AMD - Bulldozer Chitchat

[badass]

Actually not entirely true. They used a simulated E5-2630 with 33% more L3 cache - both me and scaryjim pointed this out to you. Being objective,I cannot say whether the 100MHZ clockspeed decrease they implemented will compensate for the increase in L3 cache.

Fair point. However my money's on my statement about price/performance still being valid when the actual E5-2620 and E5-2630 are tested.

[JF-AMD]

Well DDR4 probably isn't far off, mass production in H2 2012 IIRC.

I'll take the over on that bet

Which is exactly my argument.
Same performance, Xeon is cheaper
Same price, Xeon is faster.

This was demonstrated on my previous post. The Xeon E5-2620 is faster than the top bin Opteron and cheaper than the mid range one.


The comparisons are linked by me earlier in this thread. They are on Anandtech.

Nope, sorry. Anandtech results don't cut it in my book. When you are using cinebench and winzip compression to test server processors, it's not a fair fight.

SPEC INT and 1KU price

A 2620 is priced at $406, our 12-core 6238 is priced at $455, so those are pretty close. SPEC INT on a 6238 is 414. Let me know when you get a score on the 2620.

[watercooled]

I'll take the over on that bet

Hmm.

Also important to note that lots of synthetic benchmarks like Cinebench are compiled with the Intel compiler so optimised better for their architecture.

[DanceswithUnix]

Haswell is rumoured to have on-die stacked RAM for the GPU and it seems that AMD might be also working on something similar too.

SA,leaked this picture last year:

http://semiaccurate.com/2011/10/27/a...-gpu-pictured/



Some more information:

http://semiaccurate.com/forums/showp...8&postcount=26

At the back of my mind when I wrote my sideport/on package dram comment was this:

http://semiaccurate.com/2012/03/02/s...th-an-amd-gpu/

which implies the PS4 could have an AMD cpu with dram and an FPGA stacked in 1 package.

[watercooled]

I think an FPGA would be a great addition to current platforms, consoles/PCs alike. Yeah I'm biased because I'm a hardware enthusiast but still...

[badass]

Nope, sorry. Anandtech results don't cut it in my book. When you are using cinebench and winzip compression to test server processors, it's not a fair fight.

SPEC INT and 1KU price

A 2620 is priced at $406, our 12-core 6238 is priced at $455, so those are pretty close. SPEC INT on a 6238 is 414. Let me know when you get a score on the 2620.

You've completely ignored the fact the linked pages were virtualisation scores, something the vast majority of these processors will actually be doing.

Please can you explain how SPEC numbers for CPU's are more relevant than real world virtualisation scores.
I find this particularly odd as you've just used the "not real world" argument to attempt to rebuff my figures.
Secondly, I don't care about the 1KU price. I care about the price of the servers I can buy with said processors in.
You brought up the 2 Dell systems to compare, now you've moved the goalposts.

You're treating this like anyone trying to sell the inferior product does. They treat it like it's top trumps, it's their turn and they can see the other players card. They have a selection of metrics to choose from and choose only the ones where their card wins.

Instead of fighting the losing battle trying to suggest any Bulldozer processor is better than the similarly priced intel equivalent, how about picking a battle you can win, such as "Piledriver will be better in every metric than bulldozer"

EDIT: I've just realised you've chosen SPEC INT. What an absolutely meaningless benchmark to choose. About as meaningless as SPEC FP TBH. I bet however that the Xeon CPU gets a much higher score. That further reinforces my top trumps argument.

[shaithis]

A potential return to CPU slots or even those "cache slot" upgrades....

[JF-AMD]

You've completely ignored the fact the linked pages were virtualisation scores, something the vast majority of these processors will actually be doing.

Please can you explain how SPEC numbers for CPU's are more relevant than real world virtualisation scores.
I find this particularly odd as you've just used the "not real world" argument to attempt to rebuff my figures.
Secondly, I don't care about the 1KU price. I care about the price of the servers I can buy with said processors in.
You brought up the 2 Dell systems to compare, now you've moved the goalposts.

You're treating this like anyone trying to sell the inferior product does. They treat it like it's top trumps, it's their turn and they can see the other players card. They have a selection of metrics to choose from and choose only the ones where their card wins.

Instead of fighting the losing battle trying to suggest any Bulldozer processor is better than the similarly priced intel equivalent, how about picking a battle you can win, such as "Piledriver will be better in every metric than bulldozer"

EDIT: I've just realised you've chosen SPEC INT. What an absolutely meaningless benchmark to choose. About as meaningless as SPEC FP TBH. I bet however that the Xeon CPU gets a much higher score. That further reinforces my top trumps argument.

So, the virtualization benchmark is what you want to talk about.

The 2690 is ~40% faster than a 6276.

But, when you compare them, you are essentially saying a $2000+ processor is faster than a $788 processor. Or, in simple terms would you pay 150% more for a CPU that gives you only 40% more performance. The economics just don't work.

You keep talking about how much faster the 2690 is and then how cheap the 2620 is. Let's stop splitting the discussion. Pick your part, 2690 or 2620 and stick with it.

As for SPEC scores, those actually happen to be the most relevant for comparing CPUs. Integer math is 90% of what a CPU does. Floating point is the other 10%. SPEC Int tends to be less influenced by compiler, configuration and component issues outside of the CPU. It's not totally abstracted, some of those things will influence, but not to the same degree.

Also, keep in mind that Anandtech's virtualization benchmark is not really fair as it relies on an OS that is not optimized on the AMD processor. So any results there are automatically skewed anyway. The reason that SPEC works well is that every vendor submits their score. I am pretty sure that if AMDzone.com did a benchmark where Interlagos beat Sandybridge you'd call foul. When each vendor submits their own scores you have the best actual representation of performance.

