Take This Magny-Cours!!!

[BadTrip]

http://it.anandtech.com/IT/showdoc.aspx?i=3769

LOL!!!

 

[yannifb]

I think everyone was already aware of this.

 

[MU_Engineer]

I would wait to see an actual comparison of these chips vs. Magny-Cours before gloating. Magny-Cours brings twice the core count per chip than the Westmeres, and that's a very big advantage in AMD's favor in anything very multithreaded (which is many of these server benchmarks.) I would particularly wait until we know what Magny-Cours chips will cost because rumor has it that the lower-end Magny-Cours chips will be far less expensive than their current 4P-capable server CPUs. If it takes a $1000 Xeon to beat a $500 Magny-Cours (I'm just throwing out numbers to make an example), guess who will sell more chips. Just look at the desktop market where AMD pretty much cleans up the CPU recommendations in the entire <$200 market to get an idea. Also, power consumption is a huge issue with server buyers. Intel will have to lean heavily on clock speed to overcome Magny-Cours' core count advantage, and increasing clock speed a lot increases power consumption even more. I don't know how this will turn out, but I do expect AMD to be far more competitive than they are currently.

Also, BadTrip, you should want Magny-Cours to wipe the floor with Westmere. Intel very noticeably did not release any six-core Xeons costing less than $1000. I bet that's because they don't want desktop LGA1366 users to grab six-core Xeons that cost less than $1000 instead of the $1000 i7 980X. AMD outperforming the Xeon 5600s at equivalent price points would push the price of the Xeon 5600s down and either give desktop users access to <$1000 six-core Xeons and/or force Intel to release six-core Core i7s for <$1000. And if you are not a fan of six-core chips, the 32 nm quad-core Xeon 5600s with their 50% cache size increase and likely higher overclock potential relative to Bloomfield would also become less expensive. So any way you see it, you should be rooting for AMD to dethrone Intel as Intel has gotten back to their '90s and early 2000s-era attitude of "we're invincible, so we'll slowly mete out gradual improvements and keep prices high" laziness during the last few years. Intel needs a swift kick to the teeth by AMD for them pull off a sequel to Core 2's launch day.

 

[yannifb]

Also Magny cours isn't just Istanbul with double the core count, it has new tech like a quad channel memory controller and faster Hypertransport speeds. Thats all I know of but i think theres more.

 

[jimmysmitty]

Also Magny cours isn't just Istanbul with double the core count, it has new tech like a quad channel memory controller and faster Hypertransport speeds. Thats all I know of but i think theres more.

I think thats about it.

And MU, I think he mean Magny-Cours because so far the benchmarks that Anand put up show the Westmere single 6 core CPU keeping up with a dual CPU Istanbul setup and in some cases quad Optys. The Xeon setup also only had 24GB of DDR3 while the Optys had 32 GB (dual) and 64 GB (quad). Probably not that much of a difference but still.

As for Magny-Cours, Nehalem-EX will be the true competator to that. Its only a 8 core but has quad channel DDR3 as well as 4x QPI links

Will be nice to see. Oh and MU, Intel will have a 6 core for the DT market at the $500 mark, the Core i7 970.

 

[BadTrip]

Best answer selected by BadTrip.

 

[randomizer]

Give me 8 of these.

 

[fazers_on_stun]

I would wait to see an actual comparison of these chips vs. Magny-Cours before gloating. Magny-Cours brings twice the core count per chip than the Westmeres, and that's a very big advantage in AMD's favor in anything very multithreaded (which is many of these server benchmarks.) I would particularly wait until we know what Magny-Cours chips will cost because rumor has it that the lower-end Magny-Cours chips will be far less expensive than their current 4P-capable server CPUs. If it takes a $1000 Xeon to beat a $500 Magny-Cours (I'm just throwing out numbers to make an example), guess who will sell more chips. Just look at the desktop market where AMD pretty much cleans up the CPU recommendations in the entire <$200 market to get an idea. Also, power consumption is a huge issue with server buyers. Intel will have to lean heavily on clock speed to overcome Magny-Cours' core count advantage, and increasing clock speed a lot increases power consumption even more. I don't know how this will turn out, but I do expect AMD to be far more competitive than they are currently.

From the AT article:

Even worse, 12 Westmere cores are enough to come very close to the performance of a 24-core Opteron machine. This is does not bode well for the newest octal and twelve-core Opterons (Magny-cours). To be really frank, we think the SAP market is Intel owned until AMD launches the multi-threaded Bulldozer CPU. Most of the SAP server market is not very sensitive to pricing, let alone CPU pricing. SAP projects, which need expensive licenses and many consulting hours are typically in the $100K to $100M range and x86 hardware costs are most of the time only a small percentage of the total project costs. The final blow is the appearance of the Nehalem EX at the end of this month.

