AMD Bulldozer server info

[accord99]

No it doesn't. It uses the extra thread to reduce the amount of time the core spends idle whenever there's a stalled thread.

No, the execution pipeline can contain instructions from both threads.

[Apokalipse]

No, the execution pipeline can contain instructions from both threads.

The scheduler can. Not the execution pipeline.
http://en.wikipedia.org/wiki/Hyper-threading

[kl0012]

No it doesn't. It uses the extra thread to reduce the amount of time the core spends idle whenever there's a stalled thread.
The Operating system sees two cores. That's just to allow the extra thread to be issued. It can't process them simultaneously.

There are different kinds of MT. HT is SMT which means that instructions from both threads can be executed simultaneously (depending on available execution resources). Nehalem and later CPUs have 5 execution ports which in some degree reduces "threads collisions". Of cause it can be eliminated completely.

Elaborate.....

Here is a simple calculation: Penryn core: 22 nm^2, Nehalem core: 24 nm^2. Considering that there are additional improvements in Nehalem cores (relative to Penryn - such as larger OoO buffers) and cashes, IMC e.t.c, it is safe to say that HT uses less then 5% of total core size.

[accord99]

The scheduler can. Not the pipeline.

Of course it can, that's the point of Hyperthreading/SMT. You're confusing it with something like Switch-On-Event Multi-threading.

[Apokalipse]

Of course it can, that's the point of Hyperthreading/SMT. You're confusing it with something like Switch-On-Event Multi-threading.

No, the point of HyperThreading is to reduce the idle time caused by stalled threads.
Instead of waiting for the data it needs to continue the first thread, it starts processing the second thread instead.

[kl0012]

No, the point of HyperThreading is to reduce the idle time caused by stalled threads.
Instead of waiting for the data it needs to continue the first thread, it starts processing the second thread instead.

Wrong again.
http://download.intel.com/technology...iss1_art01.pdf

[Apokalipse]

That paper specifically mentions utilization of spare resources from stalled threads. As does the wikipedia link I provided.

[kl0012]

That paper specifically mentions utilization of spare resources from stalled threads. As does the wikipedia link I provided.

Stalled threads were mentioned as one of the reasons of worst application scaling, but not the only. The main reason of HT is to enable "true" TLP on single core for better utilization of execution resources with minimum die size increase.
As the whole SMT concept:
http://en.wikipedia.org/wiki/Simulta...multithreading

[Fatfool]

Be careful. You can't really overclock with these boards, and your CPUs aren't going to have that high of clock speeds. I had a dual Istanbul setup on a dual-NB SR5690 board a while back. It did pretty well at threaded work, but I eventually replaced it a year and a half later when Thuban came out. An overclocked Thuban was just about as fast in threaded work as my 12-core machine was and ungodly faster in single threaded stuff, which is important for games.

Last I checked, games didn't depend that much on the processor anyway. (all the scaling on cores controversy)

I've got a use for the many extra cores though. I encode videos and well, that scales real well with more cores.

Socket F was never cost effective anyway. Compare the price of Istanbul against Lisbon and the clocks of Istanbul vs Lisbon. Plus lisbon could actually upgrade to Valencia (bulldozer).
At the time of launch, Lisbon's clocks really weren't far from the desktop Thuban variant (2.8ghz on the 4184 vs 3.2ghz on the 1090T). Talking bout non overclocked frequencies that is. Thuban could 'turbo' though. Doesn't actually boost all cores (which are used during multithreaded number crunching) to 3.6ghz, but it's good for single threaded proggies. Let's hope Valencia actually sticks to the plan.

Overclocking wise, I'm a little hesitant to do that even though video editing is a hobby for me. Any crash while the encode is going on is rather maddening. (waste of time repeating the encode, disruption of workflow etc)

[informal]

Back to topic:

AMD's Bulldozer 'Orochi' wafer pictured
Yet another leak
by Charlie Demerjian

January 17, 2011

Opteron LogoWHEN IT RAINS, it seems to pour, only this is January in Minnesota, so it is snowing and we are talking about leaks. This time, here is a much closer shot of Orochi/Bulldozer than has been released elsewhere.

[jmm5351]

I just wanted to say that I am sick of people arguing about how it is impossible for Bulldozer to be dual socket configurable. I am sure if AMD wanted to do this then they surely can. AMD just invented a whole new type of processor that is basically unheard of. Bulldozer pushes the very foundation as to what makes a processor a processor. If AMD can design, and manufacture such an intricate piece of hardware, well common sense tells me if they actually wanted to then AMD could make Zambezi dual socket compatible. No one can argue this either, unless you actually work for AMD and you helped design Bulldozer. BUTTTTT, c'mon people, if AMD can make such a crazy processor I think they could design a socket that would allow Zambezi be configured as a mult-CPU platform. Making the processor is the hard part, making the socket is the easy part.

