AMD's "Istanbul" 6-core Opteron reviews

[fellix]

AnandTech (http://it.anandtech.com/IT/showdoc.aspx?i=3571)
TechReport (http://www.techreport.com/articles.x/17005/1)

http://img5.imageshack.us/img5/2208/6vs4core.png

[hoom]

Doing nicely in the virtualisation tests :)

Strikes me that it could have been better to just add another 4-6MB of L3? (or would that mean it'd have to be slower L3?)

There seems to be a fairly big chunk of empty die there, with bits like IO pads not moved to match the wider die -> quick & dirty copy + paste (which I've got no big issues with mind you)

Edit: Presumably the true impact of this chip should come in the 4P setup where the HT Assist kicks in, but neither review had a 4P setup to test with.

[fellix]

Yup -- without new native multi-core design, stacking more CPU cores over the old one is likely to lead to such wasting of real estate. It's a quick patch solution, but at least it works with minimum upgrade requirements. Now if only AMD has some dedicated SRAM cell design for denser L3 cache...

Edit: Presumably the true impact of this chip should come in the 4P setup where the HT Assist kicks in, but neither review had a 4P setup to test with.
And a new chip-set platform with HT 3.x support, combined together. ;)

[hoom]

Just finished reading the TechReport review, some very interesting results there :yep2:

If only they were bulldozer cores... :sad2:

And a new chip-set platform with HT 3.x support, combined together.OK, fair enough if the mobos aren't ready yet.

[3dilettante]

Kind of a neat optical illusion where the slightly offset position of the fifth and sixth cores makes it look like they are smaller.

[fellix]

RS890 is on its way to the shelfs -- Tyan S8212 (http://www.semiaccurate.com/2009/06/04/ati-890-chipset-pictured/)
A true enterprise chip-set the new Opterons deserve, now with IOMMU support... Oh, boy! Oh, boy! http://www.hardwarebg.com/forum/images/smilies/gtzvetkov/boots.gif

[Davros]

With virtualization, guest operating systems can use hardware that is not specifically made for virtualization. Higher performance hardware such as graphics cards use DMA to access memory directly; in a virtual environment all the memory addresses are remapped by the virtual machine software, which causes DMA devices to fail. The IOMMU handles this remapping, allowing for the native device drivers to be used in a guest operating system.

does this mean better hardware 3d on vm's ?

[fellix]

Yup! ;)

[hoom]

And not like the NV style where you need a dedicated card per VM.

[fellix]

A Glance into the Future: Six-Core AMD Istanbul in a Desktop Platform (http://www.xbitlabs.com/articles/cpu/display/amd-istanbul.html)

With a better more modern core logic, the thing could fly off. ;)

[Albuquerque]

Sadly, with more modern core logic, it might not have to play catch up in 92% of the benchmarked applications. From a performance standpoint, I don't see how it's going to supplant the nearly year-old Core i7 architecture. In light of the even newer and cheaper Core i5, I can only see more pain for AMD's CPU division.

Here's hoping that I'm wrong, but I'm not very inspired by those results.

[rpg.314]

Well, it wasn't an equal matchup, but yeah, it's prolly not good enough. May be AMD should just switch to bulk process for cpu's instead of sticking to SOI, as GF is transitioning to smaller nodes faster in that area. Also, only their CPU's use the SOI process (for the most part), so having both the CPU's, and GPU's made on same process will help there too.

[Mat3]

Well, it wasn't an equal matchup, but yeah, it's prolly not good enough. May be AMD should just switch to bulk process for cpu's instead of sticking to SOI, as GF is transitioning to smaller nodes faster in that area. Also, only their CPU's use the SOI process (for the most part), so having both the CPU's, and GPU's made on same process will help there too.

What about this?

http://www.arm.com/news/26070.html

It seems to show that SOI is still providing benefits over bulk (it seems to disprove the popular opinion that the benefits of SOI are progressively diminished at the smaller nodes). AMD needs any potential advantage it can.

[rpg.314]

Then why not make gpu's on SOI? At any rate, it seems that SOI processes are slower to migrate, for whatever reasons. AMD is more than a year behind Intel in process transitions now, and the gap seems to be only increasing.

[rpg.314]

I'd say a less power efficient but smaller process will still turn out better... Are they confirmed to be using SOI for bulldozer too?

[Mat3]

They have confirmed 32nm SOI, for Fusion and most certainly Bulldozer. In the earnings call script (http://seekingalpha.com/article/166870-advanced-micro-devices-inc-q3-2009-earnings-call-transcript?page=-1) if you haven't seen it, there are a few questions about using SOI beyond 32nm, and it sounds like they might have some answers to that next month.

[entity279]

At any rate, it seems that SOI processes are slower to migrate, for whatever reasons. AMD is more than a year behind Intel in process transitions now, and the gap seems to be only increasing.

It don't think that has anything with SOI.

[hoom]

Been a while since we've had any updates here but details seem to be leaking out now.
According to Dresdenboy & other sites Thuban is shaping up to be more than I'd been expecting.

Looking like they will be launching a 3.2Ghz chip with 125W ACP.
Still AM3 socket.
They've got a Turbo Boost like auto-overclock system.
a dynamic core boosting feature called Core Performance Boost. When a specific condition is present, a subset of the cores on a system are boosted beyond their P0 operating fr
Talk of up to 3.3Ghz on the 2.8 chip & 3.6Ghz on the 3.2 :)

Had been about to finally pull the trigger on an i7 860 but I think I'll wait to see how this 3.2Ghz Thuban performs.

[Miksu]

Thanks for the info.

Do you happen to know if the new 6 core processors will work in a AM2+ -motherboard? If yes, I think i'm gonna upgrade to a new processor this summer...

[Lightman]

Thanks for the info.

Do you happen to know if the new 6 core processors will work in a AM2+ -motherboard? If yes, I think i'm gonna upgrade to a new processor this summer...

They will work in AM2+ boards with BIOS update.
There are already many boards supporting Thuban, just look at Gigabyte or Asus websites.

@hoom: surely you meant 125W TDP and not ACP :wink:
ACP is used only for server CPU's from AMD.
I was very surprised to find out Thuban will fit into 125W TDP with 3.2GHz clock across all 6 cores :shock:
A little disappointed that un-core stayed at 2GHz freq. I was hoping for 2.4GHz stock, but luckily by getting Black Edition I will set it to whatever will be stable @ reasonable voltage.

