Mac Pro "Sandy Bridge-E" Performance Analysis

[Amethyst]

Sandy Bridge-E updates from VR-zone.com http://vr-zone.com/articles/what-intel-won-t-tell-you-about-sandy-bridge-e-at-idf/13552.html

1) two SATA 6Gbps ports and four SATA 3Gbps ports.
2) Intel won't even guarantee PCI Express 3.0 compliance
3) Quad-Core Model i7-3820 will delay, as it'll only go into production at the very end of 2011
4) Ivy Bridge-E isn't out until 2013

Bad news, Very baddd.

 

[Hellhammer]

In your other thread, the TH link suggested no PCIe 3.0 support (they actually had some hardware in-hand).

 

[derbothaus]

So they can't figure out how to get 6x SATA3 ports working? SB-E can't handle the PCIe 3.0 specs power? C'mon...

 

[Inconsequential]

Sandy Bridge-E updates from VR-zone.com http://vr-zone.com/articles/what-intel-won-t-tell-you-about-sandy-bridge-e-at-idf/13552.html

1) two SATA 6Gbps ports and four SATA 3Gbps ports.
2) Intel won't even guarantee PCI Express 3.0 compliance
3) Quad-Core Model i7-3820 will delay, as it'll only go into production at the very end of 2011
4) Ivy Bridge-E isn't out until 2013

Bad news, Very baddd.

My take:
1. Meh, what is more important here is that the 660MB/sec limit of ALL the SATA ports are lifted. 95% of people only have 1/2 SSDs so that is fine for the 6Gbps (not that I care anyway) and then the rest for HDs. Hope and pray for 1000MB/sec+ this time round eh?

2. Not that bothered really, I mean, current GPUs only barely start to hit the PCIe 2 limit so is it *really* worth worrying about?

3. Don't care, want six c0res
4. Well, what do you expect? SB-E then IB-E immediately after?

2.5 years out of my Mac Pro is a good run, IMO.

 

[nanofrog]

Yes. core on a single die are count. but it can't match a number from dual processor system. so that way we can completely divide an iMac from mac pro. (e.g. on next gen 6-Core iMac Mac pro user will get 12 Core as based.)

If all you're interested in is raw core count, sure.

But the issue is with current software, most of it cannot utilize all of the cores which means most of the time, they sit idle and not benefiting the user at all.

For specific applications, and assuming there's sufficient work to keep busy (make the DP model financially viable), then more cores than a SP model can provide makes sense. But this isn't all that common (i.e. spend 90%+ of their time in an application that is true n core multi-threaded).

Sandy Bridge-E updates from VR-zone.com http://vr-zone.com/articles/what-intel-won-t-tell-you-about-sandy-bridge-e-at-idf/13552.html

1) two SATA 6Gbps ports and four SATA 3Gbps ports.
2) Intel won't even guarantee PCI Express 3.0 compliance
3) Quad-Core Model i7-3820 will delay, as it'll only go into production at the very end of 2011
4) Ivy Bridge-E isn't out until 2013

Bad news, Very baddd.

The must have had some serious issues with the chipset, as even Q4 is actually a delay, and to make it, they only appear to be offering the low end P/N's for the chipset. And quite a disappointment (appears it will be the Z68 situation all over again; performance chipset SKU's shows up late to the party).

As per the PCIe Gen 3.0, it's my understanding they didn't have much in the way of cards to perform validation with. So they've carefully worded information in order to temper users' expectations ("weasel clause" because they have no idea if future PCIe 3.0 cards will work or not). Go figure.

 

[Amethyst]

If all you're interested in is raw core count, sure.

But the issue is with current software, most of it cannot utilize all of the cores which means most of the time, they sit idle and not benefiting the user at all.

For specific applications, and assuming there's sufficient work to keep busy (make the DP model financially viable), then more cores than a SP model can provide makes sense. But this isn't all that common (i.e. spend 90%+ of their time in an application that is true n core multi-threaded).

