crappy frame rates to begin with in order to see that free performance. Reply
  • Rand - Saturday, August 13, 2005 - link

    It would have been nice had you tested how this impacted AGP and PCI-E graphics cards respectively, you commented that it should of course provide a greater benefit over PCI-E.

    Assuming you don't have any motherboards natively supporting PCI-E and AGP (Not the neutered AGP over PCI) you could have used an nForce 3 and nForce 4 board, given they perform quite similarly.
    Any chance of doing any such tests?

    On another note- what system did you test the cards on anyway?
    I don't believe you made any mention of the system configuration, it's always beneficial to know the system specifications.
    Reply
  • OvErHeAtInG - Saturday, August 13, 2005 - link

    I did some extensive comparison benching with HL2 with 5.6 and 5.7 drivers. I run a 128MB 9800 Pro, 430/370, on a 4x AGP mobo, P4 2.85, 1GB DDR400 SC.

    I can second what they said about the 9700 Pro, as I had similar results. I play HL2 at 1280x1024, no AA. The only performance increases (as ATI specifies) come at that res and above WITH AA/AF enabled - which you do NOT play at with a 9800 pro.

    So, in other words, certain res/settings go from "unplayable" to "almost playable." Which is VERY impressive, but useless. I did see a 35-82% increase in frames (much larger than what AT got):

    HardwareOC Coast at 12x10, 4x8x:
    Catalyst 5.6: 36.3 fps
    Catalyst 5.7: 52.1 fps

    HardwareOC d13c17 at 12x10, 4x8x:
    Catalyst 5.6: 20.8 fps
    Catalyst 5.7: 37.9 fps

    At playable settings, we get a SLIGHT decrease in performance:

    Guru3d Demo4 12x10 noAAnoAF
    Catalyst 5.6: 103.3 fps
    Catalyst 5.7: 100.3 fps
    Reply
  • Shadowmage - Sunday, August 14, 2005 - link

    Of course, this depends on what you call "playable" and "unplayable".

    For me, anything above 40fps is considered extremely playable.
    Reply
  • OvErHeAtInG - Sunday, August 14, 2005 - link

    Yes... if it stays above 40 fps :p Reply
  • JarredWalton - Saturday, August 13, 2005 - link

    Strange... I took almost no performance hit when enabling 4xAF on my old 9800Pro with HL2. (I don't have it anymore, sorry - no new benches.) Going from 1024x768 to 4xAA brought less than a 5% decrease in FPS, while 1280x1024 was about a 30% performance decrease. Enabling 4xAA at 1280x1024 was another 5 to 10% loss. Of course, that was last year with 4.10 or so Cats, so I don't know what happened in the intervening time. Reply
  • OvErHeAtInG - Sunday, August 14, 2005 - link

    4xAF only? Never tried it. Probably keeps it playable.

    I was using the HardwareOC benches at the time which seem to get held back by my CPU (which is weird), that's why I was running 12x10 4x8x to stress the card.

    At 10x7, one can run AA AF on this card, but not 12x10... if it's still smooth enough for you, different strokes different folks
    Reply
  • pxc - Saturday, August 13, 2005 - link

    Performance went up on my XPRESS 200M 128MB w/HM laptop when I disabled the HyperMemory in the registry. :p I got rid of it before Cat 5.7 came out, so maybe the HM performance problems were fixed. Reply
  • AlexWade - Saturday, August 13, 2005 - link

    ATI has had final 64-bit drivers for a while. Is it possible to benchmark those compared to the same 32-bit? Reply

Log in

Don't have an account? Sign up now