[hoom]

Oh, I thought they use ACP for desktop too? :oops:

Oh, no bump on the uncore is a bit of a downer :???:

[Neb]

Prices seem quite incredible if correct! :eek:

http://techreport.com/discussions.x/18648

We just received the official prices from AMD:
Phenom II X6 Six Core

1090T Black Edition: 125(W), AM3 Socket, 9MB Cache, 3.2 GHz Freq @ 295 USD

1055T: 125(W), AM3 Socket, 9MB Cache, 2.8 GHz Freq @ 199 USD

[Sxotty]

It sounds good I suppose, but we will have to wait and see how the $/performance actually is once they are public. It may be the bees knees for many, but I don't know about a regular desktop. I would like 6 cores though myself :)

[ShaidarHaran]

They look like nice chips, but they *still* can't compete with i7 8xx and 9xx, let alone Gulftown.

[Sxotty]

They look like nice chips, but they *still* can't compete with i7 8xx and 9xx, let alone Gulftown.

Well at least that means it is cheap. This time maybe I will try a AMD based Mobo with one. Getting a decent IGP is a bonus for use as a HTPC later in life.
edit: (I should note that IGP decency seems to be endangered. This isn't surprising given the competition or lack there of)

[eastmen]

They look like nice chips, but they *still* can't compete with i7 8xx and 9xx, let alone Gulftown.

$300 for a 6 core chip sounds good to me. The only planed 6 core intel cpu is at $1000.

If i can get a black eddition and get it close to 4ghz it may be a worthy upgrade from my q9550 at 4.2ghz.

[hoom]

I could see the 3.2Ghz x6 being fairly competitive to i7 860.

It won't quite be there in single-thread but for 3 core stuff they'll be running @3.6Ghz while the i7 is clocking back & 6 real cores @3.2 could come out better in highly threaded stuff than 4 with HT @2.8.
If the price is right, idle power is improved on X4 & if there is good overclock headroom it could be a winner.

[Squilliam]

Will it be much of an improvement for an HD 5870 (vapour X is nice) with a 2.4Ghz C2D? :-D

[ShaidarHaran]

$300 for a 6 core chip sounds good to me. The only planed 6 core intel cpu is at $1000.

If i can get a black eddition and get it close to 4ghz it may be a worthy upgrade from my q9550 at 4.2ghz.

They're fine chips, I didn't mean to imply otherwise, they just don't have the outright performance of i7 is all. I have a Q9550 @ 4GHz and don't feel the need to upgrade. Hoping to hold off for 8-core Sandy Bridge but 6-core Gulftown 970x @ $500 just might tempt me ;)

[Silent_Buddha]

They're fine chips, I didn't mean to imply otherwise, they just don't have the outright performance of i7 is all. I have a Q9550 @ 4GHz and don't feel the need to upgrade. Hoping to hold off for 8-core Sandy Bridge but 6-core Gulftown 970x @ $500 just might tempt me ;)

It's possible in about 1 year that Intel might release a mainstream 6 core i7. It'll be interesting to see how well this does in heavily threaded apps compared to a 4 core i7.

Regards,
SB

[V3]

It's possible in about 1 year that Intel might release a mainstream 6 core i7. It'll be interesting to see how well this does in heavily threaded apps compared to a 4 core i7.

Regards,
SB

In that Anand review of the Istanbul, they mentioned that some apps like renderer only see 2^n cores, so they ignored the extra 2 cores. I wonder how prevalent is this style of coding. I can see AMD doing 6 cores on 45nm but Intel should have gone straight to 8 core i7, IMO.

I am also curious why AMD don't release a dual sockect Phenom to compete with Intel in the enthusiast desktop/workstation market. They have the technology, they're getting beat really bad in that area.

[ShaidarHaran]

It's possible in about 1 year that Intel might release a mainstream 6 core i7. It'll be interesting to see how well this does in heavily threaded apps compared to a 4 core i7.

Regards,
SB

I've heard through the grapevine that a 3.2GHz multiplier-locked 970x is due out in Q3 @ ~$500.

[3dilettante]

In that Anand review of the Istanbul, they mentioned that some apps like renderer only see 2^n cores, so they ignored the extra 2 cores. I wonder how prevalent is this style of coding. I can see AMD doing 6 cores on 45nm but Intel should have gone straight to 8 core i7, IMO.

It happens in a few places. The tri-cores sometimes encounter this as well.



I am also curious why AMD don't release a dual sockect Phenom to compete with Intel in the enthusiast desktop/workstation market. They have the technology, they're getting beat really bad in that area.
AMD at one point did announce a dual-socket Athlon FX platform for enthusiasts.
It didn't go anywhere.

[Silent_Buddha]

I've heard through the grapevine that a 3.2GHz multiplier-locked 970x is due out in Q3 @ ~$500.

Oh, now that would be interesting. Wonder if it was prompted by the iminent release of the AMD Hex Core.

Regards,
SB

[Silent_Buddha]

AMD at one point did announce a dual-socket Athlon FX platform for enthusiasts.
It didn't go anywhere.

Wasn't AMD's Spider platform able to use dual FXs? I didn't pay much attention to that platform (too much power consumption) so not sure.

Regards,
SB

[3dilettante]

I think the Spider platform was an umbrella for a Phenom+RV670+AMD chipset.

QuadFX was an earlier initiative that had a dual-socket board.

[Tchock]

In that Anand review of the Istanbul, they mentioned that some apps like renderer only see 2^n cores, so they ignored the extra 2 cores. I wonder how prevalent is this style of coding. I can see AMD doing 6 cores on 45nm but Intel should have gone straight to 8 core i7, IMO.

I am also curious why AMD don't release a dual sockect Phenom to compete with Intel in the enthusiast desktop/workstation market. They have the technology, they're getting beat really bad in that area.

Hmm, for Workstation I don't know- 16 cores at a starting price of $600 (300 each) or like the FX, 1000 each for 8 higher clocked cores?

AMD seems to be pushing 2P for workstation and 4P for servers now, albeit currently at slower clockspeeds. For WS workloads it should be mighty fine though.

