Demystifying intel's nehalem naming scheme for current and upcoming chips

[taltamir]

Intel plans on having 3 major sockets this generation. s1366 (available currently with X58 - high end, high end multi GPU, and servers), S1156 (will be available soon, mainstream). And s989 that is for laptops.

DESKTOPS:
an i9 is a six core s1366 chip with hyperthreading
an i7 for a desktop is defined as a quad core with hyperthreading (8 threads). s1366 and s1156
an i5 is either a quad core with no HT (4 threads) or a dual core with HT (4 threads), s1156 only
an i3 is like an i5, only with turbo mode disabled. s1156 only
Note: Laptop definitions are different than desktop definitions.

The P55 (s1156) will have an i7, which actually has some advantages (along with some disadvantages) over the X58 i7 (s1366). it is a slightly different design but it should compete well. and be less expensive.

Design differences:
X58:
CPU: Integrated memory controller and QPI link to northbridge:
Northbridge: A "northbridge" (actually its just a GPU splitter, all other northbridge functions have been integrated into the CPU) chip labeled X58, it has two seperate 16x version 2 PCIe video links (that can be split into four 8x links) and a DMI link to southbridge
Southbridge: ICH10/R

P55:
CPU: Integrated memory controller AND an integrated single 16x version 2 PCIe video link (that can be split into two 8x links) and a DMI link to southbridge
Northbridge: none, integrated into CPU
Southbridge: ICH10/R labeled "P55"

the differences between P55 i7 and X58 i7:
1. memory channel: X58 is triple channel DDR3. P55 is dual. This has nothing to do with the mobo chipset (X58 vs P55) as the memory controller has been integrated into the CPU, it is however linked to the SOCKET size (more channels means more socket pins). So do not expect a s1156 based triple channel CPU.
2. P55 will have a single full speed video card connection, or two half speed connections. X58 has two full speed connections. So if you run SLI/xfire X58 is better
3. P55 has the pciE video connection integrated into the CPU, improving latency. As long as you only use one slot (so you get full speed connection) than the P55 will actually be FASTER than the X58.
4. turbo mode: there are unconfirmed reports of P55 having greater turbo mode boosting (+5 instead of +2).

While triple memory channel is good for a high IO fileserver, it makes no difference in GAMES and regular programs.
So for a gamer: 1 makes no difference. 2&3 mean X58 is better for multiple cards, P55 is BETTER for single card. and 4, if true, means that P55 is better for video games (due to their tendency to be single or dual threaded).

LAPTOPS:
I really wish I knew anything about intel's LAPTOP naming plans, but I currently do not. If anyone wants to share such info, please do and I will add it to this post.

http://img32.imageshack.us/img32/2179/capturexzq.png

 

[ilkhan]

the mobile socket is 989-pin, you know. The rest...:shrug:

 

[taltamir]

thank you

 

[Denithor]

Originally posted by: taltamir
2. P55 will have a single full speed video card connection, or two half speed connections. X58 has two full speed connections. So if you run SLI/xfire X58 is better
3. P55 has the pciE video connection integrated into the CPU, improving latency. As long as you only use one slot (so you get full speed connection) than the P55 will actually be FASTER than the X58.

I've been asking about this for a while now, maybe you know.

P55 offers improved latency from CPU to GPU core(s).
X58 offers higher maximum bandwidth to GPU core(s).

How much does the x8 bandwidth really limit today's GPUs? Will the reduced latency be enough to overcome the lower bandwidth of a multiGPU setup on P55?

 

[ilkhan]

look at P45 vs X48 to see the effects of x8/x8 SLI vs x16/x16 SLI.

 

[Denithor]

Originally posted by: ilkhan
look at P45 vs X48 to see the effects of x8/x8 SLI vs x16/x16 SLI.

Not really, no - the Nehalem architecture brings a whole new level of multiGPU performance to the table. Seems to improve throughput drastically versus the older P45/X48 boards.

And I just found this tidbit in the recent AT article on Gigabyte's P55-UD6:

The Lynnfield processor series supports 16 PCIe 2.0 lanes and those lanes are assigned for graphics duty. The first x16 slot normally runs with 16 lanes and if you install a multi-GPU setup for CrossFire or SLI, then both slots run at x8. Truthfully, except for trying to set a benchmark record, the performance differences between dual x16 and dual x8 PCIe 2.0 is minimal at best in actual applications.

I demand benchmarks!

 

[taltamir]

Originally posted by: ilkhan
look at P45 vs X48 to see the effects of x8/x8 SLI vs x16/x16 SLI.

that doesn't take into account the 8x/8x having much lower latency (one third i think) than the 16x/16x configuartion. His question is which is more significant, the latency or bandwidth...