[badass]

So, the virtualization benchmark is what you want to talk about.

The 2690 is ~40% faster than a 6276.

But, when you compare them, you are essentially saying a $2000+ processor is faster than a $788 processor. Or, in simple terms would you pay 150% more for a CPU that gives you only 40% more performance. The economics just don't work.

You've all ready tried that straw man argument and it was debunked. Stop bringing it up or properly counter it.

You keep talking about how much faster the 2690 is and then how cheap the 2620 is. Let's stop splitting the discussion. Pick your part, 2690 or 2620 and stick with it.

I have not mentioned the 2690 at all. Just the 2620 and 2630. So I have chosen which one.
The 2620 and 2630 are both faster and the servers you can buy with them are cheaper.
Is that clear enough?

As for SPEC scores, those actually happen to be the most relevant for comparing CPUs. Integer math is 90% of what a CPU does. Floating point is the other 10%. SPEC Int tends to be less influenced by compiler, configuration and component issues outside of the CPU. It's not totally abstracted, some of those things will influence, but not to the same degree.

Why do I care about any of that? As you have all ready stated, server users care about cost at what level of performance they need or the best performnace at a certain cost. So why bring a synthetic benchmark into it?

Also, keep in mind that Anandtech's virtualization benchmark is not really fair as it relies on an OS that is not optimized on the AMD processor. So any results there are automatically skewed anyway. The reason that SPEC works well is that every vendor submits their score. I am pretty sure that if AMDzone.com did a benchmark where Interlagos beat Sandybridge you'd call foul. When each vendor submits their own scores you have the best actual representation of performance.

I presume you mean VMware - as in the most popular Virtualisation solution for businesses in the world.
As a user, why would I care about the optimisations - I just care about performance and cost.

However I accept that if ESXi is not optimised for AMD CPU's and is for Intel CPU's that does not reflect on the underlying architecture.
I am a little surprised if this is the case as I thought Vmware are vendor neutral.

[JF-AMD]

You talked about performance, you linked to the Anandtech article. They are using a 2690. You talk about the price of the 2620, but can't share any performance on that.

If you had both price and performance we could discuss this, but without the performance of the 2620, your straw man of "it's faster" just doesn't hold up.

I don't care to continue this.

[badass]

You talked about performance, you linked to the Anandtech article. They are using a 2690. You talk about the price of the 2620, but can't share any performance on that.

If you had both price and performance we could discuss this, but without the performance of the 2620, your straw man of "it's faster" just doesn't hold up.

I don't care to continue this.

OK. We can wait until the 2620/2630 performance becomes public. The only part of the equation missing is what effect the losing of 33% of the cache will have. In the extremely unlikely event that in real world tests the losing of the 33% of the cache from the simulated 2620 to a real 2620 results in the 6276 overtaking it, I'll concede that the 6276 based server represents better value.
I think it'll be a small miracle that losing that much cache results in that much of a loss of performance in real world applications.

[DanceswithUnix]

You've all ready tried that straw man argument and it was debunked. Stop bringing it up or properly counter it.

I have not mentioned the 2690 at all. Just the 2620 and 2630. So I have chosen which one.
The 2620 and 2630 are both faster and the servers you can buy with them are cheaper.
Is that clear enough?


Why do I care about any of that? As you have all ready stated, server users care about cost at what level of performance they need or the best performnace at a certain cost. So why bring a synthetic benchmark into it?

As "synthetic" benchmarks go, SPEC is a pretty good one. Some of us care about things like GCC and perl performance that go into that score.

I presume you mean VMware - as in the most popular Virtualisation solution for businesses in the world.
As a user, why would I care about the optimisations - I just care about performance and cost.

However I accept that if ESXi is not optimised for AMD CPU's and is for Intel CPU's that does not reflect on the underlying architecture.
I am a little surprised if this is the case as I thought Vmware are vendor neutral.

Not sure what to think of VMware these days, though Neutral isn't something they are usually accused of

[CAT-THE-FIFTH]

More details about Trinity:

http://www.tomshardware.com/news/AMD...emo,15009.html

The 35W mobile version was running DiRT3 at high settings at 1366X768 with 4XAA smoothly and the desktop version was running the same game at lower settings without AA at 5040X1050!

[Noxvayl]

I am upgrading the family PC with a trinity APU for sure, my dad has been itching to play some simulator games like Dirt3 and Microsoft Flight so this seems to be perfect for that. I hope the MOBO/APU/RAM is as affordable as the current A series systems.

[CAT-THE-FIFTH]

Wrong thread.

[CAT-THE-FIFTH]

Trinity A10 performance in Super Pi and 3DMark06:

http://semiaccurate.com/forums/showp...postcount=1302

Here are the scores for the Llano A8-3850:

http://www.computerbase.de/bildstrecke/34996/5/

http://www.computerbase.de/bildstrecke/34996/6/

Bulldozer and Piledriver lack an X87 unit,so the top end desktop Trinity is still outperforming the A8 3850 by around 19% or thereabouts. The A10-5800K has a top Turbo Core speed of around 4.2GHZ,so at least in Super Pi a 3.5GHZ Piledriver should be around the same as a 2.9GHZ Llano A8 which has an X87 unit.

However,the CPU-Z picture shows the CPU running at 2GHZ so it could be not running at its full Turbo Core speed.

The SM3.0 score in 3DMark06 is around 29% higher for the Trinity A10 too.