Memory bandwidth of A64/AFX/P4/Athlon XP

[Giscardo]

Here's what I understand so far:

P4: 64 (bits/transfer)* 200,000,000 * 4 (transfers/sec) * 8^-1 bytes/bit * 1,000,000,000^-1 (GByte/byte) = 6.4Gbytes/sec

A64: 128 (bits/transfer)* 200,000,000 * 2 (transfers/sec) * 8^-1 bytes/bit * 1,000,000,000^-1 (GByte/byte) = 6.4Gbytes/sec

AFX: 128 (bits/transfer)* 200,000,000 * 2 (transfers/sec) * 8^-1 bytes/bit * 1,000,000,000^-1 (GByte/byte) = 6.4Gbytes/sec

Athlon XP: 64 (bits/transfer)* 200,000,000 * 2 (transfers/sec) * 8^-1 bytes/bit * 1,000,000,000^-1 (GByte/byte) = 3.2Gbytes/sec

All four platforms can support dual channel (either via motherboard memory controller or on-CPU memory controller). With the same theoretical peak bandwidth on the first three chips, why does the P4 beat the A64, using the same FSB?

And what exactly is dual channel memory, it seems to do nothing at all for the XP, but on the A64/A64FX, there is a difference slightly over 25% (http://www.digit-life.com/articles2/roundupmobo/pentium4-32ghz-ee.html Why does it help so much more on AMD's 64 bit CPU platform. Does the nForce's implementation of dual channel just suck compared to the one on AMD's on-CPU memory controller?

Are there any comparisons of dual vs. non-dual channel memory configurations on the P4?

Edit: Fixed link, thanks DAPUN

 

[DAPUNISHER]

Consider this a bump so someone more technically adept can offer a detailed explanation. I will generally adress this post though. P4 needs the dual channel, it thrives on the bandwidth it produces and in single channel will get it's lunch handed to it vs dual channel so no need to hunt up comparisons because there isn't one, dual channel on P4 is a must have . The memory controller being on-die on A64 means much reduced latencies and that's why it is so much improved over the memory controllers used on nF2 if I understand it correctly.

The P4 and A64 don't use the same FSB, in fact it's not even really an FSB on the A64 but rather hypertransport speed. The HT speed is double pumped which means up to 2000HT 1000mhzx2 is now possible with the latest Via and nV chipsets. The P4 uses a quad pumped bus 200mhzx4=800fsb. HTH

 

[Giscardo]

Originally posted by: DAPUNISHER
Consider this a bump so someone more technically adept can offer a detailed explanation. I will generally adress this post though. P4 needs the dual channel, it thrives on the bandwidth it produces and in single channel will get it's lunch handed to it vs dual channel so no need to hunt up comparisons because there isn't one, dual channel on P4 is a must have . The memory controller being on-die on A64 means much reduced latencies and that's why it is so much improved over the memory controllers used on nF2 if I understand it correctly.

See, I already understand this from glancing over benchmark graphs, what I'm looking to understand is why dual channel makes such a big difference on the P4, and Athlon64, but not on the XP.

The P4 and A64 don't use the same FSB, in fact it's not even really an FSB on the A64 but rather hypertransport speed. The HT speed is double pumped which means up to 2000HT 1000mhzx2 is now possible with the latest Via and nV chipsets. The P4 uses a quad pumped bus 200mhzx4=800fsb. HTH

I see what you're saying, I'll have to read up on the Hypertransport stuff, but even if the HT were running at 2Ghz, when refering to the FSB I meant the memory, the ddr memory on the Athlon 64s runs at 200mhz * 2 = 400Mhz. The A64 and P4 have same theoretical bandwidth, but the P4 usually scores higher.

 

[DAPUNISHER]

Dual channel doesn't make a big difference on A64 vs single channel on A64, if that's what you meant then I'd like to see the numbers backing it up. If you mean in comparison to the 32bit XP then I'd refer back to latencies again. As to P4 scoring higher, in what? Sandra? For gaming even single channel A64 is better than dual channel P4 so I'm just not understanding what you are getting at, sorry

 

[Gamingphreek]

I think you have it mixed up. THere are no dual channel chipsets out for the Athlon 64 series... YET. The Athlon XP does extremely well with dual channel. Im not sure how much it will affect the A64 but we shall see when it comes out. The athlon XP is so dependant on bandwidth because it has no on die memory controller with hypertransport.