From looking at ram... DDR3 latency is equavalent to twice the latency (equal in ms not clock values, which is what improves in P55) of DDR2 due to measurement quirks AFAIK.
So 1600mhz DDR3 @ 6-6-6-18 should be equal latency and double bandwidth of 800mhz DDR2 @ 3-3-3-9
Since we are comparing double bandwith with triple the latency we should further modify those figures and say it is like comparing DDR2 @ 1-1-1-3 to DDR3 @ 6-6-6-18
I have absolutely no idea how they will compare to each other.

 

[Denithor]

Originally posted by: taltamir
I have absolutely no idea how they will compare to each other.

You and me both!

But apparently Gary Key knows the answer based on his little blurb there that I quoted above. I just want to see some benchies for validation.

I have an ulterior motive here - I'm in the position of having two GTX 260/216 cards on a P5Q Pro - strictly for folding purposes at this point (obviously no SLI going on with that board). I want to know if i5 SLI mode is neutered versus i7 or if it's a viable build.

Maybe one of the insider admins can bring this to Gary's attention?

 

[Idontcare]

taltamir, take a look at this Goto-san article that just came out reguarding i7/i5/i3 branding vs. features. Includes mobile stuff too.

Original (Japanese)

Google Translated (English)

In particular, I thought this graphic was pretty snazzy.

 

[taltamir]

interesting... those laptop pieces are all called i5 and i7, and don't seem to really match anythin in desktop sector, their ram speeds vary greatly, their L3 variest greatly, and the amount of cores and HT Statues is not even disclosed.
In the desktop it just reiterates the same information I already have posted. which is good to get confirmation of.

 

[ilkhan]

Originally posted by: Denithor

Originally posted by: ilkhan
look at P45 vs X48 to see the effects of x8/x8 SLI vs x16/x16 SLI.

Not really, no - the Nehalem architecture brings a whole new level of multiGPU performance to the table. Seems to improve throughput drastically versus the older P45/X48 boards.

I didn't say it'd be exactly the same as P55 vs X58, BUT if you want to isolate on JUST the performance of x8/x8 vs x16/x16, you can look at P45 vs X48.

 

[aigomorla]

Originally posted by: ilkhan
look at P45 vs X48 to see the effects of x8/x8 SLI vs x16/x16 SLI.

dont apply because X48 wasnt intended to handle NVIDIA cards.

Also the NF200 throws this comparison out of the window.

My understanding was i3 was a branded C2D. :X

 

[Denithor]

Somewhat confused by one factor in that table: pricing of i5 chips. Looks like the 4 thread with HT chip (which would imply dual core) will be priced in the $28x range while the actual quad core chip (no HT) will be priced at $19x??? WTF? Do they expect a dual with HT to actually outperform a true quad? Clarkdale is 32nm while Lynnfield is 45nm, correct? So the dual HT chip will actually cost them less to produce and they're going to charge more for it? Mind-boggling if this turns out to be true.

Which is interesting but still doesn't answer my question: will the lower latency between CPU & GPU cores (from having the PCIe channels built into the CPU) be enough to offset the loss of bandwidth of being limited to x8/x8?

And another question for you gurus here: how do you expect the i5 duals with HT enabled (4 virtual cores) to stack up against AMD's X3 series?

Finally when can we expect the blinders to come off and sites to begin showing all the benchmarks they're obviously sitting on?

 

[lopri]

There will be P57 early next year as well. Apparently Intel wasn't able to polish Braidwood technology (on-board SSD) for P55 launch (or never intended to incorporate it in P55), and there is a chance that P55 will be short-lived depending on the development of P57 as well as market reception of P55.

Kinda sucks because I skipped X58 in the anticipation of P55.. Though P55 + Lynnfield combo will not disappoint for those who have not jumped on i7 wagon yet. I am likely to hold out until P57 debuts. (which will come only a few months after P55)

http://xtreview.com/images/int...p55%20chipset%2002.gif

 

[lopri]

Oh and Intel now calls its southbridge "PCH" (Platform Controller Hub), instead of ICH.

 

[Idontcare]

Originally posted by: Denithor
Finally when can we expect the blinders to come off and sites to begin showing all the benchmarks they're obviously sitting on?

Good question, we haven't even heard speculation or rumors on when NDA's are supposed to expire yet...so we could be months out on this still.