[hoom]

Ooh, Techreport says Thuban is 45nm SOI with low-k :shock:
http://www.techreport.com/discussions.x/18731

I thought low-k had been ruled out on AMD 45nm?

[Silent_Buddha]

Anandtech has some interesting info on it as well. Apparently if 3 cores are idle the CPU will reduce speed on those cores by 500 mhz and boost speed on the remaining 3 cores by up to 500 mhz.

So the top model has the ability to hit 3.6 ghz if 3 cores are idle.

It's not as good as Intels Turboboost however, as it has to raise voltage to the entire CPU to do this, and idle cores aren't turned off.

Still, AMD has at least caught up a bit in the feature department.

Impressive that with lightly threaded apps, this will hit higher default clocks than the Phenom x4 965 BE. And in heavily threaded apps, it'll have more cores, so in theory it should generally be faster overall in most situations than the 965 BE (at least the top 2 models).

Supposedly 4 core derivatives (salvage parts perhaps?) will appear later with AMDs version of Turbo.

Regards,
SB

[3dilettante]

Ooh, Techreport says Thuban is 45nm SOI with low-k :shock:
http://www.techreport.com/discussions.x/18731

I thought low-k had been ruled out on AMD 45nm?

Low-k dielectrics were indicated to a feature planned on later products on the 45nm node before it was even deployed.

Hi-K gate dielectrics are a separate process feature that is waiting on a node transition.

I'm not sure techreport is correct in claiming the low-k dielectric is going to help much with leakage.
Low-K reduces the capacitance of the signaling layers, which can reduce power consumption and speed up signal propagation, but the gates below those layers aren't going to change.

[gongo]

Is this for real? Sounds great....!!....almost too great. 500mhz turbo boost with 3 cores..sure the best case scenario would be like good old Counterstrike Source or Company of Heroes(works for me!).... i7(bloomfield) Turbo only goes up to 200mhz in best case! I read several benchmarks and i7 typically is about ~500mhz faster per core, so if X6 can do the Turbo as stated, i7 is going to get a real rival! 3 Cores @ 3.7ghz each (?) against the Bloomfield would render it obsolete..the cheaper Lynnfield turns out to be the better line...but the LGA1156 socket is a deader end then LGA775...afaik..i guess this is what Intel going to get after sitting on their arses since launching Bloomfield in Nov 08..

I seen prices of the X6...and the top of the line model is priced to take on i7 930...i think AMD is on to something....6 Cores at 3.2ghz will defeat i7 930 4 cores at 2.8ghz in variety of multithreaded software...if im not wrong..6 Cores at 2.8ghz model will similar defeat Intel highest value model i5 750...guess them cutting HT of i5 is another talking point!

A preview of what a 6 core X6 can do...only with slower ddr667 ram...slower motherboard chipset...slower HT line...slower core clocks....no Turbo..Woah! x86 competition looks like coming back!
http://www.xbitlabs.com/articles/cpu/print/amd-istanbul.html

[Squilliam]

Tell me when!!!! I wanna buy this as part of my debt to Dave that cannot be repaid. :-) I also need a new processor as the Core 2 Duo doesn't cut it anymore and im pretty sure both Ruby and my otherwise quiet and reposed Vapour X 5870 are both staring at me with disappointment whenever I play games using that slow Core 2. :-(

[swaaye]

I want to see this working flawlessly and actually performing before any drool starts.

AMD's been turning off power saving features in Phenom since the beginning because they were bugged. Like say, per-core CnQ and their software C1E mode that they finally replaced now (just for Athlon II AFAIK).

Actually if you run a Phenom II in XP it will be a dog because they never disabled the per-core CnQ for XP. You end up with apps running on cores clocked down. You actually need to get 3rd party utils to get around this or disable CnQ.

Has anyone tried a Core i7 in XP?

[hoom]

Hi-K gate dielectrics are a separate process feature that is waiting on a node transition.Right you are, I was thinking HKMG :oops:

AMD's been turning off power saving features in Phenom since the beginning because they were bugged.I thought it was Windows scheduler that is bugged? or at least does some strange things.
But you'd have thought AMD would have at least checked up with MS first to confirm the behaviour &/or worked with them to alleviate the issues...

[Lightman]

Slowness in XP with C'n'Q enabled was caused primarily by MS (really bad) scheduler and relatively slow transition of AMD cores from idle to full speed. This was/is happening because most of C'n'Q states are controlled by BIOS through chipset. The new Phenoms X6 will have much faster P-state transitions thanks to on chip circuity controlling it's states (aka PCU in Nehalem).

AMD says new Thuban core will be fine even in unpleasant Windows XP environment :smile:

[swaaye]

AMD had to disable Phenom's independent clocking in NT6 too so XP isn't entirely to blame. Unlike with NT6, AMD never locked the core clocks together in XP so there you still get terrible performance (unless you are using all cores?). I ran into this first hand with my Phenom II X4 and XP, when I was playing some Dark Messiah and noticed that the load times were significantly longer than I was used to with my old C2D.

It'll be great to see everything finally working correctly though. If it all works out.

[Tchock]

Well... 4.2Ghz OC on 4 cores at 0.05v surplus (1.4) is impressive. But again it's a 32-bit score, let's see how 64b works now.

[gongo]

http://translate.google.com/translate?u=http%3A%2F%2Fwww.expreview.com%2F10081 .html&sl=zh-CN&tl=en&hl=&ie=UTF-8

Some results are out! Wow the turbo clock works, a full 500mhz overclock, Lynfield like aggression...but the vcore got a huge bump... reaching closer to PII max of 1.55v @ 69C IIRC....and also for single threaded! AMD is known to be overconservative with Vcore...i can likely set a lower vcore..wonder how the BIOS select the vcore stepup.

The good news is at default, the middle of the line PII X6 is as fast as i7 860! 2.8ghz 6C = 2.8ghz 4C/8T but with a much cheaper platform costs! I think AMD is going for a good fight back here! Intel has to cut Bloomfield prices! Yippie!

[rpg.314]

Some results are out! Wow the turbo clock works, a full 500mhz overclock, Lynfield like aggression...but the vcore got a huge bump... reaching closer to PII max of 1.55v @ 69C IIRC....and also for single threaded! AMD is known to be overconservative with Vcore...i can likely set a lower vcore..wonder how the BIOS select the vcore stepup.