-Kevin

 

[DAPUNISHER]

Originally posted by: Gamingphreek
I think you have it mixed up. THere are no dual channel chipsets out for the Athlon 64 series... YET. The Athlon XP does extremely well with dual channel. Im not sure how much it will affect the A64 but we shall see when it comes out. The athlon XP is so dependant on bandwidth because it has no on die memory controller with hypertransport.

-Kevin

Dual channel on nF2 rarely exceeds 5% over single so I wouldn't call it doing "doing extremely well" vs single channel on a SPP board. The real benefit of DC on nF2 is the IGP basically has a channel to itself since 1 64bit channel gives the XP what it needs for bandwidth. I've seen benchies of skt939 A64 vs skt754 and once again there isn't much to be gained in most usage, just as is the case with the FX vs A64 SC.

 

[Giscardo]

Originally posted by: DAPUNISHER
Dual channel doesn't make a big difference on A64 vs single channel on A64, if that's what you meant then I'd like to see the numbers backing it up.

Hi DAPUNISHER, the place where I got the impression that dual vs. single on A64 is the link in my post at the top. Check out the section called "Memory performance", you can see the comparison between the A64 and the FX, the FX scores higher on the raw memory benches.

If you mean in comparison to the 32bit XP then I'd refer back to latencies again. As to P4 scoring higher, in what? Sandra? For gaming even single channel A64 is better than dual channel P4 so I'm just not understanding what you are getting at, sorry

Yes, Sandra as well as other raw memory subsystem tests. I am not getting at anything, I was asking a question, I want to know how dual channel works, and why we see a benefit on the A64/FX, but not on the XP. I am not trying to make a point, just asking for a sort of technical explanation for the numbers I've been seeing...

 

[Giscardo]

Originally posted by: Gamingphreek
I think you have it mixed up. THere are no dual channel chipsets out for the Athlon 64 series... YET. The Athlon XP does extremely well with dual channel. Im not sure how much it will affect the A64 but we shall see when it comes out.

-Kevin

I thought the on-die memory controller of the A64 is what determined support for dual channel memory, and from what I've read, the FX does support dual channel right now, and the plain A64 doesn't, isn't that right?

If you take a look at the benchmarks here: http://www.tomshardware.com/motherboard/20040308/kt800-06.html#opengl, you'll see that there's hardly a great difference between dual and single channel on the Athlon XP.

 

[DAPUNISHER]

Fixed that link for you fixed

Those memory results which make the FX look very impressive against the SC controller on the A64, unfortunately just don't show up in most usage.

 

[Apologiliac]

Can someone tell me how the athlons utilize the higher bus speed of a DDR memory bus if their buses are relatively lower?

 

[jagec]

Originally posted by: Gamingphreek
The Athlon XP does extremely well with dual channel. Im not sure how much it will affect the A64 but we shall see when it comes out. The athlon XP is so dependant on bandwidth because it has no on die memory controller with hypertransport.

-Kevin

You must be thinking of the P4...the athlon XP barely makes a difference in dual vs. single.

 

[klah]

The AXP can only transfer 3.2GB/s from the cpu to the northbridge. Single channel ddr400 can transfer 3.2GB/s to the northbridge. Adding more memory bandwidth is therefore useless since it is choked at the fsb. It is only useful in this platform if you are using integrated video, which when combined with the fsb can exceed 3.2GB/s memory transfer.

The P4 can transfer 6.4GB/s from the cpu to the northbridge and therefore benefits greatly from the 6.4GB/s that dual ddr400 provides. This is the only cpu that benefits from the additinoal memory bandwidth. There are several articles at aceshardware.com that go into detail about how the p4's long instruction pipeline leads to additional cache misses which requires all of this memory bandwidth. Look at the p4 articles from ca.2000 for more information about this.

The A64 architecture uses a drastically different arrangement and does not route memory transfer through the northbridge, where traditionally the memory controller has been located, but has a memory controller built into the cpu die which can access memory directly. Right now the A64 does not benefit much from the additional memory bandwidth afforded by dual channel ddr400.