 

[Lonyo]

Originally posted by: Denithor
Somewhat confused by one factor in that table: pricing of i5 chips. Looks like the 4 thread with HT chip (which would imply dual core) will be priced in the $28x range while the actual quad core chip (no HT) will be priced at $19x??? WTF? Do they expect a dual with HT to actually outperform a true quad? Clarkdale is 32nm while Lynnfield is 45nm, correct? So the dual HT chip will actually cost them less to produce and they're going to charge more for it? Mind-boggling if this turns out to be true.

If it's 32nm and 2010 maybe we're getting into the realm of the integrated GPU chips? So Core i5 Clarkdale is dual core/HT + GPU on board for $28x and Lynnfield is obviously just CPU+NB/etc, but then that's a large premium for an IGP which probably won't be up to much, so that seems weird as well.

The other possibility is that they simply got the prices mixed up and they should be reversed.

 

[ilkhan]

We'll see on *die* GPUs when sandy bridge hits, not before then.
Clarkdale is 32nm dual core with GPU, lynnfield is 45nm quad core without GPU.

This is one of those strange generations when each variant of the chip has nothing in common but the processing core itself.

 

[taltamir]

Originally posted by: Lonyo

Originally posted by: Denithor
Somewhat confused by one factor in that table: pricing of i5 chips. Looks like the 4 thread with HT chip (which would imply dual core) will be priced in the $28x range while the actual quad core chip (no HT) will be priced at $19x??? WTF? Do they expect a dual with HT to actually outperform a true quad? Clarkdale is 32nm while Lynnfield is 45nm, correct? So the dual HT chip will actually cost them less to produce and they're going to charge more for it? Mind-boggling if this turns out to be true.

If it's 32nm and 2010 maybe we're getting into the realm of the integrated GPU chips? So Core i5 Clarkdale is dual core/HT + GPU on board for $28x and Lynnfield is obviously just CPU+NB/etc, but then that's a large premium for an IGP which probably won't be up to much, so that seems weird as well.

The other possibility is that they simply got the prices mixed up and they should be reversed.

i think the case is that the clarkdales are dual cores + HT + GPU

 

[IntelUser2000]

Originally posted by: Lonyo

If it's 32nm and 2010 maybe we're getting into the realm of the integrated GPU chips? So Core i5 Clarkdale is dual core/HT + GPU on board for $28x and Lynnfield is obviously just CPU+NB/etc, but then that's a large premium for an IGP which probably won't be up to much, so that seems weird as well.

The other possibility is that they simply got the prices mixed up and they should be reversed.

What? You don't like the 3.46GHz base clock speed for the $28x Clarkdale?? Since most applications are most of the time 4 thread versions, Hyperthreading on Clarkdale is going to do a lot here. HT still seems to leave a bad taste on some folks, but unlike P4 its going to do much better.

Plus its a Westmere core so add a few % per clock in performance over the Nehalem derivative Lynnfield.

Conclusion is that a 3.46GHz Clarkdale is going to be fairly close to a 2.66GHz Lynnfield without HT.

 

[Denithor]

HT didn't leave a bad taste in my mouth - just the opposite in fact - I loved my P4C 2.4GHz Northwood.

But I think your reasoning is somewhat flawed.

Yes, HT allows a chip to process work in a much smoother flow (both cores will stay pegged at 100% instead of rising and falling like cores without HT behave). But in CPU-intensive duties (video encoding, for example) two cores clocked slightly higher running 100% are still not going to match the raw processing power of four full cores.

I think AMD is going to be the big loser here - I think these dual HT chips will be able to beat the X3 series in most cases, giving Intel a distinct pricing advantage (32nm dual has to be cheaper to produce than a 45nm triple [quad], right?).

 

[Idontcare]

Originally posted by: Denithor
HT didn't leave a bad taste in my mouth - just the opposite in fact - I loved my P4C 2.4GHz Northwood.

Same here, albeit a 2.8GHz P4C. Served me well until I jumped to kentsfield.

Originally posted by: Denithor
But I think your reasoning is somewhat flawed.

Remember he's trying to explain Intel's reasoning for creating the market segmentation that Intel is doing, at best you should say Intel's reasoning is flawed. I thought Inteluser2000 made a fair attempt to rationalize what Intel's marketing dept thinks it is going to do come Q4.

Originally posted by: Denithor
Yes, HT allows a chip to process work in a much smoother flow (both cores will stay pegged at 100% instead of rising and falling like cores without HT behave). But in CPU-intensive duties (video encoding, for example) two cores clocked slightly higher running 100% are still not going to match the raw processing power of four full cores.

I agree, only 2+2=4 after all, but the distinction to be made here is that 2cores + HT at higher clockspeed won't match the processing power of four otherwise identical cores.

