Monday, October 13th 2008

Core i7 940 Review Shows SMT and Tri-Channel Memory Let-down

As the computer enthusiast community gears up for Nehalem November, with reports suggesting a series of product launches for both Intel's Core i7 processors and compatible motherboards, Industry observer PC Online.cn have already published an in-depth review of the Core i7 940 2.93 GHz processor. The processor is based on the Bloomfield core, and essentially the Nehalem architecture that has been making news for over an year now. PC Online went right to the heart of the matter, evaluating the 192-bit wide (tri-channel) memory interface, and the advantage of HyperThreading on four physical cores. In the tests, the 2.93 GHz Bloomfield chip was pitted against a Core 2 Extreme QX9770 operating at both its reference speed of 3.20 GHz, and underclocked to 2.93 GHz, so a clock to clock comparison could be brought about.

The evaluation found that the performance increments tri-channel offers over dual-channel memory, in real world applications and games, are just about insignificant. Super Pi Mod 1.4 shows only a fractional lead for tri-channel over dual-channel, and the trend continued with Everest Memory Benchmark. On the brighter side, the integrated memory controller does offer improvements over the previous generation setup, with the northbridge handling memory. Even in games such as Call of Duty 4 and Crysis, tri-channel memory did not shine.
As for the other architectural change, simultaneous multi-threading, that makes its comeback on the desktop scene with the Bloomfield processors offering as many as eight available logical processors for the operating system to talk to, it appears to be a mixed bag, in terms of performance. The architecture did provide massive boosts in WinRAR and Cinebench tests Across tests, enabling SMT brought in performance increments of roughly 10~20% with general benchmarks that included Cinebench, WinRAR, TMPGEnc, and Fritz Chess. With 3DMark Vantage, SMT provided a very significant boost to the scores, with about 25% increments. It didn't do the same, to current generation games such as Call of Duty 4, World in Conflict and Company of Heroes. What's more, the games didn't seem to benefit from Bloomfield in the first place. The QX9770 underclocked at 2.93 GHz, outperformed i7 940, both with and without SMT, in some games.

Source: PC Online
Add your own comment

91 Comments on Core i7 940 Review Shows SMT and Tri-Channel Memory Let-down

#1
DarkMatter
FordGT90Concept said:
I think what it really boils down to is the IMC in Nehalem is over engineered and slower because of it. The designed the IMC to handle not just four cores, but eight, and maybe even more (as much as 72). I doubt we will see any significant changes to the IMC as far off as Sandybridge chips.

But yeah... maybe there is a bug in it. When AMD shifted to IMC, they got a huge memory performance boost. Intel appears to be taking a hit instead. CIS is a more complex scheme but you'd think that would show throw in more than just gaming benchmarks.

I just wonder how what Nehalem runs with FB-DIMMs.
What are you talking about? A hit? Core2 peaks at 8000 MB/s while Nehalem does 15000 MB/s, both in dual channel, where do you see a hit there?
Posted on Reply
#2
Wile E
Power User
masterbw2000 said:
I knew you'd say that.
So take a look at the results from certified ones later.
They will be exactly the same. ;)

DarkMatter said:


I also still fail to see how you could do quad pumped SDRAM, that is that each memory cell performs 4 ops per clock cycle. And I also don't understand what would be the benefit of that, versus a DDR RAM with double the speed. I.e if your memory cells can perform 1600MT/s wouldn't it be better (simpler, easy to implement, cheaper...) a DDR running at 800Mhz than a "QDR" at 400Mhz?
I don't know much about the technical side of ram, but what about GDDR5? It's rated as QDR, and as far as I was aware, it gets it's roots from SDRAM as well.

As a side note, do you think triple channel would come in handy on lower speed modules? Say, DDR3 1066 cas7?
Posted on Reply
#3
PP Mguire
DarkMatter said:
Well, I wouldn't say so. Just because Nehalem doesn't seem capable of using all the bandwidth it has unlocked by himself, that doesn't mean it's a fail. It's still way faster than Core2 clock for clock and has a lot more bandwidth even on single channel mode! SMT also seems to work very well.

It's maybe not worth it for games, but for everything else is faster enough to justify the expenses for many people (not me TBH), just as any other new CPU. Don't forget it's suposed to be aimed at the server market. Off course an upgrade from a C2Q is not worth it either, but if you were to build a completely new PC and you care about more than gaming, Nehalem is worth a look or two.
I only game so its not worth it to me to even have a quad really. 4+ghz dualy is ideal for me :rockout:
Posted on Reply
#4
FordGT90Concept
"I go fast!1!11!1!"
DarkMatter said:
What are you talking about? A hit? Core2 peaks at 8000 MB/s while Nehalem does 15000 MB/s, both in dual channel, where do you see a hit there?
Game performance which tighter timings yield better FPS than higher bandwidth. Games don't need a lot of bandwidth (only enough to satisfy the engaged cores) but, they need very quick response times.

