AMD shows off six-core Istanbul
AMD HAS SHOWN that it can recover from the Barcelona disaster with Shanghai and the consumer variants thereof. To prove it is not a one-off fluke, the company was showing off working Istanbul six-core silicon a few days ago.
It had three demos to show off, two with four sockets, the last with two. Each one showed off a different feature of the new CPUs, cores, virtualisation and compatibility. Sadly, no speeds, prices, or availability were discussed.
Count the cores, 4x6 = 24
The first demo was the core count. What can you say? Four sockets each running a six-core Istanbul CPU is 24 cores, enough to drive a scheduler mad. The system it was running on is an internal AMD development chassis nicknamed 'Warthog' because it is ugly but functional, and the program is an internal stress-testing tool. Keeping 24 cores nearly at full boil isn't exactly easy.
Four OSes on 12 cores
Next we moved on to a virtualisation demo on a two-socket desktop system. Once again, nothing shocking, just that the silicon is up, running, and works at least as well as Shanghai for virtualisation.
What you are seeing is three hosted OSes on a Windows Server 2008 box running the almost-ready-for-prime-time (if you ask someone from MS that is) Hyper-V. There is a copy of Server 2003 in the top left, then Redhat, SLES and lastly the native OS. In case you can't tell, it works.
Board swaps are functional
Last up we had an interoperability demo that turned out to be much more. On the surface, they took two 4S systems, one with Shanghai's and the other with Istanbul boards, and swapped them. If you simply upgrade the BIOS of any S1207 board, assuming they followed the rules and put in split voltage planes, it should work just fine with 6 cores. Istanbul needs split voltage planes, Shanghai didn't require it.
The AMD demo crew ran stream on the machines, and the Shanghai system scored much lower than the Istanbul box. No shock there. Then they powered down the machines, swapped the CPU daughter boards, and powered them up. It went smoothly, and the scores reflected the new CPUs perfectly.
Shanghai scores about 25K
Istanbul scores a bunch more
Swapping is all fine and dandy, but look at the scores for a sec. Stream is basically a bandwidth tester, and Istanbul scores 1.6x Shanghai on this test, a massive increase. Keep in mind that the only change between the machines is the CPU, the bandwidth should not have changed much. You could say that this is a best case scenario for threading, but there is more to it, the secret weapon of Istanbul is called a Probe Filter (PF).
Probe Filter is the AMD name for a snoop filter, and what it does it to limit the number of times the CPU has to wait while a remote core returns cache data. This cache-induced latency is what killed the 8S Opteron market, the latency really hammered that machine.
Now that AMD has a snoop filter (Intel had it since the late P4 days with the game-changing Blackford chipset) it is seeing the same benefits. The PF has three modes, 'no probe needed', 'directed', and 'broadcast'. The old way was all broadcast, and you can see the difference the filter makes on a HT3 system, it gets you about 50 per cent more effective bandwidth.
The way it works is simple, they take a portion of the L3 cache and dedicate it to keeping track of what is in the other CPUs caches. This way, the cache snoop is simply an L3 lookup in many cases, a directed request in a lot of others, and worst case, the same as it was before. From the numbers, it is easy to see that it is a clear win, even if it is in the best case scenario for the test.
Other than that, Istanbul is pretty damn close to Shanghai, or in the words of one AMDer, "about as close to cut and paste as you can get". The differences are minor and mostly to the uncore. The only one that is worth mentioning is Remote Management Interface (RMI). This is exposing some APMC (server external management controller) data to the system so you can look at it from outside the box. It is a first step, and starting with Socket G34 in a few months, you will see more and more of this type of data exposed.
Overall, Istanbul is more than 1.5 Shanghais, but not much more. The PF is a huge win, the rest is just more of the same. That said, it is currently very healthy silicon and upgradable with a BIOS flash on almost any S1207 box.
What's not to like? µ
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
"The differences are minor and mostly to the uncore. The only one that is worth mentioning is Remote Management Interface (RMI)."
Crikey Crumbs, Batman/Charlie, since when has Remote Management ever been a minor difference as opposed to a major facility?
Is that Enhancement/Faculty as a Result of the Fab Foundry Company move with Mubadala and the Advanced Technology Investment Company (ATIC)?
And would any reply to that question be purely speculative comment on Proprietary Intellectual Property Issues and/or Potential Investment Opportunities in Virtual Markets/ZerodDay Trading Places and Spaces?
So, I understand that a little thingie like "Probe Filter" actually gets 60% MORE bandwidth??
That means that AMD had a 97% awesome processor 2 years ago :) ... but with some little missing features that actually made it a cripple!
And now they should be on the right track. I have one question: will they include Probe Filter in any version of Phenom 2 processors?
When I saw a "fully loaded" task manager demo... my gut reaction was, yeah I remember the Barcelona "task manager demo"...