The good news is at default, the middle of the line PII X6 is as fast as i7 860! 2.8ghz 6C = 2.8ghz 4C/8T but with a much cheaper platform costs! I think AMD is going for a good fight back here! Intel has to cut Bloomfield prices! Yippie!

Clock for clock, thuban needs 50% more cores and a bit larger cache (9M total vs 8M in Lynnfield). I think I need to underclock my expectations from Bulldozer. :cry:

[swaaye]

Bulldozer isn't a bunch of boring Phenom K10 cores though. It could very well be a horrible disaster, but that is the chip I want to hear about. :D

This hex core Istanbul chip reminds me of the huge Intel Dunnington hex core chip in that it is just an interim solution.

[gongo]

http://www.xtremesystems.org/forums/showpost.php?p=4336220&postcount=36

http://forum.coolaler.com/showthread.php?t=235891&page=3

More benches! Let us not forget the point that this chip is priced to take on i5 750 and it does handily beat i5 750 in multithreaded benches! ...i like to see some games ....AFAIK games likes the big L3 cache...

[swaaye]

It has the same amount of L3 as the other Phenom IIs and there are only a few games that even do > 2 cores. That's why they duped Intel's Turbo Boost, because adding more cores isn't helping performance very often and there's obviously a problem with heat (and surely yields) when clocking more and more cores really high.

[gongo]

http://forum.coolaler.com/showthread.php?t=235972&page=3

MOAR! Thuban runs circles around the higher priced i7 in multithreaded apps! Thuban overclocks well, 3.85Ghz at only 1.424v, at this kind of overclock, you will need i7 quad cores to be at 4Ghz to beat Thuban real cores! AMD has a winner in their hands! I wonder how the review sites will grade Thuban...has to be as big win like i5 750....speaking of which...it will be obsolete at $199...now for Intel to speedstep the release of 6 cores i7 at LOAR prices! Bloomfield obscenely stubborn prices has to go right away! Anyone heard of the quarterly Intel price cuts coming?

Im frankly surprised AMD pulled this off.....its still 45nm...its still PII microarchitecture...kinda defy logic like their GPU brothers did with HD48/5800 series!

[ShaidarHaran]

http://forum.coolaler.com/showthread.php?t=235972&page=3

MOAR! Thuban runs circles around the higher priced i7 in multithreaded apps! Thuban overclocks well, 3.85Ghz at only 1.424v, at this kind of overclock, you will need i7 quad cores to be at 4Ghz to beat Thuban real cores! AMD has a winner in their hands! I wonder how the review sites will grade Thuban...has to be as big win like i5 750....speaking of which...it will be obsolete at $199...now for Intel to speedstep the release of 6 cores i7 at LOAR prices! Bloomfield obscenely stubborn prices has to go right away! Anyone heard of the quarterly Intel price cuts coming?

Im frankly surprised AMD pulled this off.....its still 45nm...its still PII microarchitecture...kinda defy logic like their GPU brothers did with HD48/5800 series!

Performance in synthetic benchmarks can by no means be equated to real-world performance. If you want to believe that Phenom II x6 is going to out-perform any 8-thread or greater i7 in real-world apps, you're in for a disappointment.

[Silent_Buddha]

Performance in synthetic benchmarks can by no means be equated to real-world performance. If you want to believe that Phenom II x6 is going to out-perform any 8-thread or greater i7 in real-world apps, you're in for a disappointment.

In games it's highly unlikely that Istanbul will beat out Lynnfield (unless limited by GPU), however in server and workstation apps, I wouldn't be surprised if it traded blows with Bloomfield and/or Lynnfield depending on how heavily threaded an app is.

In heavily threaded apps, HT doesn't really do much compared to real cores. If the app is only lightly threaded (low CPU load for each thread) HT can help to some degree.

Regards,
SB

[ShaidarHaran]

In games it's highly unlikely that Istanbul will beat out Lynnfield (unless limited by GPU), however in server and workstation apps, I wouldn't be surprised if it traded blows with Bloomfield and/or Lynnfield depending on how heavily threaded an app is.

In heavily threaded apps, HT doesn't really do much compared to real cores. If the app is only lightly threaded (low CPU load for each thread) HT can help to some degree.

Regards,
SB

When discussing server and workstation workloads we should be discussing server and workstation chips, i.e. Xeon and Opteron. Anandtech reviewed Magny Cours (12 core Opteron) back in March here (http://www.anandtech.com/show/2978/amd-s-12-core-magny-cours-opteron-6174-vs-intel-s-6-core-xeon) and compared to Intel's then top-of-the-line Xeon SKUs, based on 6-core Nehalem derivatives (now super-ceded by 8-core Nehalem EX).

The conclusion of the article:
The beancounters will probably point out that AMD’s strategy of bolting two CPU dies at 346 mm² together is quite costly. But this is the server CPU market, margins are quite a bit higher. Let AMD worry about the issue of margins. If AMD is willing to sell us - IT professionals - two CPUs for the price of one, we will not complain. It means that the fierce competitive market is favoring the customer. The bottom line is: is this twelve-core Opteron a good deal? For users waiting to use it in a workstation we have our doubts. You’ll benefit from the extra cores when rendering complex scenes, but in all other scenarios (quick simple rendering, modeling) the higher clocked and higher IPC Xeon X5600 series is simply the better choice.

Applications based on transactional databases (OLTP and ERP) are also better off with new Xeon. The SAP and our own Oracle Calling Circle benchmark all point in the same direction. Intel has a tangible performance advantage in both benchmarks.

Data mining applications clearly benefit from having “real” instead of “logical” cores. For datamining, we believe the 12-core Opteron is the clear winner. It offers 20% better performance at 20% lower prices, a good deal if you ask us. Intel’s relatively high prices for its six-core are challenged. The increased competition turns this into a buyers market again.

