For many years now, processors have been working not with physical memory addresses, but with virtual addresses. Among other advantages, this approach lets more memory be allocated to a program than the computer actually has, keeping only the data necessary at a given moment in actual physical memory with the rest remaining on the hard disk. This means that for each memory access a virtual address has to be translated into a physical address, and to do that an enormous table is put in charge of keeping track of the correspondences. The problem is that this table gets so large that it can’t be stored on-chip—it’s placed in main memory, and can even be paged (part of the table can be absent from memory and itself kept on the hard disk).
If this translation stage were necessary at each memory access, it would make access much too slow. As a result, engineers returned to the principle of physical addressing by adding a small cache memory directly on the processor that stored the correspondences for a few recently accessed addresses. This cache memory is called a Translation Lookaside Buffer (TLB). Intel has completely revamped the operation of the TLB in their new architecture. Up until now, the Core 2 has used a level 1 TLB that is extremely small (16 entries) but also very fast for loads only, and a larger level 2 TLB (256 entries) that handled loads missed in the level 1 TLB, as well as stores.
Nehalem now has a true two-level TLB: the first level of TLB is shared between data and instructions. The level 1 data TLB now stores 64 entries for small pages (4K) or 32 for large pages (2M/4M), while the level 1 instruction TLB stores 128 entries for small pages (the same as with Core 2) and seven for large pages. The second level is a unified cache that can store up to 512 entries and operates only with small pages. The purpose of this improvement is to increase the performance of applications that use large sets of data. As with the introduction of two-level branch predictors, this is further evidence of the architecture’s server orientation.
Let’s go back to SMT for a moment, since it also has an impact on the TLBs. The level 1 data TLB and the level 2 TLB are shared dynamically between the two threads. Conversely, the level 1 instruction TLB is statically shared for small pages, whereas the one dedicated to large pages is entirely replicated—this is understandable given its small size (seven entries per thread).
I regard being late as a quality seal really. No point being first, if your info is only as credible as stuff on inquirer. Better be last, but be sure what you write is correct.
Perhaps, if you count being translated from French.
Nice article, good depth, well written
I don't know french, so no idea if it actually works. But I've tried from english to germany and danish, and viseversa. Also tried from danish to german, and the result is always the same - it's incomplete, and anything that is slighty technical in nature won't be translated properly. In short - want it done right, do it yourself.
You claimed the article on toms was a copy paste from another article. He merely stated that the article here was based on a french version.
I actually read the whole thing.
I just don't get TLP when RAM is cheap and the Nehalem/Vista can address 128gigs. Anyway, things have changed a lot since running Win NT with 16megs RAM and constant memory swapping.
1) How's the loop detection feature know when it is a loop ? The diagrams posted don't show any connection between it and the 'front' of the pipeline, so how can it know that the next operation is the same if it hasn't yet entered the loop?
2) On page 8 there's a diagram with a 4 socket setup showing 2 io hubs. Are they connected to the same pcie bus and whatever else they interface with? or are only 2 of the sockets able to directly access a given resource?
3) With the modular design, would one risk buying a cpu that doesn't work in a motherboard because it is intended for a 2 or 4 socket system? or are they all the same, simply with some qpi's disabled?
4) Am I right assuming that qpi replaces fsb when it has to communicate with an i/o hub only? (as shown in one of the top diagrams on page 8) Or is it used for every one of the 'blue' lines on the lower diagram (10 total in a 4 socket layout). The latter would mean 4 qpi's are barely enough to satisfy bandwidth needs in a server enviroment. I imagine an esx server with 4 processors (32 threads) can easily demand memory from dram pools not linked to the local core the threads are running on, and use 96GB/s (3x32) of the 102GB/s (4x12,8x2) total theoretical bandwidth in addition to some of the local 32GB/s bandwidth from the socket a given core/thread is running on. So if this scenario is correct, is it possible to increase the speed of the qpi (read: oc the link) to increase available bandwidth? And what happends if one would successfully find ddr3-1600 modules that would run within the 1,65v limitation? Wouldn't that mean the qpi was already at its limit? (38,4GB/s per dram pool x 3 sockets not local to the core that runs a thread). I know memory isn't truely the bottleneck in modern computers, but I still find it wierd that they put so much effort into the memory controller if it isn't actually the problem. Simply adding a few qpi links between the sockets and the chipset would've solved the bandwidth issue without limiting usable memory types by choosing a certain cpu. Sure it wouldn't have improved latencies, but honestly, who cares? neither in a gaming pc, netbook or any number of common server configurations is it the memory lantecy that is the bottleneck.
5) How much time should one assume is wasted when a core on conroe flushes the l2 cache? they seem to have solved the issue and as consequence increased cache latency (which should turn into slower overall cache performance). In english : can we expect any gain from this change?
6) Would the immensely increased tlb size improve performance in newer games which precache loads of data? (thinking quicker retrieval of texture data etc)
7) Page 12 mentions unalligned memory access, which I've never heard of before. Appearently compilers already try to avoid this situation, so can we expect the improvement to handling such to be of interest? What's the point of improving a feature to handle a situation that hardly ever arises in the first place?
http://www.xbitlabs.com/articles/cpu/display/replay.html
This was a very good article and is not a copy ... well done Fedy !!
perhaps it will burn out the IMC within the chip since its all done at 45nm, 1.6+v would be deadly, imagine air cooling a 3ghz quad core chip at ~2v? i take it it shares the rail even within the cpu so
depends on how connected that ram is, there might be advantages etc this way, and it also makes you wonder if AMD suffers from this - iv heard of extreme overclockers killing ram channels on AMD's etc
on the other hand who cares about high performance memory - 3 x 1333mhz is going to be better then 2 x 1600+mhz channels etc, along with the fact its an IMC based setup etc and average maximum bandwidths of ~32gb/s vs the current average maximum of ~12.8gb/s etc
as for the memory issue. Who'd want to run 3x1333 if they could run 6x1600 ? any enthusiast will only be satisfied with the best, and 1333 just isn't it. Not even 1600 is. ddr3-1333 is basicly obsolete, and it's not even mainstream yet. It's a disaster really.
A case of perhaps minimising reflected impedence?
Just my theory anyway ... remember ... I am only here for the humour ... not the technology.
AMD4LIFE