Maybe the IMC/memory has little to do with poor Nehalem game performance which reminds me of something else. Nehalem's architecture has strong ties to Pentium 4 w/ Hyperthreading more so than Core 2 architecture. We all remember how Athlon 64 was the better gamer but Pentium 4 w/ Hyperthreading took the cake in terms of multimedia. That pretty much explains everything.
Posted on Reply
#5
e6600
right now, a Q9550 E0 should last us many years. very good price:performance chip
Posted on Reply
#6
DarkMatter
Wile E said:
They will be exactly the same. ;)


I don't know much about the technical side of ram, but what about GDDR5? It's rated as QDR, and as far as I was aware, it gets it's roots from SDRAM as well.

As a side note, do you think triple channel would come in handy on lower speed modules? Say, DDR3 1066 cas7?
I have noticed there's an insanely amount of things which are called QDR and are not exactly. In this regard Quad/Octal Pumping is a much better description to say it can do 4/8 operations per cycle. AFAIK GDDR5 uses two DDR streams that are conbined (multiplexed) to form a 2x faster memory module. Again it's not that each memory cell performs 4 ops per clock cycle, but 2 cells perform 2 ops per cycle and send the data together. In practice is almost the same as you are doubling the frequency per pin, but it's important to note that the difference with a QDR signal is evident: you don't have control to every operation being made, you only can control it by pairs. For GPUs that kind of parallelism is not a problem, but in CPUs it could create an insanely and undesired amount of latency (it already does in GPU BTW, but it's masked by the parallelism of the GPU.). In this regard is as if someone called a 2.4 Ghz Dual core CPU a 4.8 Ghz CPU. It's not the same. In the case of GDDR5 that's exactly what we do. Although as I understand it, the memory controler must work at twice the speed of the memory cells, so from that point of view it is twice as fast. It's what I said, if your memory is twice as fast (by any method), make the external clock twice as fast.

This is how I think it works, just from the clock being use perspective:



Consider the input the external clock generator.

- In DDR a spike is created for every rising and falling edge.

- QDR would create 4 spikes. I don't know how, that's what I'm being asking all the time.

- Note how GDDR5's input clock is twice as fast. A completely different thing would be if the input was 100Mhz and it was doubled inside the memory itself and not externally. That is what I said it would be pointless IMO.
Posted on Reply
#7
Hayder_Master
anyone expect more performance in games with tri channel ram , why it is go down
Posted on Reply
#8
Poisonsnak
DarkMatter said:

... This is how I think it works, just from the clock being use perspective: ...
Yeah you're definitely on th right track there. I have a bit of background in digital signaling so I know how this stuff works. Interestingly enough if you read up on AGP (yes AGP) it's the same concept. AGP = SDR, AGP2x = DDR, AGP4x = QDR, AGP8x = ODR.

Some references here:
http://en.wikipedia.org/wiki/Agp
http://en.wikipedia.org/wiki/Quadruple_data_rate

I can summarize it here though:

SDR = transmits data on the rising edge of each clock
DDR = transmits data on the rising and falling edge of each clock
QDR = uses 2 clock generators, same frequency, one is 90° ahead (or behind) of the other (e.g. clock #1 has a rising edge, then halfway before its falling edge clock #2 has a rising edge). Transmits data on the rising and falling edge of both clocks.

In practice you wouldn't actually use 2 generators but instead delay the second signal by 90° somehow but you get the idea. ODR follows the same sort of pattern except you're using 4 clocks instead of 2.

The 2 clocks 90° apart thing can be hard to visualize so if I have time I'll draw a picture later today. Another way to think of it (maybe easier) is that the falling edge of the clock could be considered to be a clock signal that is 180° behind the original clock. In that case:

SDR = 1 clock 0°
DDR = 2 clocks, 0° and 180°
QDR = 4 clocks, 0°, 90°, 180°, and 270°
ODR = 8 clocks, 0°, 45°, 90°, 135°, etc.
Posted on Reply
#9
Swansen
FordGT90Concept said:
We see this going from DDR, to DDR2, to DDR3. Ultimately, I think it's time to stop with DDR and move to something that accomplishes more per work cycle like QDR or ODR.
my vote goes for a completely different memory architecture, my choice being XDR.
Posted on Reply
#10
DarkMatter
Poisonsnak said:
Yeah you're definitely on th right track there. I have a bit of background in digital signaling so I know how this stuff works. Interestingly enough if you read up on AGP (yes AGP) it's the same concept. AGP = SDR, AGP2x = DDR, AGP4x = QDR, AGP8x = ODR.

