Win a 48-core Magny Cours system

[cbn]

Low clock speeds? Let's just say that at clock to clock comparisons, MC is a lot faster than Istanbul. And our 8-core and 12-core products will have higher clocks than beckton.

How much does switching to DDR3, four memory channels, and increasing hyper-transport link have to do with the improved Clock per clock performance of MC over Istanbul?

Does Lisbon have the same L1/L2/L3 cache as Istanbul? I tried looking this information up but couldn't find it on my search.

P.S. How many volts does MC (model 6174) need compared to a higher clocked Istanbul?

 

[cbn]

This of course assumes that Intel pushes the limits to get the highest clock speeds and does not drive to lower clocks and lower power consumption. History has proven that they take this path 100% of the time.

What sort of challenges (economic, process technoplogy, green laws) could Intel face in future by always taking this pathway?

Is there anything predictable that could make them want to change? If Intel decided switch strategies to "performance per watt" rather than "performance per mm2 silicon" how would that impact AMD?

Of course, I don't know the answers to these questions, but just looking at why some companies are switching to water cooling really adds a twist to the developments. I never would have thought getting power to a building's air conditioning system could be the limiting factor.

 

[JFAMD]

I would sell it.
honestly, I don't think my house's wiring can provide enough power to run it.

Actually, it probably could. I'd give the winner a choice of the standard power or low power processors (12-core as well) if they wanted.

 

[JFAMD]

How much does switching to DDR3, four memory channels, and increasing hyper-transport link have to do with the improved Clock per clock performance of MC over Istanbul?

Does Lisbon have the same L1/L2/L3 cache as Istanbul? I tried looking this information up but couldn't find it on my search.

P.S. How many volts does MC (model 6174) need compared to a higher clocked Istanbul?

Some of the performance is processor, some is platform. Not at liberty to say how much is attributed to each.

Same cache structure, but might be some tweaks.

Not sure on the voltage. Wattage is about the same.

 

[JFAMD]

What sort of challenges (economic, process technoplogy, green laws) could Intel face in future by always taking this pathway?

Is there anything predictable that could make them want to change? If Intel decided switch strategies to "performance per watt" rather than "performance per mm2 silicon" how would that impact AMD?

Of course, I don't know the answers to these questions, but just looking at why some companies are switching to water cooling really adds a twist to the developments. I never would have thought getting power to a building's air conditioning system could be the limiting factor.

I doubt that there is any issue that they would face. Our design goals revolve around performance per watt vs. raw performance.

Here's an example. Barcelona was 65nm, Shanghai was 45nm. Power dropped between them. Top speed Shanghai was 2.7 at launch. We could have pushed it to 2.8 or 2.9 if we wanted to match the power of Barcelona. Instead we took the power drop over the performance gain. That is not always 100% the case though, we matched the Shanghai and Istanbul power (didn't take it down) but gave everyone 2 more cores. But generally we look for opportunities to give better performance per watt (which has been our focus.)

Intel has basically pushed for more performance at every turn, they have generally not focused on pulling power down. Look at max power on their processors and you'll see that since everyone got "power religion" back in the early 2000's, their max power keeps creeping up.

Just different philosophies on how to achieve the same goal. To get better performance per watt you have 2 levers: performance and power. As an example, from idle to ~3% utilization you see them jump ~10% where we go up by ~1%. A low idle is interesting, but if you ramp power quickly to get to low utilization, you are wasting power. We'll have a new platform with some interesting power features. Remember that all comparisons today are their 2009 platform to our 2006 platform.

 

[TuxDave]

Heh, this sounds like a contest we had a while ago with "How would you show off a Nehalem system?" where then you get funded to go build that project and actually demo it. There was some pretty good stuff that I had no chance in hell of competing against. I assume it'll be pretty messed up if I participated huh.

 

[Accord99]

I doubt that there is any issue that they would face. Our design goals revolve around performance per watt vs. raw performance.