Yes, In this situation ;Same arch (Sandy bridge), Same base core number (e.g. 4c/8t on i7 iMac, 4c/8t on Entry-level mac pro) I really care about core count. Ok most apps aren't optimized for more than 4 core. But as we learn (other threads) the people who are main mac pro buyer can easily max out 12 core. That thing important. When you launch a product. you must launch something that point to your market. And SP mac pro is point to nowhere. (It doesn't cheap, It doesn't faster much more than Top-of-the-line iMac (Unless you pay $3699 for 6-core model)).


All of the line are only IMO.

 

[Umbongo]

Yes, In this situation ;Same arch (Sandy bridge), Same base core number (e.g. 4c/8t on i7 iMac, 4c/8t on Entry-level mac pro) I really care about core count. Ok most apps aren't optimized for more than 4 core. But as we learn (other threads) the people who are main mac pro buyer can easily max out 12 core. That thing important. When you launch a product. you must launch something that point to your market. And SP mac pro is point to nowhere. (It doesn't cheap, It doesn't faster much more than Top-of-the-line iMac (Unless you pay $3699 for 6-core model)).


All of the line are only IMO.

UP workstation units make up a considerable portion of PC workstation market, why do you think Apple are so different? Yes they are expensive, but all the Mac Pros have a similar Apple premium of $1,000. Processor power is not the only consideration. Memory capacity and bandwidth, I/O performance, powerful graphics, internal RAID, multiple display support/graphics cards, expansion cards, noise output, reliability; these are all things that a user may want without needing 12, slower, cores or the added expense of it.

 

[deconstruct60]

Yes. core on a single die are count. but it can't match a number from dual processor system. so that way we can completely divide an iMac from mac pro. (e.g. on next gen 6-Core iMac Mac pro user will get 12 Core as based.)

Eh? They are all the same die. The "SP" versions have some of the 8 cores flipped off. One some of the "DP" versions all 8 are flipped on . If targeted toward "GHz" jockeys then Intel permanently flips off at least half of the cores.

 

[Tanax]

Sandy Bridge-E will be able to handle PCIe 3.0.

Socket B2 = LGA1356
Socket R = LGA2011

It also depends on which chipset you get. Patsburg(aka X79) will come in 4 different versions; Patsburg -A, -B, -D and -T and they'll all have different amount of functionality.

At least if I understood everything correctly. Someone could perhaps verify this.

 

[Eidorian]

It is PCI-Express 2.0 for now. The dedicated storage link to the X79 PCH is another victim along the way. It is SATA only and not SAS as mentioned in that older slide. Compare it to P67.

 

[Hellhammer]

Sandy Bridge-E will be able to handle PCIe 3.0.
Image

Socket B2 = LGA1356
Socket R = LGA2011

It also depends on which chipset you get. Patsburg(aka X79) will come in 4 different versions; Patsburg -A, -B, -D and -T and they'll all have different amount of functionality.

At least if I understood everything correctly. Someone could perhaps verify this.

Earlier reports suggested PCIe 3.0 but looks like Intel got into problems and decided to stay with PCIe 2.0. After all, SB-E is aimed at servers mainly so there is absolutely no room for errors.

The chipset shouldn't play any role because the PCIe controller is in the CPU. The only way would be that the PCIe 3.0 support is there in the CPU but motherboards are limited to 2.0, though that doesn't sound too likely IMO.

 

[deconstruct60]

Sandy Bridge-E updates from VR-zone.com http://vr-zone.com/articles/what-intel-won-t-tell-you-about-sandy-bridge-e-at-idf/13552.html

1) two SATA 6Gbps ports and four SATA 3Gbps ports.

Apple quite likely was going to pick C600 Platsburg-A anyway. The fact that X79 slid back to being the A version would not have any impact on Apple's plans. The "ultra" , "extreme" board hackintosh builders perhaps.