 

[MU_Engineer]

I think thats about it.

And MU, I think he mean Magny-Cours because so far the benchmarks that Anand put up show the Westmere single 6 core CPU keeping up with a dual CPU Istanbul setup and in some cases quad Optys. The Xeon setup also only had 24GB of DDR3 while the Optys had 32 GB (dual) and 64 GB (quad). Probably not that much of a difference but still.

As for Magny-Cours, Nehalem-EX will be the true competator to that. Its only a 8 core but has quad channel DDR3 as well as 4x QPI links

Nehalem-EX is targeted more at IBM's POWER line than anything, according to their marketing guys. They continually tout the high-availability features of Beckton and the term "RISC replacement" gets bandied about a bit. I would agree with their marketing guys as the feature set of Beckton seems much more in line with the big-iron market than the typical x86 server market. Magny-Cours is apparently being marketed more as a high-end 2P platform rather than a 4P platform, and if the pricing rumors are true, it will be priced more in line with 2P chips rather than 4P+ chips. Thus Magny-Cours will mostly compete with Westmere and not Beckton.

Will be nice to see. Oh and MU, Intel will have a 6 core for the DT market at the $500 mark, the Core i7 970.

I haven't heard any official word on the Core i7 970. Most of what I read traces back to a Fudzilla article. Fudzilla says the i7 970 is a six-core 3.2 GHz part planned for Q3 of this year, so it may be a while before Intel releases <$1000 six-core CPUs on the desktop. We'll just have to see how that plays out, and also how AMD introducing the Phenom II X6 Thuban affects Intel's plans. The Phenom II X6 certainly won't beat the Gulftown, as the comparison is basically a Deneb with two more cores versus a die-shrunk Bloomfield with two more cores and 50% more cache. If you take overclocking into consideration, then it especially won't as a guy on another forum with an EVGA W555 and two Westmere 6-core Xeons got them to about 4.2 GHz on air, and I believe his Westmeres were even engineering samples. Denebs generally top out right around 4.0 GHz, and I doubt the six-core variants on a similar process will do a whole lot better, even if they are a new stepping. However, the Thubans may very well clean up against the Bloomfields in thread-heavy applications and pressure Intel to release something like the i7 970 earlier than the rumored Q3 to fill the performance gap between the Bloomfields and the i7 980X. That also remains to be seen.

@fazers_on_stun:
Anandtech is making a different argument than I am. They say that companies that spend $100 million on running SAP won't flinch if they have to pay somewhat higher hardware costs for faster hardware. I won't dispute that. What I was saying is that buyers won't want to spend more on hardware to get similar performance. Even if you are dropping $100M to run a single application, no accountant is going to turn down saving a few hundred to a few thousand dollars per server to get similar performance. A hundred here, a thousand there, and pretty soon it adds up to real money.

 

[jimmysmitty]

^I meant Beckton to compete in terms of actual tech, such as quad channel DDR3. We don't know for sure really since Intel has not let very much info out about Beckton beyond the specs. Still a 4 x QPI would be devestating to anything AMD has if Intel released it in 1P and up, so maybe they are just trying to keep that from happening and avoid anymore monopoly issues with AMD.

As for the 970, I would love to see AMD push Intel to release more than one 6 core variant at near $300 pricing. That would be a dream to have a 6 core that cheap but it will probably take time to see that and by then, 8 cores will become high end.

That or if Intel just goes for home and pushes out something based on their many core arch such as that 48 core one set for Cloud computing or even Terascale....

 

[fazers_on_stun]

@fazers_on_stun:
Anandtech is making a different argument than I am. They say that companies that spend $100 million on running SAP won't flinch if they have to pay somewhat higher hardware costs for faster hardware. I won't dispute that. What I was saying is that buyers won't want to spend more on hardware to get similar performance. Even if you are dropping $100M to run a single application, no accountant is going to turn down saving a few hundred to a few thousand dollars per server to get similar performance. A hundred here, a thousand there, and pretty soon it adds up to real money.

What you posted earlier was "rumor has it that the lower-end Magny-Cours chips will be far less expensive than their current 4P-capable server CPUs. If it takes a $1000 Xeon to beat a $500 Magny-Cours (I'm just throwing out numbers to make an example), guess who will sell more chips. "

What I've found on MC pricing :

8-core models

•6124 HE, 1.8GHz, 65W ACP, $529
•6128 HE, 2.0GHz, 65W ACP, $599
•6128, 2.0GHz, 80W ACP, $309
•6134, 2.3GHz, 80W ACP, $599
•6136 2.4GHz, 80W ACP, $849
12-core models