[Dimitriman]

I just wanted to say that I am sick of people arguing about how it is impossible for Bulldozer to be dual socket configurable. I am sure if AMD wanted to do this then they surely can. AMD just invented a whole new type of processor that is basically unheard of. Bulldozer pushes the very foundation as to what makes a processor a processor. If AMD can design, and manufacture such an intricate piece of hardware, well common sense tells me if they actually wanted to then AMD could make Zambezi dual socket compatible. No one can argue this either, unless you actually work for AMD and you helped design Bulldozer. BUTTTTT, c'mon people, if AMD can make such a crazy processor I think they could design a socket that would allow Zambezi be configured as a mult-CPU platform. Making the processor is the hard part, making the socket is the easy part.

i dont think amd ever said they cannot make dual socket zambesi. they have been saying all this time that dual socket desktoo segment is nearly dead and the more cores are released on single socket the more dead that segment is. it makes sense as you simply not find any games anywhere that will benefit from 16 cores. 8 cores is plenty for games and physics and its more important to have them clocked high than to stack up so many.
on the other hand if amd had a zambesi with quad channel ram on single socket g34 boards (aka athlon fx 51) that would be an interesting setup for hardcore gamers, and if they wanted more cores they could simply buy opterons. (amd could also limit the g34 zambesi to 1 socket so not to cannibilize the c32/g34 server segment.

thats just my way of seeing a super premium amd product.

[cky2k6]

And a Bulldozer module being 12% larger than one core (or what one core would be if they didn't use modules) and actually have the ability to process another thread (instead of just using idle resources of the one core)

You are comparing 8 virtual threads with 8 actual threads.
Do you understand how hyperthreading works? It doesn't make each core process two instructions simultaneously.

And that's how Bulldozer modules are superior. An even smaller increase in die area (HT needs +30%, BD module needs +12%) for the ability to actually process two threads simultaneously.

HT certainly improves performance to the extent that 4 core 8 threaded 920s are comparable to 6 physical cores on thuban. This is just based on a rumor, but it said 8 core bulldozers compete with a westmere, which is fine and dandy, except that thats not what they will be up against. Bulldozer I'm sure will do fine against 1155 sandybridges even if they don't oc quite as well, but people currently on 1366 are probably not looking for a bargain considering x58 is not the best bang for the buck compared to 1156 or amd's am3 chips. Its not a small market like 2011 will be, and I don't think AMD should just dismiss it because of the "professionalism" aspect of letting g34 boards being equipped with enthusiast features. A $300-400 12 core cpu will put a real wrench in 1356's sales, as that was the i7 920's launch price, and I don' think intel lost money on those chips

[jeremyshaw]

Remember, you are talking to AMD's server PM, he will have a professionally (not to say enthusiasts are not in a sense) oriented view of what the G34 platform is oriented towards. That being said I don't see G34 being taken to overclocking. I'm just hoping AMD slips in a Bulldozer G34 SKU that has an unlocked multiplier (but BIOS limitatons on board support?).

[cky2k6]

Yes, all thats really needed is a "black edition" g34 part, and the mobo makers can take care of the "rest" like putting on consumer trinkets and lollerskate heatsinks. The two markets will never come into contact, and intel was perfectly fine with using 1366 for enthusiast chips as well as xeons. Hell, amd does it as well with opterons using the same socket as a consumer sku just with a server chipset. You have the 12 and 16 core cpus, just connect the dots amd.

[JF-AMD]

Yes, all thats really needed is a "black edition" g34 part, and the mobo makers can take care of the "rest" like putting on consumer trinkets and lollerskate heatsinks. The two markets will never come into contact, and intel was perfectly fine with using 1366 for enthusiast chips as well as xeons. Hell, amd does it as well with opterons using the same socket as a consumer sku just with a server chipset. You have the 12 and 16 core cpus, just connect the dots amd.

It's about $5M to bring any project to market.

The revenue opportunity is significantly under $5M, so you end up with a loss.

[Solus Corvus]

Eventually a 16 core on AM3+ would be nice.

[mAJORD]

Back to topic:

didn't seem to help the OT!, but interesting. Shows symetry between the 4 modules, look fwd to higher res shots.

Wonder what Hans will make of this shot though

[jeremyshaw]

It's about $5M to bring any project to market.