Here's an example. Barcelona was 65nm, Shanghai was 45nm. Power dropped between them. Top speed Shanghai was 2.7 at launch. We could have pushed it to 2.8 or 2.9 if we wanted to match the power of Barcelona. Instead we took the power drop over the performance gain. That is not always 100% the case though, we matched the Shanghai and Istanbul power (didn't take it down) but gave everyone 2 more cores. But generally we look for opportunities to give better performance per watt (which has been our focus.)

In other words, all you did was stop selling two or three tiers of higher-clocked products that you would have sold previously and that you still sell for the desktop market. On the other hand, Intel merely kept selling new products at the same rough power tiers as before, though with obviously massive increases in performance/watt.

Intel has basically pushed for more performance at every turn, they have generally not focused on pulling power down. Look at max power on their processors and you'll see that since everyone got "power religion" back in the early 2000's, their max power keeps creeping up.

So does AMD, with the Phenom II 965 which is a very high power consuming CPU. Both companies push their designs to achieve the maximum performance while remaining within the absolute highest acceptable limit for x86 products. After which, they scale down with frequency and voltage to fit other power thresholds.

But personally, I think AMD clearly recognized the might of Nehalem in multi-processing throughput, and came to the logical conclusion that the only chance to match it in throughput is to back down the voltage-frequency curve and throw more cores at it.

 

[IntelUser2000]

How much does switching to DDR3, four memory channels, and increasing hyper-transport link have to do with the improved Clock per clock performance of MC over Istanbul?

Does Lisbon have the same L1/L2/L3 cache as Istanbul? I tried looking this information up but couldn't find it on my search.

P.S. How many volts does MC (model 6174) need compared to a higher clocked Istanbul?

Remember JFAMD is a server guy.

In servers the 2x cores will benefit lots of applilcations greatly, aside from other changes like memory and Hypertransport.

 

[cbn]

Here's an example. Barcelona was 65nm, Shanghai was 45nm. Power dropped between them. Top speed Shanghai was 2.7 at launch. We could have pushed it to 2.8 or 2.9 if we wanted to match the power of Barcelona. Instead we took the power drop over the performance gain.

It sounds like getting another 100-200 Mhz resulted in a poor trade-off in terms of energy used.

Just different philosophies on how to achieve the same goal. To get better performance per watt you have 2 levers: performance and power. As an example, from idle to ~3% utilization you see them jump ~10% where we go up by ~1%. A low idle is interesting, but if you ramp power quickly to get to low utilization, you are wasting power. We'll have a new platform with some interesting power features. Remember that all comparisons today are their 2009 platform to our 2006 platform.

How much time do these servers spend at idle? (I have seen some water cooler reviews where the idle temps are actually higher for Water coolers, but in all fairness they were very small water coolers)

But personally, I think AMD clearly recognized the might of Nehalem in multi-processing throughput, and came to the logical conclusion that the only chance to match it in throughput is to back down the voltage-frequency curve and throw more cores at it.

I would imagine the trade-off is greater manufacturing costs, but AMD gets more performance from less watts while Intel gets more performance from less silicon.

I guess the question in my mind is which approach will be more flexible for rapidly developing markets? Are we going to see an explosion in cloud computing in the near future (10-15 years out) where whoever has the best power efficiency vs upfront costs makes the most sales? How important will server density be if this happens? How will Fusion and heterogenous computing play a role in this type of scenario?

As it stands now it seems Intel has the more profitable CPUs, but I suspect AMD/ATI will be more profitable than Intel on the GPU side of things of APU/System on a chip.

The profitability obviously will affect the how low each company can set their retail prices.

Remember JFAMD is a server guy.

In servers the 2x cores will benefit lots of applilcations greatly, aside from other changes like memory and Hypertransport.

I didn't realize Magny Cours had the same L1/L2/L3 cache as the 45nm hexcore server chips. This is why I keep thinking being able to bump clocks up to say 2.5-2.6 (an arbitrary number) might still be in the efficiency range of other "uncore" components of the chip.