•6164 HE, 1.7GHz, 67W ACP, $879
•6168, 1.9GHz, 80W ACP, $849
•6172, 2.1GHz, 80W ACP, $1,149
•6174, 2.2GHz, 80W ACP, $1,349
•6176 SE, 2.3GHz, 105W ACP, $1,599

According to the AT article it might take 24 Opteron cores to come very close to 12 Westmere cores on SAP, to paraphrase. Looking at the supposed 12-core MC prices, my guess is that it would be the 2.2 or 2.3GHz parts to compete with the 3.3GHz Westmere Xeon part, which according to AT costs $1663. So the actual price difference for 2 CPUs seems to be either $628 for the 2.2GHz MC or $128 for the 2.3GHz MC. And comparing the 95W Xeon 5670 2.9GHz part, with a price of $1440, it might actually be cheaper to buy Intel. At least for SAP.

 

[MU_Engineer]

What I've found on MC pricing :

8-core models

•6124 HE, 1.8GHz, 65W ACP, $529
•6128 HE, 2.0GHz, 65W ACP, $599
•6128, 2.0GHz, 80W ACP, $309
•6134, 2.3GHz, 80W ACP, $599
•6136 2.4GHz, 80W ACP, $849
12-core models

•6164 HE, 1.7GHz, 67W ACP, $879
•6168, 1.9GHz, 80W ACP, $849
•6172, 2.1GHz, 80W ACP, $1,149
•6174, 2.2GHz, 80W ACP, $1,349
•6176 SE, 2.3GHz, 105W ACP, $1,599

That information is also apparently incorrect. We only know one price for sure, which is that the 6172 is supposed to be $989. That was buried in the fine print of the Analyst Day slides.

According to the AT article it might take 24 Opteron cores to come very close to 12 Westmere cores on SAP, to paraphrase. Looking at the supposed 12-core MC prices, my guess is that it would be the 2.2 or 2.3GHz parts to compete with the 3.3GHz Westmere Xeon part, which according to AT costs $1663. So the actual price difference for 2 CPUs seems to be either $628 for the 2.2GHz MC or $128 for the 2.3GHz MC. And comparing the 95W Xeon 5670 2.9GHz part, with a price of $1440, it might actually be cheaper to buy Intel. At least for SAP.

Sure, maybe the Xeons will be faster in SAP. That is entirely possible. Some programs just run better on certain CPUs, be it because of a working set that fits into one CPU's cache and not another, structure or optimizations that take more advantage of a certain CPU's features over another CPU's features, or using a compiler that compiles code that runs much better on one CPU than another. But other programs run faster on other CPUs. If you want to really get into depth about AT's results and assume rightly or wrongly that a dual 12-core Magny-Cours setup performs roughly similarly to a quad Istanbul 8435 setup, there are several tests that a dual MC setup would beat a pair of Xeon X5670s or X5680s:

1. Oracle: There are no quad Istanbul figures. Going from one X5670 to two X5670s is a 31% performance increase, so either Oracle isn't all that well threaded or there is some sort of bottleneck somewhere. Where the bottleneck exists could make a huge difference in the performance of Xeons vs. Opterons, so trying to estimate the scaling from two Istanbuls to two MCs would be a real shot in the dark. I'll defer on this one.
2. SAP: Four 2.6 GHz Opteron 8435s are beaten by about 10% by two Xeon X5680s. AMD would have to pull about a 20% per-clock boost from the increased HT bandwidth and DDR3 bandwidth to tie the Xeons, considering a top-bin launch speed in the low 2 GHz range. I'd probably say the Xeons will edge out MC on this one.
3. MS SQL Server: Four Opteron 8435s beat two Xeon X5670s by nearly 20%. MCs should also be faster than the Xeons here.
4. VMMark: It's a virtual dead heat between dual X5670s and quad 8435s. The 10% higher-clocked Xeon X5680 would beat the 8435s, but if it would beat the top MCs is too close to call. Maybe it would, but again it will depend on MC's actual clock speeds and performance improvements over the Istanbuls.
5. vApus ESX: Quad 8435s are 7% faster than two X5670s. X5680s vs. MCs would again be too close to predict.
6. vApus Hyper-V: Dual Xeon X5670s are 15% faster than dual Opteron 2435s. No quad Istanbul figures. Scaling from dual X5570s to dual X5670s gives a 30% boost, so the program appears well-threaded and dual MCs should be significantly faster than any of the Xeons.

So at the end of the day, it looks like MCs will be clearly faster in MS SQL Server and Hyper-V virtualization. The Xeons will likely remain a touch faster in SAP. VMMark and ESX virtualization appear very close and it could go either way. I can't even begin to predict for Oracle. It looks like the Xeons may go from the current situation of kicking around any AMD 2P system to tying or losing to them in most applications if my estimation is anywhere near accurate. Only time will really tell, but it certainly looks like AMD is going to become much more competitive after MC launches.