Some references here:
http://en.wikipedia.org/wiki/Agp
http://en.wikipedia.org/wiki/Quadruple_data_rate

I can summarize it here though:

SDR = transmits data on the rising edge of each clock
DDR = transmits data on the rising and falling edge of each clock
QDR = uses 2 clock generators, same frequency, one is 90° ahead (or behind) of the other (e.g. clock #1 has a rising edge, then halfway before its falling edge clock #2 has a rising edge). Transmits data on the rising and falling edge of both clocks.

In practice you wouldn't actually use 2 generators but instead delay the second signal by 90° somehow but you get the idea. ODR follows the same sort of pattern except you're using 4 clocks instead of 2.

The 2 clocks 90° apart thing can be hard to visualize so if I have time I'll draw a picture later today. Another way to think of it (maybe easier) is that the falling edge of the clock could be considered to be a clock signal that is 180° behind the original clock. In that case:

SDR = 1 clock 0°
DDR = 2 clocks, 0° and 180°
QDR = 4 clocks, 0°, 90°, 180°, and 270°
ODR = 8 clocks, 0°, 45°, 90°, 135°, etc.
Yeah thanks, that's what I thought, phased signals and multiplexed input/output data (or something similar). I now just need someone to explain why would be better to use QDR, instead of a faster bus, when AFAIK only reason FSBs (or the like) are not faster is because of the slow memories. With DDR the benefit is clear as you can use the rising and falling edges of same 1 signal, but QDR and ODR require an additional signal (be it a different one or the same one phased out) and I don't see much benefit there for main memory where good latency is must*. Maybe the answer is really simple and I'm just missing it out, I dunno.

* One of the requirements to convince me is that the advantage of using QDR is not use for memory cell/bank parallelization (like in GDDR5) as that wouldn't be a good solution for main memory.
Posted on Reply
#11
Morgoth
psst remember where using now imc goes no longer trough nortbridge
Posted on Reply
#12
DarkMatter
Morgoth said:
psst remember where using now imc goes no longer trough nortbridge
DarkMatter said:
FSBs (or the like)
I used FSB as a generic term. I should have said bus interconnect or something like that. QPI or HT are still buses and in some way it's still in the "up front" of the architecture. FSB is a very specific technology, but IMHO it also describes more or less what HT or QPI is in a generic way. Much like HD means 720/1080p, but the word itself could mean any high resolution.

Anyway IMC indeed helps my point. As I understand it IMC allows for much faster inerconnects between the CPU and the memory, so whenever faster memory (by any method) is available, the bus should be made faster instead of using "multiple instances" of the same one. Am I right or not?
Posted on Reply
#13
Poisonsnak
DarkMatter said:
... I now just need someone to explain why would be better to use QDR, instead of a faster bus ...
I think (?) the main reason to use a lower clock frequency is for signal synchronization over long distances (PCB traces). If 1 1GHz signal is travelling along 10cm of PCB trace at the speed of light then it takes 3 ns to make the trip, or more importantly 3 clock cycles.
Posted on Reply
#14
DarkMatter
Poisonsnak said:
I think (?) the main reason to use a lower clock frequency is for signal synchronization over long distances (PCB traces). If 1 1GHz signal is travelling along 10cm of PCB trace at the speed of light then it takes 3 ns to make the trip, or more importantly 3 clock cycles.
That doesn't make sense at all. Those numbers don't make sense to me. An electron or a hole travelling at light speed will cover 10 cm in 0,33 ns = 10cm/(300.000km/s*1000m/km*100cm/m). So it would have the time to do 3 travels.

Anyway, why would that matter? As I see it, it doesn't. It would be like saying that in a chain montage, you can't have a product every second because it takes 4 hours to each of them to go from start to finish. It's the production rate which matters, and unless I'm missing something important the same happens in electronics. Besides clock speed limits are constituted by much slower elements, such as the gate's (NAND, NOR...) state change delay, or the delay in transistors state change (one depends on the other really). Following the analogy, we can compare that to the time it takes to fullfill the trailers that will carry the goods to another place.

I understand there are limitations on clock speed, but considering the speeds at which GDDR5 runs, doubling what we have in our mobos several times wouldn't be a problem yet.
Posted on Reply
Add your own comment