And then there is the most important segment: the virtualization market. We estimate that the new Opteron 6174 is about 20% slower than the Xeon 5670 in virtualized servers with very high VM counts. The difference is a lot smaller in the opposite scenario: a virtualized server with a few very heavy VMs. Here the choice is less clear. At this point, we believe both server CPUs consume about the same power, so that does not help either to make up our minds. It will depend on how the OEMs price their servers. The Opteron 6100 series offers up to 24 DIMMs slots, the Xeon is “limited” to 18. In many cases this allows the server buyer to achieve higher amount of memory with lower costs. You can go for 96 GB of memory with affordable 4 GB DIMMs, while the Intel server is limited to 72 GB there. That is a small bonus for the AMD server.

The HPC market seems to favor AMD once again. AMD holds only a small performance advantage, and this market is very cost sensitive. The lower price will probably convince the HPC people to go for the AMD based servers.

You can be sure Nehalem EX with 8 cores, 16 threads, and 24MB L3 is faster than the Xeon models used in the above comparison. For most heavily-threaded workloads, Intel's architecture just plain beats AMD's. This doesn't make AMD's chips bad, it just makes them not as fast.

[Xenus]

I went to a small presentation about Magny Cours last friday at my university. It was somewhat interesting but it was more performance per watt based than anything else. Learned some interesting things about heat throttling was about it.

[Silent_Buddha]

You can be sure Nehalem EX with 8 cores, 16 threads, and 24MB L3 is faster than the Xeon models used in the above comparison. For most heavily-threaded workloads, Intel's architecture just plain beats AMD's. This doesn't make AMD's chips bad, it just makes them not as fast.

Sure 8 core Nehalem (Beckton) will erase much of the gains Magny-Cours had in some application scenarios over 6 core Nehalem/Westmere (Gulftown), but at what cost? With prices ranging from ~2,500 - ~3,700 USD (for the 8 core models), it had better perform rings around the competition.

So an 8 core will be higher performance, but since Magny-Cours was already faster in some applications at a 20% lower price than the Gulftown (~1450 USD) that was used, I'm not sure 8 core Beckton will change the conclusion much.

The Magny-Cours would still hold some significant advantages. And oddly enough despite losing in performance to an 8 core Beckton, it's attractiveness in situations where it won against Gulftown would only grow against an 8 core Beckton. And those situations where it was unattractive to Gulftown would either get worse or oddly again get better due to price.

Coming in at half the price of the cheapest 8 core Beckton CPU is going to get a lot of design wins, despite being a lower performing part.

Anyway, I digress. As you quite well pointed out (directly or indirectly) I was most definitely wrong in mentioning performance in server applications as that presumably isn't what Istanbul is targetted at.

In the desktop and workstation space, Istanbul will have similar advantages and disadvantages. Although probably more disadvantages than advantages in the desktop space where there's not a lot of heavily threaded applications. Then again the desktop space is also where one is least likely to notice the speed differences in CPU's with many people being limited by their GPU when gaming, and most other applications that aren't niche products not fully utilising all of current or past CPUs.

And I certainly don't disagree that AMD is quite definitely behind Intel with regards to CPU design and implementation currently, and I'm not convinced Bulldozer will change that much at all.

However, given the position AMD is in, they are doing what they can to make their CPU's attractive. And so far, Intel is willing to "help" them out in this regard. It doesn't hurt that "helping" AMD doesn't hurt Intel.

AMD basically has to play the cost effectiveness card, and in certainly applications they definitely win out in the cost/performance. That's fine and dandy as that still leaves the majority of the market to Intel. And while Intel could certainly crush them with prices if they wished, they are more than happy to have extremely large margins and allow space for a CPU competitor to survive, thus reducing their chances of being branded a monopoly.

Regards,
SB

[ShaidarHaran]

33% more cores/threads with an L3 twice the size of a processor that already out-performs the competition in many benchmarks should yield pretty sizable gains. Cost matters to some markets, but not all. If performance is king, I think the clear choice is Nehalem EX.

That being said I'm glad AMD isn't out of the game by any means and truly do look forward to Bulldozer.

[JardeL]

Best codename ever. ^_^

[V3]

So this X6 got released date around 27th of this month, where are the reviews ? The price seems competitive if the performance is there.

Also is the new mobo AMD 890GX meant to replace the 790FX or will there be a 890FX ?

[ShaidarHaran]

Despite my earlier skepticism, I am interested in Thuban for its potential @ encoding media. I may just end up replacing one of my Core 2 Duo systems with an AMD X6 if the price is right and the media encoding performance is there.

[Lightman]

So this X6 got released date around 27th of this month, where are the reviews ? The price seems competitive if the performance is there.

Also is the new mobo AMD 890GX meant to replace the 790FX or will there be a 890FX ?

Yes, there is 890FX coming the same day as Phenoms X6.
Here in UK end of last week you could buy Asus Crosshair IV based on 890FX, but obviously AMD's ninjas took care of that early bird quickly :wink:

I'm looking forward to this upgrade as I will be jumping from 790FX/SB600 DDR2 platform to 890FX/SB850 DDR3 one. Also my good old Phenom 940BE finally will have new home at my brothers gaming rig because prettier 1090T (I hope available in UK next week) is going to power my gaming and media encoding fun :razz:

[V3]

Thanks Lightman. X6 seems to be on retails hand already. My buddy who is in the business offered me X6, I think it was the 3.2 GHz version. The price is about the same as i7 930, which is not too bad if the performance is there but he only have the 890GX mobo currently. Can't wait for proper reviews, but those early impressions look promising.

[swaaye]

I think the only really exciting aspect of these CPUs is that AMD copied Turbo Boost so we'll get some single threaded app boosts hopefully. But there's no way I'd replace a quad from the past year or two with one of these chips.

I want to see Intel's next "tock" and just what Bulldozer ends up being.

[Sxotty]

What I wonder is will the AM3 socket take whatever CPU amd releases after this?

I need to get another CPU and motherboard as I gave away my old HTPC, but I don't want to break the bank. At the same time I would not mind something faster than my X3 phenom 720BE or whatever stupid name it has. I think it is something like that.

I am tempted to get a cheapo AMD processor for the HTPC that I can put in my current setup and buy a new motherboard and ram for my current processor. Or I suppose I should just splurge and get an i7 of some kind since single threaded performance is probably the most important to me. I can't code for crap so all my programs I need to run are single threaded and take forever to finish (like 5 hours run time).

