@Bob: Used to run OHMYTRANCE's shoutcast servers back in the day (streaming vp3/4) over shoutcast, I had access to all their content and just downloaded and enjoyed.. I have a good TB drive sitting around of old videos that needs repairing, but I still remember a lot of the good stuff which I enjoyed listening to from my collection.
I'm getting extremely frustrated with Umbraco.
I have inherited a mess of a site from a long gone developer who did not leave any notes, code comments or even have the site in source control. It has been built in Umbraco 6.2 and is using something called Vega Site Builder (which appears to be hi...
See:
In Solution Explorer, double-click the My Project node for the project.
In the Project Designer, click the References tab.
Click the Add button to open the Add Reference dialog box.
In the Add Reference dialog box, select the tab indicating the type of component you want to reference.
S...
Notes on how Sandra reports cores counts, sockets, etc.
- If the system has multiple sockets, it is reported as "2x <processor>", "4x <processor>", etc. - If there are multiple NUMA nodes in one socket, it is reported as "2N", "4N", etc. (This was the case for certain Opteron processors; see http://ranker.sisoftware.net/top_run.php?q=c2ffc9ef82efd2e5c3b18cbc9af3cefbddb588b89ee6dbebcda8cdf0c0e695a898&l=en) - If the chip is based on Bulldozer or related architectures, the cores are reported as M (modules) and T (threads). Otherwise, the usual C and T are used to indicate core and thread count.
@JonClements why do I get the feeling there's a mountain of Private jokes that mods have somewhere, and that you give each other snide comments related to it? — ItamarG3Aug 1 at 20:41
I would think if you do this or the equivalent in any modern programming language, you'd be able to use all cores/threads/nodes on the system without any special code, and the performance would come down to how much all data crosses between nodes
public class Foo implements Runnable { @ThreadLocal private int x;
@Override public void run() { while(true) { x++; } }
public static void main(String[] args) { for(int i = 0; i < 32; i++) { Thread t = new Thread(new Foo()); t.start(); } }
it's pretty easy to create 32 threads and give them all work to do
but if they incur the latency cost of the fabric "too often" it's going to be slow
@rahuldottech Cache coherence is basically all the processor cores in the system having consistent data in their caches. Because each core has its own cache and runs separately from the others, but share the same overall memory space, the local caches for each core inevitably fall out of sync with each other. If this isn't properly handled, you'll wind up with nasty things like race conditions which can result in data corruption.
There are protocols to allow the data on the other caches to be invalidated or updated efficiently. It's an extremely complicated subject; I had briefly touched upon cache coherence in a graduate-level computer architecture class but the specifics are very murky. You'll want to do some reading of your own.
The Infinity Fabric is what holds them together. AMD's architecture is designed to scale out, rather than scale up.
Applications tend to be designed around a scale-up system architecture. AMD's approach has significant manufacturing cost advantages thanks to massive economies of scale and high yields, but suffers from degraded performance on applications not optimized for this sort of scale-out architecture.
unfortunately if your compiler generates native code that hits worst-case behavior of the infinity fabric / caching logic, you could create torture tests that saturate the fabric and give very skewed results of how fast the chip is actually capable of running, by poorly optimizing your code for the architecture
what this means, in practice, is that anyone using more than a handful of often-busy threads has to take NUMA into account for performance, or a ThreadRipper CPU will behave around the same as a Ryzen 5 at best
this means there's yet another hardware oddity about somewhat-common CPUs these days that software devs have to care about, and old software that's no longer maintained will never be able to accommodate
in fairness to AMD, Intel has had NUMA CPUs in the Xeon range for a long time
I do believe that programs such as the Linux kernel, Windows kernel, nginx, memcached, mysql, postgresql, etc. are at least nominally NUMA aware, with quite a lot of attention being paid to it in the Linux kernel in the mid to late 2000s
Intel: Sit on our laurels for half a decade raking in the dough and laying waste to housing costs in Silicon Valley and Oregon, then act offended when someone dares (gasp!) to compete with us! I thought competition was illegal!
Yup. Intel is getting hit really hard. As more and more apps, especially games, are optimized for >4C processors, AMD's advantages will only become more apparent with time.
Intel still has the advantage of high per-core performance, both in clock frequency and in IPC. However, with Zen being a completely new core architecture, there's probably more room for AMD to grow.
...unlike Intel, who seems to be running out of ways to squeeze more IPC out. Zen is a completely new uarch and it is reasonable to expect that it has more room to grow.
I doubt Intel will be able to get all cores to run that fast. The i9-7900X (10C/20T Skylake-SP LCC @ 3.3-4.3 [4.5] GHz, 140W) has all-core turbo at 4.0 GHz, but absent a 240mm or larger liquid cooling system, will overheat and experience thermal throttling if these clocks are sustained for extended periods of time.