For example, if Apple changed the two Optical bays into hot swap SSD bays they would only need two 6Gbps connections. The number of folks who need a 4 drive SSD set up are few and far between. Two SSD and four HDDs (presuming only minor tweak to current 4 drive sleds) would be plenty for the vast majority if Mac Pro users. So that is likely what Apple had already picked.


A DMI interface wasn't going to hold up to 4, 6, or 8 6Gps lanes. It was always a bit of an inbalanced I/O skew that would place too many high Gbps lanes into the limit I/O of the design. I doubt Apple "bought into" the whole "extreme" feature war design approach.

The workaround to use CPU package PCI-e v3.0 lanes to unblock the bottleneck is a bit odd. A separate SATA III controller with all the fancy RAID features work fine. Furthermore, if adding something like Thunderbolt to the C600's PCI-e v2.0 lanes was only going to help limit the DMI bandwidth too.

2) Intel won't even guarantee PCI Express 3.0 compliance

Before the board vendors ship out 3.0 cards most of these comparison/rumor/hobbyist sites are all going to go bake-offs that typically include using the same cards.

Until October I don't anyone large trust other folks to have done their homework right. I think the 3.0 "official" compliance test were suppose to come online till late September , early October timeframe anyway.

"...The PCI SIG will publish a list of compliant products starting in the fall of 2011. ... "
http://www.eetimes.com/electronics-news/4200588/PCI-Express-3-0-delayed-until-2011?pageNumber=1 ( a story from over a year ago. I'm not sure why folks are rolling around on the ground talking PCI-e 3.0 doom and gloom when announced a long time ago there wouldn't be a official until later this year. )

This should have cleared up by a mid-November launch.

3) Quad-Core Model i7-3820 will delay, as it'll only go into production at the very end of 2011

That seems to match up with the rumors about Xeon E5 sliding out longer than the. Or that the E5 are being swapped into the production lines and pushing out the 3820. That would be good news (for Mac Pro folks).

4) Ivy Bridge-E isn't out until 2013

Well, given that the first Ivy Bridge launch looks to be sliding out till Feb-April that will push back the Xeon launch.

 

[Tanax]

Earlier reports suggested PCIe 3.0 but looks like Intel got into problems and decided to stay with PCIe 2.0. After all, SB-E is aimed at servers mainly so there is absolutely no room for errors.

The chipset shouldn't play any role because the PCIe controller is in the CPU. The only way would be that the PCIe 3.0 support is there in the CPU but motherboards are limited to 2.0, though that doesn't sound too likely IMO.

As far as I understood it, LGA1356 is the new LGA1366 which will be for enthusiasts(pro-sumers) while LGA2011 is what we would want in the new Mac Pro since that's the socket for servers and workstations.

No, Im sorry you must've misunderstood me. I meant that the socket plays a role in wheather or not PCIe 3.0 is supported or not. LGA1356 had problems like you said but they wanted to get it out for the enthusiasts and then "make it right" with LGA2011 which is for servers and workstations.

At least if I understood it correctly. All these platforms, sockets, processing-architecture and naming-conventions are so damn hard to understand

 

[Eidorian]

As far as I understood it, LGA1356 is the new LGA1366 which will be for enthusiasts(pro-sumers) while LGA2011 is what we would want in the new Mac Pro since that's the socket for servers and workstations.

The "consumer" X79 boards have Socket 2011.

 

[Tanax]

The "consumer" X79 boards have Socket 2011.

Then I'm at a loss. What would 1356 be good for?

 

[deconstruct60]

The chipset shouldn't play any role because the PCIe controller is in the CPU. The only way would be that the PCIe 3.0 support is there in the CPU but motherboards are limited to 2.0, though that doesn't sound too likely IMO.

In part it will. That's because most of these "extreme" boards all have PCI-e switches on them. Unless there is a good, inexpensive set of PCI-e 3.0 switches out there that is going to be a kink in the plans for the designs that try to max out the number of PCI-e slots. Apple doesn't really have that problem; limits themselves to 4. (although do currently use a switch) But that's about what is natively supported without layering switches (the 40 lanes makes 4 physical and 1 for embedded features relatively easy to do 16-8-8-4-4 ).