[Tchock]

33% more cores/threads with an L3 twice the size of a processor that already out-performs the competition in many benchmarks should yield pretty sizable gains. Cost matters to some markets, but not all. If performance is king, I think the clear choice is Nehalem EX.

That being said I'm glad AMD isn't out of the game by any means and truly do look forward to Bulldozer.

Too bad the EX gets outperformed by a dual Gulftown like there's no tomorrow.

Keep in mind that for 1 (ONE) EX you're going against 2 Gulftowns or 2-4 Magnys (depending on cost of the EX and Magnys).

If performance is king I'd definitely choose the latter 2. Nehalem EX is just here for the nice RAS features in case people don't want RISC/IA-64 by large.

[Blazkowicz]

What I wonder is will the AM3 socket take whatever CPU amd releases after this?

I need to get another CPU and motherboard as I gave away my old HTPC, but I don't want to break the bank. At the same time I would not mind something faster than my X3 phenom 720BE or whatever stupid name it has. I think it is something like that.

I am tempted to get a cheapo AMD processor for the HTPC that I can put in my current setup and buy a new motherboard and ram for my current processor. Or I suppose I should just splurge and get an i7 of some kind since single threaded performance is probably the most important to me. I can't code for crap so all my programs I need to run are single threaded and take forever to finish (like 5 hours run time).

with that "Black Edition" processor, i.e. unlocked, remember that you can just up the CPU multiplier in the BIOS to get higher single-thread performance, without needing good ram and a good motherboard.
else look at the i5 650 for your stuff, it will be awesome (but overpriced)

[Sxotty]

with that "Black Edition" processor, i.e. unlocked, remember that you can just up the CPU multiplier in the BIOS to get higher single-thread performance, without needing good ram and a good motherboard.
else look at the i5 650 for your stuff, it will be awesome (but overpriced)

I did have it sped up a bit 2.8GHz-->3.2 GHz, but there were a few stability issues so I lowered the speed. However I don't think they are CPU related so I might turn up the multiplier again. When you are running at 100% CPU for hours on end though you don't want a crash at hour 4.5/5 :) After looking more at what Intel has done with their chipsets though I do not feel inclined to go that route. (in otherwords I am annoyed with Intel now after trying to see how much a mobo for i7 would be compared to a $35 AM3 mobo).

[Blazkowicz]

eek, a $35 mobo has a crappy weak power stage, and the phenom aren't a low power offering. low end full ATX mobos are MUCH better (with nforce 720d or amd 770), though perhaps on the crappy side of bits in some ways.
I can't overclock 1% of fsb (used to do 25%, then 20%, before trying bad sticks of ram). but I am 100% stable. I wish I had a BE CPU such as the X2 550. At least there's undervolting left to me, to better use the CPU.

[ShaidarHaran]

Too bad the EX gets outperformed by a dual Gulftown like there's no tomorrow.

Keep in mind that for 1 (ONE) EX you're going against 2 Gulftowns or 2-4 Magnys (depending on cost of the EX and Magnys).

If performance is king I'd definitely choose the latter 2. Nehalem EX is just here for the nice RAS features in case people don't want RISC/IA-64 by large.

1) what apps?
2) where's the link?

[Albuquerque]

Well, in terms of pure performance, I wouldn't be surprised that a dual Gulftown (twelve cores, twenty four threads) would generally outperform a single EX (8 cores, sixteen threads) in a workload that's heavily multithreaded. I might be able to buy that a quad Magny could also win against EX also..

But when you talk about multicore platforms, you're talking about seriously expensive motherboards (and perhaps FB-DIMMs too.) And on the Intel side of the fence, you're talking about Xeon-branded processors which add even more. Where do you find a quad-capable Magny-compatible motherboard? And what's the cost on that? Are there Gulftown Xeons out yet? What are the prices on those?

I'm not convinced that you're going to win a price vs perf argument when you go in the multi-physical-CPU conversation -- even with the added performance that they will invariably bring along.

[Albuquerque]

Damn, perhaps I spoke too soon ;)

So, as an addendum to my own post above and Tchock's post. I found one place selling the Xeon 7560 (Nehalem EX Octal-core 2.26Ghz) for around four thousand dollars. Yikes. You'll need the board that can do LGA 1567, which I cannot find immediate pricing for. Figure it won't be cheap ;)

The six-core, twelve thread Xeon 5820 (2.26Ghz) you can get for around $400 each, and you can get DP motherboards for around $300.

The octal-core Magnycors (2.0Ghz) you can find for $300 each, and you can get quad processor boards for about $800.

Looks like the sweetspot (considering only processor hardware and motherboard) will probably be Gulftown at this point in time.

[Sxotty]

eek, a $35 mobo has a crappy weak power stage, and the phenom aren't a low power offering. low end full ATX mobos are MUCH better (with nforce 720d or amd 770), though perhaps on the crappy side of bits in some ways.
I can't overclock 1% of fsb (used to do 25%, then 20%, before trying bad sticks of ram). but I am 100% stable. I wish I had a BE CPU such as the X2 550. At least there's undervolting left to me, to better use the CPU.

I am not saying I will get a $35 board. I am saying the low end of AM3 is cheap. The low end of x58 is Frickin expensive. I can get a great AM3 board for about 100. That is what happens thanks to AMDs woeful inability to compete with intel i7 chips and no 3rd party intel chipsets. Intel can charge whatever they want for the boards for their fast chips and enthusiasts pay through the nose.


BTW I did turn mine back up to 3.2Ghz and it is fine. It was a AHCI issue before from swapping after windows was installed. Given the above sentiments though it is unlikely I will go the intel route.

[ShaidarHaran]

There's only one scenario in which Westmere EP ("Gulftown Xeon") beats Nehalem EX, and that's at the two socket level. Westmere EP processors aren't available in configurations greater than two sockets so Nehalem EX gets an automatic win when you scale beyond 2 sockets. Keep in mind I'm not talking at all about the cost of any of the platforms in question, just the performance. If we take cost into account then clearly Westmere EP and Magny Cours are better solutions than Nehalem EX.

[Blazkowicz]

maybe Power 7 kicks its ass, if cost becomes negligible.

it's clearly meant for "minicomputers", or high end enterprise server, however you call them. Who knows, at $4000 per CPU, maybe it's cheaper than IBM, and it's dwarfed by the software that runs on it (Oracle or similar).