To a lesser extent, this is true of Ryzen, but Intel seems to be having some trouble dissipating heat from the core due to their skimping and using thermal paste instead of soldering the IHS to the die.
If this mistake continues with the HCC parts, then they may have trouble competing with AMD's fastest processors.
Somewhat underwhelming, poor performance per watt. However, any competitor against NVIDIA's high end is welcome. I for one am strongly evaluating using RX Vega for cost reasons (FreeSync is far cheaper than G-SYNC, but generally less flexible with refresh rate ranges).
It'll come down to the benchmarks. Radeon RX Vega will be available next week.
Threadripper and X399 ships in just three days.
TechRadar releases some preliminary benchmark results from the Threadripper 1950X: techradar.com/news/…
200+ W is far more than the 140 W TDP specified for the processor. Furthermore, even though this is under torture test conditions, the other processors are operating somewhere close to TDP. The i9-7900X is not.
> Pushing all of the Core i9-7900X’s cores with Prime95 or LuxRender propels power consumption to incredible heights. You do get 48 percent more rendering performance in LuxRender, but at the expense of 58 percent-higher power use. This approach has the elegance of a sledgehammer. Then again, if you need speed at any cost, Core i9-7900X is top-notch.
Arctic Silver 5 has not aged well, but has the advantage of widespread availability and a long track record of decent performance. Thermal Grizzly Kyronaut is clearly the best of the conventional (not liquid metal) thermal pastes, but I'm having trouble finding any significant quantity of it at a reasonable price (>$10 for just one gram!).
Truth be told, AS5 (and several other products) might have been at a disadvantage during this test due to lack of curing time.
> 1x four-hour burn-in, followed by break of at least two hours 4x one-hour measurement, with one-hour breaks Total time at least 16 hours per thermal product and cooler
> Because of Threadripper's two-die layout and its distributed memory controllers, AMD is offering owners two memory operating modes with differing performance characteristics. The distributed mode provides uniform memory access characteristics to the operating system at a higher average latency. Through testing, AMD discovered that some games are sensitive to memory latency, and for the best performance with those titles, the company is offering a "local mode" in firmware that will transform Threadripper into two non-uniform memory access (NUMA) domains. Each of those local domains will ru…
Also, Windows licensing? Windows 10 Home does not support more than one socket, and I'm wondering if enabling NUMA mode would cause licensing issues, especially if Windows winds up seeing the processor as two sockets.
Then again, whoever is building a system with a 16C processor would probably have no problem shelling out the extra $80 for Windows 10 Pro.
hey guys, a couple of years ago there was a firewall software that popped up a notification everytime an application made an outbound connection, allowing you to confirm or block that connection. does something like that still exist?
@allquixotic it's easy to create threads; it's very difficult to give them work to do
if the unit of work that can be completed without having to access shared data is small, you lose performance against a serial implementation
in other words, there's few cases where you can actually wave a magic multithread stick
@bwDraco I'm curious how it performs with limited cooling; presumably it would not be operating in turbo mode and the lower clocks could make all the difference
@bwDraco I know what will happen. I'm talking about relative performance.
The 140W TDP isn't necessarily a lie, but it does make benchmarks harder to compare because you need a benchmark performed with limited cooling (or turbo disabled)
Not too sure, but Skylake-X already has a reputation for being very power-hungry, a problem compounded by Intel's use of thermal paste rather than solder.
Ryzen thus far has been shown to have excellent performance per watt.
@bwDraco ...cooling performance doesn't make it more or less power hungry afaik
paste cools slower (does not imply) higher power consumption
it does mean it might hit internal temp limits and throttle earlier
but moving heat from core to package slower doesn't affect the total heat generated (= power consumption), except perhaps reducing it if it hits a limit
I'm not exactly sure what will happen if the processors were using coolers with low cooling capacity, but chances are good that the Ryzen processors will perform better in a thermally-constrained situation.
Among other things, Ryzen processors have the IHS soldered to the dies.
@bwDraco Hm. A bit more thought, and I was going about that the wrong way. Slower heat transfer means the CPU starts throttling before the heatsink hits that temp. So the question is what the delta is when throttling starts - if it's small, all good, if it's big, then we have a problem.
@Burgi I think the Blade Runner universe is supposed to sound sleazy :P
Or so my english (literature) teacher would've said
See:
In Solution Explorer, double-click the My Project node for the project.
In the Project Designer, click the References tab.
Click the Add button to open the Add Reference dialog box.
In the Add Reference dialog box, select the tab indicating the type of component you want to reference.
S...
