Replies to post #56918 on Advanced Micro Devices Inc (AMD)

[sgolds]

Tenchu, not exactly...

"Joe, I think he means that CPU communication to memory and I/O devices goes over the same bus (FSB), which is true."

Other than snoop cycles, that's not true. I/O devices access memory directly.

CPU communication to I/O devices would go over the FSB in Intel's design. There is no other path. However the bulk of data goes between memory and I/O through the memory controller (in the North Bridge), bypassing the CPU. I think that is what you are trying to say.

The only stuff that goes between the CPU and I/O devices (in any reasonable design) would be the occasional control byte. That does indeed go over the FSB, but the impact is negligable.

[Dan3]

Re: I/O devices access memory directly.

There certainly are cases where DMA operations support a fair amount of I/O activity that isn't dependent on the CPU, like a simple disk copy, but they are periodic cases, and almost always cases where the system is otherwise basically idle and we don't care about CPU bandwidth. The northbridges on AMD systems enjoy the same advantages as you are well aware.

But when TCP control or encryption or compression or sorting or comparison or rendering etc. is going, on, the data is streamed into the CPU and out to memory (it may be being concurrently streamed as DMA block I/O moves, but that's handled the same way in both architectures and has nothing to do with CPU bandwidth).

On an Athlon 64 and Opteron, data streams in through hypertransport data lines and out through the additional memory data lines. On Pentium M, Pentium 4, and Xeon, it streams in on the same data lines it's trying to stream out on - a bottleneck.

I understand that you're pretty upset that Intel appears to have totally screwed the pooch in mobile as well as desktop, but there's no need to start posting nonsense because of it.

"Intel is promising EM64T support in Yonah" http://67.19.9.2/?article=19149

Too bad.

[oc detective]

Correct me if I am wrong but the entire context of this was in the server sector (i.e Blackford) NOT your desktop ! So what exactly does you playing on your P4 have to do with anything?
Are you really saying that all I/O work has no impact on cpu utilization icluding TCP/IP? - I have to clarify if that is what you are actually claiming!
I guess iSCSI TOE cards (even Intel have their own for sale) must be a marketing gimmick then lol.
http://www.iscsistorage.com/intelhba.htm
Oh and what exactly is Intel's I/O AT supposed to achieve if I/O devices dont eat into cpu utilization?

[jhalada]

Tenchu,

DMA does bandwidth does outweigh the bandwidth of I/O transaction sent from the CPU, but I am not sure how universal DMA is in low end stuff. For example, I haven't checked this for a while, so it may have changed in latest chipsets, but last time I looked into it, IDE drive reading and writing used to result in high CPU utilization. So I would guess that the data is going through the CPU.

Joe