[Tchock]

There's only one scenario in which Westmere EP ("Gulftown Xeon") beats Nehalem EX, and that's at the two socket level. Westmere EP processors aren't available in configurations greater than two sockets so Nehalem EX gets an automatic win when you scale beyond 2 sockets. Keep in mind I'm not talking at all about the cost of any of the platforms in question, just the performance. If we take cost into account then clearly Westmere EP and Magny Cours are better solutions than Nehalem EX.

Yawn. That's not a convincing argument for a lot of workloads.

It's probably easier to you know, get two+ units of Westmere/Magny servers and run the software in separate instances. Sure, you *can* get even more EX servers, but it'll be lovely looking at how this war of attrition plays out when the costs go into the 5/6-digits. :wink:

The factoring in of cost makes it possible to scale multiple workloads to multiple systems giving more performance for still less $. Considering AMD's 2P/4P strategy and slot-in Bulldozer (along with no visible upgrade for LGA 1366 after Westmere) which we have no/"some" idea in performance wrt Nehalem EX, I'm pretty sure they're getting the design wins for anything that does not need RAS or does not have Westmere fit the criteria (or is just too expensive).

As for server purchases it should go like this regarding performance/cost:

16-Magny, 12-Westmere, 32-Magny (only ~1000USD more than 16-Mg disregarding more RAM), 2x12-Westmere, 2x32-Magny, (here is where 16-Beckton comes in at cost).

If the app doesn't scale, remove everything and leave Westmere.

[Blazkowicz]

Yawn. That's not a convincing argument for a lot of workloads.

It's probably easier to you know, get two+ units of Westmere/Magny servers and run the software in separate instances. Sure, you *can* get even more EX servers, but it'll be lovely looking at how this war of attrition plays out when the costs go into the 5/6-digits. :wink:


Unless it's cheaper to not split your database, thanks to a bigger SMP machine with a single adressable memory pool?

but sure, it's useless for the common mortal or companies like google.
it's amazing that IBM PC compatibles have come so far. I'm picturing out getting a 8-way nehalem EX, then booting MS-DOS 5.0 instead of that boring Unix/Windows stuff to play good old Arkanoid and Doom 2 on it :razz:

[ShaidarHaran]

Yawn. That's not a convincing argument for a lot of workloads.

It's probably easier to you know, get two+ units of Westmere/Magny servers and run the software in separate instances. Sure, you *can* get even more EX servers, but it'll be lovely looking at how this war of attrition plays out when the costs go into the 5/6-digits. :wink:

The factoring in of cost makes it possible to scale multiple workloads to multiple systems giving more performance for still less $. Considering AMD's 2P/4P strategy and slot-in Bulldozer (along with no visible upgrade for LGA 1366 after Westmere) which we have no/"some" idea in performance wrt Nehalem EX, I'm pretty sure they're getting the design wins for anything that does not need RAS or does not have Westmere fit the criteria (or is just too expensive).

As for server purchases it should go like this regarding performance/cost:

16-Magny, 12-Westmere, 32-Magny (only ~1000USD more than 16-Mg disregarding more RAM), 2x12-Westmere, 2x32-Magny, (here is where 16-Beckton comes in at cost).

If the app doesn't scale, remove everything and leave Westmere.

Is the number of workloads relevant? I wouldn't think workloads in the server market would be extremely diverse...

[hoom]

So Thuban reviews are out now, haven't read any yet (lost a HDD & been out getting a new 2TB Caviar Green) but some are:
http://www.anandtech.com/show/3674/amds-sixcore-phenom-ii-x6-1090t-1055t-reviewed
http://www.xbitlabs.com/articles/cpu/display/phenom-ii-x6-1090t.html
http://www.guru3d.com/article/phenom-ii-x6-1055t-1090t-review/
http://www.legitreviews.com/article/1289/1/
http://www.hexus.net/content/item.php?item=24332
http://www.bit-tech.net/hardware/cpus/2010/04/27/amd-phenom-ii-x6-1090t-black-edition/1

Die pic from Anandtech
http://images.anandtech.com/reviews/cpu/amd/phenom2/x6/Phenom_II_x6_die_clear_48078C.jpg
Seems to be nearly identical to Istanbul die O_o
I'd thought they might have tidied up the large scale layout a bit given the extra time.
But then, higher clocks, reduced power draw, turbo & price matter. Aesthetics of die layout won't affect whether people buy this.

[Erinyes]

Looking at the reviews and the performance of Thuban against Gulftown, AMD better be targeting a Pentium 4 to Core 2 type jump for Bulldozer, they're gonna get steamrolled by Sandy Bridge otherwise (pun intended :lol:). I seriously hope they're targeting at least a 30-40% increase in IPC tho. And 32nm HKMG will help too

[V3]

Finally some reviews. It looks pretty good. Anandtech don't paint a good picture of X6, and what's what with their idle power consumption ? It seems to be way higher than other reviews. Other reviews is better though.

What was the process tweak that reduced power consumption BTW ? High k ?

[hoom]

Yeah, some weird performance variations around.

Seems to be some apps that don't properly trigger the turbo or something so x6 is sometimes scoring way less than the x4.
Also either the AMD fusion software seems to not be showing turbo clocks properly?

Some nice things like that BE models get BIOS level turbo multiplier tweaking & new 980FX chipset with the IOMMU :grin:

edit: ugh, bit-tech WTF moment
Power Consumption
For all of the performance tests, we disable all power saving technology in order to give us a consistent set of results, and also best-case performance numbers - even though technologies such as Intel's SpeedStep might only take microseconds to kick in, that can make a difference in some tests.

However, for the power consumption tests we re-enable everything in order to get a real-world power draw.
This means that their tests are actually of no practical value & they even complain about turbo not working during their testing (hint to bit-tech guy, it uses the cool'n'quiet tech that you turned off)

[fehu]

AMD suffer from the "good product, by no way exciting" syndrome :(
ehm... the famous GPBNWES...

[Erinyes]

Yeah, some weird performance variations around.

Seems to be some apps that don't properly trigger the turbo or something so x6 is sometimes scoring way less than the x4.
Also either the AMD fusion software seems to not be showing turbo clocks properly?

