Back to Article

  • Freakie - Saturday, June 01, 2013 - link

    Ah yes, yet again, not VT-d for K parts.

    Maybe I'm just the only one who thinks that people concerned about performance enough to get a K part would actually want and use VT-d? I mean sure, businesses wont be equipping their workstations with K parts, but there are plenty consumers who do visualization AND like high performance parts. I can understand leaving out TXT and vPro but common Intel... Excluding VT-d just seems like a move to purposely make sure there is still ill-will with customers.
  • AccountCreationUnsuccessful - Saturday, June 01, 2013 - link

    Agree. I'm currently in the AMD camp and was looking forward to switch to Intel with the arrival of Haswell.

    I didn't know they disabled VT-d on K processors, that's a real bummer for me. Of all the people I know IRL I'm the only one who uses either unlocked processors or virtualization, and I use them both together. I'm now doubting wether I should still go with the locked Haswells or stick in the AMD camp a little longer to see if things get better here, with the understanding that I'll not hold my breath...
  • DigitalFreak - Saturday, June 01, 2013 - link

    To force you to buy Xeon models. Reply
  • AccountCreationUnsuccessful - Saturday, June 01, 2013 - link

    That makes no sense. Any Xeon is by definition a locked processor, and the locked i7/i5 models have virtualization just like the Xeon.

    If they really wanted to push Xeon they'd do the opposite of what they're doing: VT-d only on K processors (which are not available in Xeon) and only Xeon for all else.
  • coolbho3k - Saturday, June 01, 2013 - link

    Or, maybe to force you to buy Enthusiast models (which are now an unprecedented TWO microarchitecture generations behind) if you want an unlocked CPU AND VT-d? Reply
  • Klimax - Sunday, June 02, 2013 - link

    Price of Xeon validation. (Same as eq Xeon, but with some things fused off)
    LGA 2011 is same track as workstation/server chips.
  • Zinc64 - Saturday, June 01, 2013 - link

    VT-d is disabled on the current K's like the 3570K.
    Virtual machines run just fine with VT-x enabled on the 2600K, 3570K and 3770K.

    What if any benefit would VT-d add for the average user?
  • AccountCreationUnsuccessful - Saturday, June 01, 2013 - link

    For the average user? Maybe little. The point Freakie and me are making is that someone with a K processor is NOT likely an average user, and we are the most likely to need VT-d at least.

    In my case it's niche hardware with Windows-only software.
  • patrickjchase - Monday, June 03, 2013 - link

    VT-d allows virtual machines to directly access "dedicated" peripherals like network controllers, disk controllers, etc. By "dedicated" I mean that each such peripheral can only be used by either the host or by one virtual machine. As such it's mostly applicable to server virtualization under ESX/Vsphere and similar "bare-metal" platforms.

    VT-d also tends to add overhead to transfers to/from PCIe devices (for example I've seen ~40% performance hits to zerocopy GPU<->host transfers with VTd enabled) so you really don't want to enable it unless your machine is dedicated to server virtualization. I therefore expect that 99.999% of customers for "K" series processors would be better off disabling it even if it were supported.

    Going a bit deeper, the problem with allowing VMs to directly access peripherals is that a "rogue VM" can configure the DMA in a peripheral to read or write arbitrary memory addresses, and thereby violate VM isolation. VTd addresses that my putting an IOMMU (basically a paged memory management unit with TLB) between the peripheral and memory. The hypervisor (a fancy name for a virtualization platform like VMware) then configures the IOMMU so that the peripheral can only access memory owned by its assigned VM.

    The benefit to "dedicating" peripherals to VMs is that you no longer have to "trap out" to the hypervisor in order to do I/O, so in principal it reduces the CPU overhead associated with VM I/O ("in principle" because that overhead is already fairly low for devices that can do vectored I/O). As noted above, the cost is that all peripheral accesses (whether from VMs or the host OS) now go through an MMU, and it's entirely possible to thrash that MMU's TLB with "unfriendly" access patterns.
  • iLloydski - Sunday, July 14, 2013 - link

    stupid question, im planning to game on VMs - do I need these VM features to benefit from the optimal performance? Reply
  • Vinas - Sunday, September 01, 2013 - link

    No, VT is included with most (all?) intel cpus now a days. VT-d is a subset of commands that gives direct I/O to guest OSes. It's not required to use virtualization technology, and in most cases will drastically impact the performance of your non-server host OSes as VT-d can create competition for direct access to disk, graphics, memory, etc. Reply
  • Vinas - Sunday, September 01, 2013 - link

    Thank god someone finally pointed out how useless VT-d is for non-dedicated workstations. Reply
  • critical_ - Sunday, June 02, 2013 - link

    Took the words right out of my mouth.