Well this time they actually showed some performance #'s which just goes to show what a SCAM that first Barcelona demo was. There was all this talk about not "letting Intel see how good it was", yet this time around they show data (I guess the secrecy is no longer a concern). Perhaps (and this is the cynic in me) it has more to do with the fact that they actually have something respectable to show this time, as opposed to the Barcy demo.
It's good to see AMD come around to actually showing some #'s when doing a demo (as opposed to the hypocracy of Barcelona). The 1.6X conclusion - that is comparing chips of similar CPU speeds? Were the speeds of the respective systems disclosed? (I really hope they did not use an intentionally low speed quad core chip!)
"will they include Probe Filter in any version of Phenom 2 processors?"
No, since the problem doesn't exist on a single socket system.
... AMD: the GM of CPU. Or is it Chrysler, with Intel being the GM of CPU?
Gawd, I hope not, Doug. Both GM and Chrysler are both highly mis-managed groups of short-sighted, inbred idiots who should be allowed to go bankrupt and disappear. I would hope Intel and AMD stick around for awhile. I hear the ChiCom CPUs are that good yet.
About time AMD got around to a snoop filter; their designs really lacked without it.
And RMI is only for anal corporate IT types anyway; the kind that just has to know what everyone is doing with "their" precious little corporate beige boxes at all times.
Thanks for the update, Charlie. It's good to see that AMD is getting their act back together. Perhaps this doffing of their fab plants will allow them to focus on bringing some real value to the process.
...does that mean that we'll see more 8S servers now? 48 cores, mmmm. I doubt it's that cost effective compared to a whole bunch of 2S and 4S servers but it's still a happy thought and would make partitioning resources in a VM server farm, the "cloud", a little easier.
Nifty! That looks very nice.
hr /
<hr/>
Forty eight cores in a single box is a different programming challenge than 48 cores spread across several boxes.
So, how will this 6 core chip help me run my single core app faster compared to the 4 core chip?
That said, it is currently very healthy silicon and upgradable with a BIOS flash on almost any S1207 box.
I'm sorry I'm not sure I buy it, or I was very unlucky to have the kind of motherboard with a S1207 socket that didn't upgrade to Barcelona. I purchased a HP xw9400 with 2210 (1.8Ghz) waiting to replace it with the Barcelona cpu when they came out. Well, 2 years later I tried to drop in 2x 2352 (@2.1Ghz). I did the BIOS upgrade, but no luck. Worse for me, the HP xw9400 where using OEM re-branded Tyan 2915 motherboards. I kinda lost confidence in Tyan and HP for such a simple upgrade...
So I'm at a dead end with this nice HP xw9400 Workstation. I did purchase a new Tyan motherboard to use the 2352, but now with Shanghai and with HT3 around the corner, it's difficult decision time... Embrace Nehalem or take one more risk on AMD ?
So, are AMD engineers dancing in the isles in their undees?
yet when they bring the same van to the USA it can only do 23mpg. Ford Transit Connect, google it for yourself.
AMD 6 cores, cool. Lets hope this takes the fight to Intel, so we can all win.
... If the motherboard maker made the socket s1207 board right... @Erik Bussink i got bit by the same Tyan s2915 bug. but i bought the s2915-e and it rocks with 2x2376 Opterons in it, and it should work fine with Istanbul... Tyan, for all its faults still mops the floor with Asus when it comes to server boards, its Tyan or bust...
That's 50mpg for the European *diesel* version of the Transit Connect using a 1.8-liter Duratorq engine, not the standard 2.0-liter 4-cylinder *gasoline* engine in the US version for which the 23mpg is rated for. It's also harder to directly port European diesel vehicles here as the emission standards are lower in Europe.
AMD if you're reading this. You need to release this processor for $400 on the Socket AM2 platform. If that's done, you can bet your ass you'll regain market share. Put your R&D dollars to use and give Intel some real competition. Instead of charging 8X more for something with 1/8 the product run. Make 8X more and charge 1/4 of the price. Alright give me my inflated salary now for using common sense.
Hope it will be in market fast. Intel i-7 will have to reduce their prices.
Check out Istanbul demo videos here: http://links.amd.com/IstanbulUpgrade & http://links.amd.com/IstanbulDemo
See images of the demos here: http://links.amd.com/Images
Tested Istanbul configuration is of a 4 sockets / 4 processors. 41GB/s is still on the lower end comparing with Nehalem EP 2 sockets configuration, which is about 34GB/s, considering the testing-configuration has an extra two sockets, that is extra 4 memory channels or two extra MCs.
A two socket Istanbul configuration should be on par or less comparing with a Nehalem EP 2P configuration, I think, it will be more likily end up "less".
A 2P configuration has less snooping traffic, a 2P Istanbul would not benefit from its new method of snooping as a 4P configuration, by following what is said here.
And more, the cost of snooping has to be resolved if any one wants to move from 4 cores to 6 cores, or more sockets.
Nehalem EP's "snooping" is very costly too, but its higher QPI helped a lot.
But ... anyway