Some nice things like that BE models get BIOS level turbo multiplier tweaking & new 980FX chipset with the IOMMU :grin:

edit: ugh, bit-tech WTF moment

This means that their tests are actually of no practical value & they even complain about turbo not working during their testing (hint to bit-tech guy, it uses the cool'n'quiet tech that you turned off)

Im sure you meant 890FX :) Apart from now having SB850, seems to be the same old 790FX silicon, hell its still on 65nm! What is IOMMU though?

Yea you're right if they turned off power saving would that let any of the cores turbo up? I guess even Intel's turbo mode wouldnt work either. This needs to be investigated further

[Blazkowicz]

Seems to be nearly identical to Istanbul die O_o
I'd thought they might have tidied up the large scale layout a bit given the extra time.
But then, higher clocks, reduced power draw, turbo & price matter. Aesthetics of die layout won't affect whether people buy this.

Is there any difference? I'm under the impression they use the exact same core.
if so the same core could span litterally every socket, depending on configuration/packaging : AM2 and AM3, F, G34, C32 (ddr3 socket F derivate for two-CPU systems)

[Erinyes]

Yea i think the turbo mode is now implemented in hardware right? That would need new silicon. Else they could have been shipping six core desktop chips for almost a year now. Isnt G34 only for Magny cours?

[Blazkowicz]

yes, but Magny-Cours is two Istanbul cores on the same package (or two Shangai for the 8-core I presume)

We can find reasons for the dektop 6-core coming now :
- it was needed first on the opteron's market, where performance/watt is a big deal
- letting the 965 BE be the flag ship instead of cannibalizing sales
- let the frequencies ramp up, so you can launch high frequency X6s rather than 2.2GHz and 2.4GHz ones

[Erinyes]

yes, but Magny-Cours is two Istanbul cores on the same package (or two Shangai for the 8-core I presume)

We can find reasons for the dektop 6-core coming now :
- it was needed first on the opteron's market, where performance/watt is a big deal
- letting the 965 BE be the flag ship instead of cannibalizing sales
- let the frequencies ramp up, so you can launch high frequency X6s rather than 2.2GHz and 2.4GHz ones

Ahh yes that is of course true, but i think its the same die only for the server sockets. AFAIK the desktop chips are different silicon

Yes true definitely needed in the server market but it could have sold reasonably well in the desktop market too. From a marketing point of view 6 cores > 4 and at that time they were shipping at 2.6 ghz. They could have simply priced it higher than the 965 BE, that wouldnt have been an issue. I think they delayed it to add the turbo core feature in hardware as intel had done and of course as you said to let the process mature and launch with higher frequencies.

[hoom]

Cool, seems like 890FX is a pretty decent chipset.
http://techreport.com/articles.x/18825/1
I haven't really paid all that much attention to AMD chipsets lately but have been under the impression that they suffered sub-par performance in stuff like USB throughput.
It would appear that this is well fixed with the 890 northbridges/850 southbridge :smile:

I'm thinking my new build may be a good 890FX mobo with a 1055T to keep price down & leaving me primed for upgrade to Bulldozer next year.
That is assuming that Bulldozer is still AM3 compatible? I think I saw something about AM3+ :-|

[Erinyes]

Cool, seems like 890FX is a pretty decent chipset.
http://techreport.com/articles.x/18825/1
I haven't really paid all that much attention to AMD chipsets lately but have been under the impression that they suffered sub-par performance in stuff like USB throughput.
It would appear that this is well fixed with the 890 northbridges/850 southbridge :smile:

I'm thinking my new build may be a good 890FX mobo with a 1055T to keep price down & leaving me primed for upgrade to Bulldozer next year.
That is assuming that Bulldozer is still AM3 compatible? I think I saw something about AM3+ :-|

Nah the UBS issue was way back, it was pretty much fixed with SB 7xx onwards and SB 850 has improved it further. But the SATA controllers still have poorer performance than Intel. Heres a link to a review if you have any other doubts :smile: http://www.anandtech.com/show/2952

The question is, do you really need six cores? It makes sense only if you primarily run well threaded apps. Abt compatibility, considering AMD's track record till now, i think we can expect AM3 at the very least to be compatible but as ever we cant say for sure till closer to the launch. Intel's moving to new sockets with Sandy bridge though

[Sxotty]

From what I remember on anandtech the SATA controllers do worse than prior gen SATA controllers as well. That is too bad. I am still tempted to do this as a person can run multiple single threaded program on the CPU. Last night I had three models running on my computer for about 8 hours each. With this I could double my throughput which is tempting.

[Kej]

Finally some reviews. It looks pretty good. Anandtech don't paint a good picture of X6, and what's what with their idle power consumption ? It seems to be way higher than other reviews. Other reviews is better though.

What was the process tweak that reduced power consumption BTW ? High k ?

One tweak is probably low-k (http://en.wikipedia.org/wiki/Low-k_dielectric). It's used for inter connects.
Hans de Vries (http://www.xtremesystems.org/forums/showthread.php?p=4330334#post4330334) has said that he believes Thuban has low-k down to 2.4.

AMD aren't supposed to use high-k until the 32nm process. High-k is used in the transistors.
So you can use high/low-k independently.

[bearmoo]

Finally some reviews. It looks pretty good. Anandtech don't paint a good picture of X6, and what's what with their idle power consumption ? It seems to be way higher than other reviews. Other reviews is better though.


:roll: It's Anand after all, his words brighten with emotion whenever he mentions that some other company's name

[hoom]

I'm thinking my new build may be a good 890FX mobo with a 1055T to keep price down & leaving me primed for upgrade to Bulldozer next year.OK, so I've gone for 1055T + 890GX + 4GB of DDR3 1600.
Cost about the same as i5 750 + P55 mobo + 4GB DDR3 1600.

Pricing on 890FX mobos is quite high (NZ$1-200 more than 890GX!) & there is nearly NZ$200 between 1055T & 1090T :shock:
Still quite comparable to i7 860 & good P55 mobos I guess, but if I was going to push my budget that high I'd probably actually go for the i7 860.

Next year Bulldozer (unless it sucks &/or isn't compatible with AM3) & a replacement for my 4870!