Who is to say where 2 westmere cores with their architecture improvements and cache latency improvements is going to end up IPC-wise compared to the yorkfield quads and phenom II quads it will be competing with in Q4.

Given Intel's track record since 2006 I am actually willing to, at this point, give Intel the benefit of the doubt and assume they did their homework and they have good reason to believe the performance/price aspects of clarkdale are going to be compelling enough to warrant the creation of a product with the precise specs that clarkdale has been created to meet.

Originally posted by: Denithor
I think AMD is going to be the big loser here - I think these dual HT chips will be able to beat the X3 series in most cases, giving Intel a distinct pricing advantage (32nm dual has to be cheaper to produce than a 45nm triple [quad], right?).

Bingo! Clarkdale, if clocked to the right GHz, will give even the lower GHz X4's a run for their money as well. But the real killer here is the diesize, weighing in at around 80mm^2 the dual-core westmere chip is just going to be cheap as chips to manufacture so gross margins are going to be delicious for Intel's shareholders.

When it comes to IPC/mm^2 AMD's K10.5 architecture on 45nm is really going to struggle against westmere on 32nm. 2010 is going to be a dreadful year for AMD's financials IMO as they compete 45nm against 32nm.

 

[Denithor]

Originally posted by: Idontcare

Originally posted by: Denithor
HT didn't leave a bad taste in my mouth - just the opposite in fact - I loved my P4C 2.4GHz Northwood.

Same here, albeit a 2.8GHz P4C. Served me well until I jumped to kentsfield.

Great chips, weren't they?

I agree, only 2+2=4 after all, but the distinction to be made here is that 2cores + HT at higher clockspeed won't match the processing power of four otherwise identical cores.

Who is to say where 2 westmere cores with their architecture improvements and cache latency improvements is going to end up IPC-wise compared to the yorkfield quads and phenom II quads it will be competing with in Q4.

Given Intel's track record since 2006 I am actually willing to, at this point, give Intel the benefit of the doubt and assume they did their homework and they have good reason to believe the performance/price aspects of clarkdale are going to be compelling enough to warrant the creation of a product with the precise specs that clarkdale has been created to meet.

But we haven't typically seen this much improvement in die-shrink transitions. Consider Penryn versus Conroe - we saw roughly 3-5% improvement at the same clockspeed for this transition. If you consider the overall picture here:

3.46/2.66 = 1.3 x 2 = 2.6

So the two Clarkdale cores at higher speed is theoretically equal to 2.6 Lynnfield cores. Allow even 10% improvement for die shrink (2.86) and you're still hitting below the point of three slower physical cores. Unless HT can improve performance by 40% there's simply no way that two cores can match four.

EDIT: I think actually you have to look at the HT in reverse of what I said here. A quad without HT will run at say 80% efficiency (90%? 98%? No idea what it actually is) when under load - OS juggles threads onto the different cores and there is some lag created here. A dual with HT on the other hand has four virtual cores - the OS still distributes threads to each but in essence a queue is created for each physical core and work flows much more efficiently through as a result. So if you assume 80% efficiency for a nonHT quad you have basically 3.2 cores at 100% versus the 2.86 cores I calculated above for the dual (best case and assuming 100% efficiency from HT). That would narrow the gap significantly and you'd get similar performance at a much lower power load. However - and this is a huge however - the 80% efficiency may be completely wrong. If it's 90% or even higher for the nonHT quad the comparison breaks completely and the true quad will easily stomp the HT dual in any/all multithreaded apps.

A couple of question about the integrated GPU. If these chips contain that feature will there be a way to disable it if you run discrete graphics? Otherwise it seems like it would simply increase the power consumption and thermal generation for zero purpose. Also - I suppose we will be able to purchase like a "G55" board that features DVI/HDMI that will be driven by the integrated GPU, right? This one feature is kinda confusing - used to be if you wanted integrated graphics you just chose a board that had that built in - now you will have to pick out both a board and specific CPU with the correct support.

 

[taltamir]

I would imagine you can disable it, considering intel already has full core shutdown implemented.
It will also be MCM. the cpu die will have a built in QPI that connects to a GPU on the same socket but seperate die, and a DMI that goes through the socket to connect to the southbridge.

 

[taltamir]

got a new chart, it shows the lynfield will be +5 while current is +2...

so looking at the same priced i7 for each socket:
S1366 i7: 2.66ghz with 2.93ghz turbo and 130 watt TDP
S1156 i7: 2.80ghz with 3.46ghz turbo and 95 watt TDP