That being said would it be possible for some form of watercooling to allow higher clocks on less watts? Or would some form of water cooling be more effective for the GPU side of the server APUs? (I am obviously thinking about the future, not the present chip design)

Is there is anything about GPU architecture that could make them more receptive to watercooling efficiency gains than CPUs? If so, any water cooling benefit for the CPU side of the APU would be icing on the cake.

In my mind I am imagining AMD/ATI having higher performance per mm2 silicon for the GPU, but possibly more performance per watt on the CPU side of the Fusion/APU equation.

 

[JFAMD]

How much they spend at idle is really a function of the workload.

Most servers are ~20% utilization and as we continue to get more cores, you get more idle time because you are running multiple threads concurrently vs. in a singular serial fashion. These will be at idle often.

Those are your standard network loads, the loads that are all being virtualized.

Virtual servers are generally running 40-60% utilization and they will generally be at idle for very short periods of time.

HPC servers are generally ~80-90% utilization, so, generally little or no idle.

Right now ~50% of the server workloads are utilized, the workloads that are at higher utilization generally don't end up being virtualized because there is little gain. Unless you are doing it for disaster recovery, load balancing or for agility.

 

[cbn]

As far as APU server chips go, when do you think we will see those being made? What advantage would Fusion have over a competitor using a combination of Intel CPUs and Nvidia Accelerators?

What type of server would be able to make the best use of a Fusion APU?

 

[JFAMD]

You probably won't see APU in servers for a few years still. For client PCs there is a market for APU wih a CPU and lower/middle range GPU.

In the server world, for GPGPU loads you are looking for the highest speed GPUs that you can get. Those are harder to integrate into a CPU. Most of the customers that I talk to using GPGPU are looking at a 2:1 GPU to CPU ratio. So for them, there is less benefit immediately from an APU. It will happen, but look for it first on the client side.

 

[jvroig]

Most of the customers that I talk to using GPGPU are looking at a 2:1 GPU to CPU ratio

I can understand that. That is also mostly the story with the servers of most of my clients. Either they need CPU power (most cases) or they need lots of GPGPU-enabled video cards, and Llano/Westmere types won't cut it, not by a longshot.

In that regard, when is AMD going to target that portion of the server market? I have no doubt your newly rolled-out server procs (Magny Cours) are in fact a marvel, given that it has the same TDP as Istanbul and yet has 2x the cores. Green data centers will love this. But when will the other half of AMD (the graphics division) roll-out something for the servers (HPC) as well? Since this is still about servers, are you also The Man when it comes to AMD's HPC efforts using GPGPU, or does ATi have a counterpart of you for it? I don't see how ATi is taking HPC seriously, I don't even remember any mention of it in roadmaps.

 

[JFAMD]

There is someone on the ATI side that is in charge of the FireStream products. We are already delivering products to market. Even the 5th fastest supercomputer in the world is using ATI technology. It's not our CPUs, but it is a testament to our open strategy.

 

[cbn]

When I did a search last night for "AMD Fusion APU servers" I got some hits for AMD 3D Cloud rendering.

I don't know how it works (other than it is on a cloud). However, anytime piracy can be controlled (through subscription services) I would imagine more programmers would be interested.

I would be great to see some educational/puzzles/games/simulations being created that are beyond the realm of a console or regular PC.

 

[cbn]

You probably won't see APU in servers for a few years still. For client PCs there is a market for APU wih a CPU and lower/middle range GPU.

In the server world, for GPGPU loads you are looking for the highest speed GPUs that you can get. Those are harder to integrate into a CPU. Most of the customers that I talk to using GPGPU are looking at a 2:1 GPU to CPU ratio. So for them, there is less benefit immediately from an APU. It will happen, but look for it first on the client side.

When you say 2:1 GPU to CPU ratio, what exactly do you mean by that?

Does the ratio refer to calculation power, power usage, silicon area or something else?

 

[JFAMD]

I mean that with a typical 2P system customers would generally want 4 GPU cards.