The revenue opportunity is significantly under $5M, so you end up with a loss.

How much would it cost to bring in an unlocked multi part? Intel doesn't seem adverse to unlocked multis on Xeon parts (W3570 being an example). For that matter, in in the Phenom II range (consumer, not server), there seems to be an evergrowing collection of "Black Edition" parts with unlocked multis, to the point of 3 Denebs (if you don't want to count the harvested X3 and X2 pieces) with unlocked multis, compared to 3 without ("e" editions never lasted long on the open market, iirc).

[Apokalipse]

HT certainly improves performance to the extent that 4 core 8 threaded 920s are comparable to 6 physical cores on thuban.

Thuban cores don't have as high IPC as Nehalem cores, so HT isn't exactly the sole reason for that.

This is just based on a rumor, but it said 8 core bulldozers compete with a westmere

Fudzilla.
I think I'd like a more reliable source.

[Fatfool]

I just wanted to say that I am sick of people arguing about how it is impossible for Bulldozer to be dual socket configurable. I am sure if AMD wanted to do this then they surely can. AMD just invented a whole new type of processor that is basically unheard of. Bulldozer pushes the very foundation as to what makes a processor a processor. If AMD can design, and manufacture such an intricate piece of hardware, well common sense tells me if they actually wanted to then AMD could make Zambezi dual socket compatible. No one can argue this either, unless you actually work for AMD and you helped design Bulldozer. BUTTTTT, c'mon people, if AMD can make such a crazy processor I think they could design a socket that would allow Zambezi be configured as a mult-CPU platform. Making the processor is the hard part, making the socket is the easy part.

Read the previous page, Dual sockets for bulldozer already exists in the form of socket C32(with attractive prices) You can buy the Asus board now, the MSI board should be shipping next month.

[cky2k6]

It's about $5M to bring any project to market.

The revenue opportunity is significantly under $5M, so you end up with a loss.

The revenue for mid-high end cpus cannot be less than $5 mil, x58 is a damn popular platform, as will be x68 and that is something you can adress with a simple unlocked multi. I want to see the evidence where even 20% of the x58 market is under 5 million.

[JumpingJack]

The revenue for mid-high end cpus cannot be less than $5 mil, x58 is a damn popular platform, as will be x68 and that is something you can adress with a simple unlocked multi. I want to see the evidence where even 20% of the x58 market is under 5 million.

In 2009 the Enthusiast class high end hardware was some $9.2 billion (that is with a B not an M)... http://jonpeddie.com/press-releases/...ulates-worldw/

This includes everything, from graphics cards to monitors ... let's say 1/10th of that is on the CPU .... this is 920 million just on CPUs, let's say AMD holds or takes just 1/3 of that market share, thats around 306 million in revenue for AMD annually. Mind you this is not DIYers, low end buyers who over clock, this is the highest end hardware. Let's say a hypothetical dual socket C34 enthusiast class desktop product fetches just 5% of that highest end stuff... that's roughly $15 million.

A lot of hand-waving, but there is money there to be had.

But what is dangerous is that now you need to support that platform for several life cycles, and dual socket desktop systems are, in actuality, not very popular so that hand-waving 5% assumption could be way way off. Sure, lots of people posting here have some 2P server related boards as desktops or you could look at the ill-fated QuadFX. So the actual share of market such a system might could certainly fall below the cost to develop and market the product.

But then again, if BD really delivers as the rumor mill is suggesting, people will buy it, if anything, to just brag.

[Fatfool]

It's about $5M to bring any project to market.

The revenue opportunity is significantly under $5M, so you end up with a loss.

What about keeping the opteron SKUs as it is and using the same chipsets. In fact, why not buy a few dozen of these boards from Asus and MSI (those ATX dual lisbon ones), then send them out for reviews with a pair of Opteron 4184s, emphasising their price against performance and their ability to upgrade to bulldozer. That won't cost 5mil. The only resistance would be 'we can't make these server parts seem like enthusiast parts!'

Then judging from the response, there could be a version tailored for enthusiasts shortly after bulldozer's introduction. Basically, move this segment into the 'high performance enthusiast' one where intel's models are currently sitting.

[jeremyshaw]

I guess it's more complicated than that, but Intel can do stuff like this since it's only adding all the more to the "halo" effect. You have the best, people are going to buy from the best, with lessor reguard as to how well the lower end parts actually perform. So even if Intel doesn't make a single cent in direct profit from their i7 980 (highly doubtable ), the fact that they have this awesome CPU associated with their name can drive people to buy something simply because it says "Intel" on it...