 

[Eidorian]

Then I'm at a loss. What would 1356 be good for?

LOL DUNNO

On a more serious note, in the earlier Sandy Bridge-E threads it was surmised to be an intermediate single socket that offers more PCIe lanes than the 16 offered on Sandy Bridge-DT (LGA 1155). With the loss of the PCI-Express 3.0 support, you are going to want a hexcore compared to the Core i7 2600K and its Xeon counterpart.

Here is Tom's Hardware's take on the Core i7 3960X based on an engineering sample. LGA 1356 might not even show up at all in the end.

In part it will. That's because most of these "extreme" boards all have PCI-e switches on them. Unless there is a good, inexpensive set of PCI-e 3.0 switches out there that is going to be a kink in the plans for the designs that try to max out the number of PCI-e slots. Apple doesn't really have that problem; limits themselves to 4. (although do currently use a switch) But that's about what is natively supported without layering switches (the 40 lanes makes 4 physical and 1 for embedded features relatively easy to do 16-8-8-4-4 ).

PCI-Express 3.0 boards are out for LGA 1155. They just await Ivy Bridge. The X79 boards previewed over the past year all appear to support 40 PCI-Express 3.0 lanes and can fail back to 2.0. I do not expect a miracle from Intel in the next few months but Round 2 Sandy Bridge-E or even all the way to Ivy Bridge-E are more than likely going to provide a drop-in upgrade to get that PCIe 3.0 controller.

The appropriate switches should be there just waiting.

 

[deconstruct60]

It is SATA only and not SAS as mentioned in that older slide. Compare it to P67.

Apple chucked SAS drives from the BTO configs later summer ( or before? ).

There was nothing past the "-A" option that had Apple name written on it.

 

[Eidorian]

Apple chucked SAS drives from the BTO configs later summer ( or before? ).

There was nothing past the "-A" option that had Apple name written on it.

On-board or a PCIe slot controller card? I sadly do not follow these things as closely as I should but take a more broad view. It is not going to branch off the current ICH10R and you are going to want the bandwidth of that dedicated storage channel.

That leaves just leeching some available PCIe lanes and putting the chip on the motherboard.

 

[deconstruct60]

Then I'm at a loss. What would 1356 be good for?

I suspect the Xeon 2400 series was going to be "cheaper". They lower the limit of memory channels back down to 3.

http://www.cpu-world.com/news_2011/2011050601_Intel_Xeon_E5-2400_CPU_series_details.html


Also fewer PCI-e lanes. And lower TDP. They would good for the $1,200 mini-tower workstations that consist of having one graphics card and maybe one storage I/O card (so only really needed two slots. ). The smaller boxes would likely have smaller fans ( so lower heat budget. ).

But yeah the gap between the E3's and E5 2600's doesn't to be very large for the E5 2400's to fit into.

The single QPI link helps the 2400 get around the limited PCI-e lanes. I suspect low end 1U white boxes would use it too. If you just add two 2400 you end up with 48 PCI-e lanes; which plenty for many uses. Just wouldn't see it for boxes with high thread + high I/O issues.

Again there is no 1U Apple box nor mini-tower so not very useful here.

 

[deconstruct60]

On-board or a PCIe slot controller card?

On-board.

That leaves just leeching some available PCIe lanes and putting the chip on the motherboard.

The Patsburg -C and -D option don't say lanes. A separate PCI-e controller hooked to the CPU packages PCI-e lanes or the -C/-D versions hooked to the same PCI-e lanes doesn't "solve" any lane conservation problem. That's why I'm saying the solution is a bit dubious.

The only thing it saves is the board space for a SATA III controller. You can very marginally save space by making the core chipset bigger and adding more pins. However, the trade off making is that creates bigger bandwidth and complexity problems. Intel is trying to suck PCI-e raid cards into the core chips black hole the way GPU got sucked into the CPU package black hole. For some markets that makes sense and for others it does not.

