|
adorai posted:Does anyone really care about "typical" single threaded workloads any more? More and more apps are being written with multithreading in mind, and even desktop computing seems to be moving towards virtualization (see: XP mode). Going forward single thread performance will be less and less important across the board. Most of the money is made in the enterprise, where max memory and number of cores are king. Until Autodesk actually gets off their rear end and writes their software to utilize the hardware that has been in every computer for the last 6 years, per-core performance is still going to matter tremendously.
|
#
¿
Jan 12, 2011 19:15
|
|
|
|
| # ¿ Apr 19, 2024 00:21 |
|
|
LiftAuff posted:Isn't Zambezi aimed at the enterprise/enthusiast market? Rumor has it that they are even bringing back the FX name for it. Sounds to me like they are going after the high end market first. The problem with AMD going after the high-end market is that product price points are down on the list of consideration after performance and over-clockability, where it looks like Sandy Bridge is going to win out. The profit margins are much better there and all they have to do is bin the better chips appropriately so it is a no-brainer for them to do it, but I seriously doubt it will compete with Intel's high end in any serious way.
|
|
#
¿
Jan 18, 2011 14:38
|
|
|
rscott posted:The problem with Sandy Bridge and the OC market is they're putting some weird rear end features together. Wanna use the faster integrated graphics in the K series chips and overclock at the same time? Gotta wait for the super expensive z67 or w/e chipset to come out! It doesn't make sense and you know Intel wants to kill overclocking and force everyone to buy the more expensive chips but they can't because they would be ceding mindshare (not really marketshare) to AMD. Do you really see a situation where people are going to be looking to overclock and use integrated graphics? Virtually anyone that is going to be overclocking is also doing 3rd work that will warrant discrete graphics in the first place.
|
|
#
¿
Jan 19, 2011 14:44
|
|
|
rscott posted:Isn't the whole point of the improved integrated graphics to make it not necessary to have a discrete graphics card unless you want to do high end gaming/CAD work or whatever? I know a fair (maybe 20% of my "nerdy" friends) amount of people who have pretty high end CPUs and poo poo GPUs because they don't really play games, they just run a lot of multithreaded applications. And regardless, why include the 3000 series GPU with the K series processors when, if you what you say is true they're rarely going to be used in the first place? The whole thing just doesn't make a lot of sense. I saw the point as raising the low-end of graphics hardware threshold while lowering cost and giving better overall value to customers. Integrated GPUs on the K series strikes me as an easy way to cash in on chips that are binned higher, nothing more.
|
|
#
¿
Jan 19, 2011 16:17
|
|
|
Aleksei Vasiliev posted:This article is utterly terrible. It doesn't bother to test x264, despite x264 being the best encoder for H.264. But the second file name (higher quality one) is quicksync.png e: Hahaha nevermind, I just saw this bit quote:Quick Sync maintains color fidelity but loses the sharpness of the x86 path Mmmm yes all those wonderful aliasing details BangersInMyKnickers fucked around with this message at 16:37 on Jan 21, 2011 |
|
#
¿
Jan 21, 2011 16:34
|
|
|
GRINDCORE MEGGIDO posted:Isn't that before and after the same as disabling HT? I suspect Game Mode is doing something to limit background process CPU activity or bump process priority for things it identifies as games in addition to avoiding the HT cores. There is a lot of CPU activity occurring on logical cores 8 and 9 on that first screenshot which makes me thing there was something bad with the test methodology, though it does fairly accurately represent a real world scenario. That's going to cause a lot of latency problems with multithreaded loads as HT pipeline optimizations do their thing. The HT cores seem to be impacting performance primarily on cores 0,1,4,5 which is probably an artifact of the cache layout of the processor that MS also optimized for.
|
|
#
¿
Mar 20, 2017 22:44
|
|
|
Blorange posted:I've done some benchmarking on Ivy bridge Xeons which we bought to make up for some terrible query writing, and disabling power management boosted performance by 25%. Even with the hardware monitor telling you that all of your cores are running at boost clocks, if you let it the system clock itself it will hamstring your performance by downclocking cores for milliseconds at a time. I don't know if Intel has improved on this with newer processors, but it isn't at all surprising to see it hit AMD too. You generally need a very poorly written application where its executing in incredibly small increments with pauses to have that kind of performance penalty. Even back on the first processors started supporting speedstep the time to transition between states was in the nanoseconds and since then intel has done more work to optimize against "fluttery" workloads that have this kind of problem. Its not impossible, but it is uncommon and an edge case where most people are better off leaving the power management to the OS or whatever OEM power management chip they put on the board. Turning power management off will make things idle at a higher temp and impair access to P-states when it needs to ramp up for a sustained burst workload.
|
|
#
¿
Mar 21, 2017 03:20
|
|
|
redeyes posted:4k Q1 is where nearly all home systems sit in terms of i/o load. ATTO on their benchmark indicated half the sustained reads that Intel gets. No, not really. Typical IO size for a desktop is going to float around 24-32kb/op. If you break down the throughput test on crystalmark to ops you'll see the AMD chipset is pulling 16k read and 51k write which is exceeding the specs for samsung SSD they're using. Intel is probably doing some kind of ram buffering/caching to get those numbers, which is concerning since it increases the risk of data loss. And that is magnitudes more IO performance than most home users will ever need, the impact in any real world use case is going to be minimal. Maybe if you're running a large SAP dev instance on your home computer you should think about it but for everyone else: doesn't matter.
|
|
#
¿
Apr 2, 2017 18:48
|
|
|
|
| # ¿ Apr 19, 2024 00:21 |
|
|
RST buffers to another block device, so that doesn't make sense here. It's something in the driver and using some kind of ram cache or write combining scheme seems most likely. You don't just magic 50% over the OEM's write OP spec out of nowhere.
|
|
#
¿
Apr 2, 2017 19:33
|
|