    With the advent of "K" series processors I have been stuck with trying to address this issue. I rotate old hardware to the role of a home server. These days "home server" also implies virtualization. My use cases necessitate passthrough functionality provided by VT-d for all-in-one boxes with large storage arrays. Like many enthusiasts, we dabble with overclocking, virtualization, and can use ECC memory. I don't want or need a Xeon CPU but Intel has found an effective way to upsell me to this line of processors. Here are the non-ideal solutions I'm stuck with:

    1) Buy Intel Exteme Editions CPUs. This will get you VT-d but won't get you ECC. Both are functions of the memory controller. I have bought laptops with the 2920XM, 2960XM, and now the 3940XM (I'd kill for ECC in my laptop). I find Intel's market segmentation to be annoying (the CrystalWell issue is a manifestation of the same problem but in a different area) but, like I said above, it effectively leads to #2...
    2) Buy Xeon E3 processors. Xeon E3 gets you VT-d and ECC but the CPU is locked. I have a Xeon E3-1240v2 and it is just as quick as a 3770. Still not good enough?
    3) Buy a Xeon E5-1660. This is an unlocked Xeon (it is a Core i7 3960X) with VT-d and ECC. It has 6-cores and can run gobs of memory. I have one with 128GB of RAM running at stock frequency.
    4) Get a massive multi-CPU 8-core system in the Xeon E5-2600 series but they are all multiplier locked but have all the other features.

    Intel will offer a Haswell-based Xeon E3-1200v3 series eventually. I'm sure (or I'm hoping) an Ivy Bridge-E replacement for the E5-1660 will come. However, Intel has no 8-or-more core "ultimate" part which has every feature under the sun. If they did then it would probably cost well above Xeon E5-2687W levels but some of us might buy it. Quite frankly, I wish Intel would throw enthusiasts like me a bone with a Core i7-4770KX part with vPro, TXT, VT-d, ECC memory, and an unlocked multiplier.

    Back to the original issue. A "K" part can run VMs on it later on in its life as a home server but do the fancy ECC and VT-d stuff on a separate Xeon box. For example, ECC and VT-d benefit a ZFS box which can run a storage array that can be shared over iSCSI or NFS with a Core i7-4770K box running only VMs. It will work but it will require two computers and some compromises. Right now, Intel gets extra money out of us. Perhaps that will change. Perhaps not.
  • Ktracho - Friday, June 07, 2013 - link

    How can you tell if a Xeon CPU is unlocked? I didn't see anything on Intel's website indicating the E5-1660 is unlocked. Reply
  • Vinas - Sunday, September 01, 2013 - link

    If you're buying non xeon CPU's to run i/o heavy servers with 64GB + RAM, you need to re-evaluate your hardware priorities. I call BS on your home server vm needs and reasoning for "needing" a XK CPU. Expand horizontally and provision your CPUs correctly and you will not need a single CPU @ 5GHz on your host. Reply
  • stennan - Saturday, June 01, 2013 - link

    Jarred, if i remember correctly Intel is loosening up the possibility to overclock the "non-k" SKU:s, .
    If i understood the article correctly, "Clock Speed = BCLK * Ratio", so we can increase the BCLK and thus increase the clockspeed. But if the CPU cant handle that frequency we have the option to lower the ratio (Multiplier?) too reach a stable result.

    Question (for future article perhaps) : Will the non-k variants have the same potential as the k-variants, or is BCLK overclocking more inferior compared to increasing the multiplyer if one is aiming for 4-4.5 GHZ?
  • Omoronovo - Saturday, June 01, 2013 - link

    BCLK in Haswell, just like BCLK in sandy and ivy bridge processors, isn't just used to determine the CPU frequency. If you increase this frequency by more than a few mhz, you would be pushing the limit of these components. As such, it is not the cpu that is the limiting factor in BCLK overclocking, it's everything else that's tied to this frequency.

    Right now, you can overclock any non-k intel processor by increasing the BCLK, but the tight integration of components that run off (directly or indirectly) the BCLK frequency since sandy bridge means the system becomes unstable very quickly, regardless of which multiplier you set.
  • Meaker10 - Saturday, June 01, 2013 - link

    Except there is a new ratio.

    CPU = Base clock x CPU multi
    Everything else = Base clock x their multi x ratio

    So with the 5:6 ratio the CPU sees 125mhz x CPU ratio where as everything else sees 5:6 x 125 x ratio which = 100 x their ratio.
  • stennan - Saturday, June 01, 2013 - link

    Exactly, (100/125/167MHz), but some sites mention that intel scraped BCLK overclocking on non-k models. Probably was to effective :/ Reply
  • Omoronovo - Saturday, June 01, 2013 - link

    I misunderstood the original poster - modifying the BCLK strap in haswell still requires a K-Series CPU, which is the point I was making. Regardless of the modifications made to the bclk in increments, you'll still it a wall after a few mhz. Reply
  • todlerix - Saturday, June 01, 2013 - link

    BCLK in sandy/ivy is only adjustable by a small percent, 10% is the most I've seen stable usually it is 3-6%. Reply
  • LancerVI - Saturday, June 01, 2013 - link

    What's the street date for retail?? Reply
  • 8steve8 - Sunday, June 02, 2013 - link

    any products coming to market with 4770R? Reply
  • vailr - Tuesday, June 04, 2013 - link

    Any information about Xeon Haswell CPU's compatibility with desktop Z87 boards?
    I'm thinking they probably are compatible, but can't confirm that by looking at Gigabyte's web site.
    Performance comparison between Xeon & "desktop" Haswell CPU's would also be interesting to see?
  • tomsailor - Wednesday, June 19, 2013 - link

    Now that I can buy a Core i7-4770 with VT-d, TSX and hardware RDRAND (e.g., NewEgg), I'd really appreciate a review of that full-featured Haswell desktop CPU! These features are more important to this geek than overcloc-King; your readers may include like-minded geeks who really do want to try XBEGIN or send truly random bit-streams to friends and drive NSA crazy:) Reply

Log in

Don't have an account? Sign up now