Intel having glitches in a couple of chipsets this year may be a sign they aren't really managing the complexity growth (subsuming larger set of features into core chipset ) all that well.

 

[Eidorian]

On-board.

Thanks.

The Patsburg -C and -D option don't say lanes. A separate PCI-e controller hooked to the CPU packages PCI-e lanes or the -C/-D versions hooked to the same PCI-e lanes doesn't "solve" any lane conservation problem. That's why I'm saying the solution is a bit dubious.

The only thing it saves is the board space for a SATA III controller. You can very marginally save space by making the core chipset bigger and adding more pins. However, the trade off making is that creates bigger bandwidth and complexity problems. Intel is trying to suck PCI-e raid cards into the core chips black hole the way GPU got sucked into the CPU package black hole. For some markets that makes sense and for others it does not.

Intel having glitches in a couple of chipsets this year may be a sign they aren't really managing the complexity growth (subsuming larger set of features into core chipset ) all that well.

Correction, for the current LGA 1366/X58 derived platform not LGA 2011/Patsburg. The higher tier Patsburg options had the dedicated controller for storage.

I am somewhat surprised at the teething issues with chipsets in recent memory. At least Aladdin and SiS are not still around...

 

[nanofrog]

Yes, In this situation ;Same arch (Sandy bridge), Same base core number (e.g. 4c/8t on i7 iMac, 4c/8t on Entry-level mac pro) I really care about core count. Ok most apps aren't optimized for more than 4 core. But as we learn (other threads) the people who are main mac pro buyer can easily max out 12 core. That thing important. When you launch a product. you must launch something that point to your market. And SP mac pro is point to nowhere. (It doesn't cheap, It doesn't faster much more than Top-of-the-line iMac (Unless you pay $3699 for 6-core model)).

Your original question seem general to me, not for a small number of workstation users (all platforms, not a subset of Mac Pro users).

Since most workstation software won't be able to utilize an unlimited number of cores, an SP system with say 8 cores will be sufficient for more workstation users. There will still be a few that can utilize a DP system, but those numbers will be small compared to the entire workstation market.

Then I'm at a loss. What would 1356 be good for?

Mid-level server (need cores, but not I/O at the same performance as an LGA2011 can provide).

 

[VirtualRain]

Sandy Bridge-E updates from VR-zone.com http://vr-zone.com/articles/what-intel-won-t-tell-you-about-sandy-bridge-e-at-idf/13552.html

1) two SATA 6Gbps ports and four SATA 3Gbps ports.
2) Intel won't even guarantee PCI Express 3.0 compliance
3) Quad-Core Model i7-3820 will delay, as it'll only go into production at the very end of 2011
4) Ivy Bridge-E isn't out until 2013

Bad news, Very baddd.

Bad news indeed. Intel clearly needs to shift focus from copy and pasting cores on the lithography to their I/O subsystems. This stunted chipset will definitely reduce the appeal of SB-E systems.

 

[deconstruct60]

Bad news indeed. Intel clearly needs to shift focus from copy and pasting cores on the lithography to their I/O subsystems. This stunted chipset will definitely reduce the appeal of SB-E systems.

Stunted??? Maybe for the single core oriented "extreme" , "elite" gamer but on a budget crowd. For the workstation folks there is very little lost here. A dual package E5 2600 has 80 , yes eight zero , PCI-e lanes. Some of the current Xeon boards could throw all the embedded switches away. There is zero reason to hook to the low level I/O hub if you really have large I/O throughput issues.

The low level I/O hub is just that... a hub for the slower I/O. Turning it into a major intersection for double digit Gbps traffic is a flawed notion. The X58 went into the CPU package; not the I/O hub.


P.S. No I don't think that Apple will roll out a fully hooked up 80 lane monster dual package model. There is enough in 40 lanes to toss the switch in the current Mac Pro's oriented design though.