Sent: Tuesday, December 18, 2001 11:10 AM
Subject: PCI Latency Patch

(((To George Breese - i hope the email-adress is ok)))

Dear George,

i'm an editor from IDG Germany and we read about your newest PCI Latency Patch. And we made interesting Benchmarks.

Ok, first the story:

We made a lot of benchmarks with the new ATA/133-PCI-Controller on mainboards with Intel- and VIA-Chipsets. On mainboards with Intel-Chipset we measured Burst-Rates between 110 and 115 MByte/s. We made the benchmarks with the Maxtor Drives D540X and D740X and on a Promise Ultra133 TX2 and Highpoint RocketRAID 133. If the controller is working in mainboards with VIA-Chipsets, the burst-rates of the harddisks slow down to 65 - 75 MByte/s.

To verify the problem, we checked the IDE-timings of the PCI-controller with our Logic-Analyzer. The ATA/133 cycle-time of 30 ns was correct (on mainboards with intel and via chipsets) --> the controller worked in ATA/133-Mode. But only on mainboards with Intel-Chipset we saw long bursts without interrupts. You see it on the following picture:

If we put the ATA/133-Controller in a mainboard with VIA-Chipset (we measured with a P4X266A), the burst begins correct and then continues always with interrupts every 1,2 Ás. Look at the next picture:

These interrupts slow down the Burst Rate on mainboards with VIA-Chipsets. In a next step we measured the timings of the PCI-Bus with the Logic-Analyzer. And this was very interesting. Intel-Chipsets can transfer 1024 data-packages in one burst on the PCI-Bus without interrupts (1024 x 32 Bit --> Burst-Size of 4094 Bytes), then a new adress comes. But VIA-Chipsets only transfer 24 data-packages in one burst (24 x 32 Bit --> Burst-Size of 96 Bytes) and then a new adress is on the PCI-Bus. So, the bandwith of the PCI-Bus from VIA-Chipsets is lower as from Intel-Chipsets, because they kill bandwith with the more frequently adresses (smaller bursts) on the bus.

As the ATA/133-PCI-Controllers allow Burst-Rates of over 100 MByte/s (without overhead), they reach the limit of the PCI-Bus. And as VIA-Chipsets have a smaller Burst-Size on the PCI-Bus, they slow down ATA/133-Controller. Intel-Chipsets allow much longer Bursts on the PCI-Bus.

We already published an article about the slow PCI-performance of VIA-Chipsets last week. You will find it here (sorry, german only):

Many people acknowledged us the problem of VIA's slow PCI-Bus (especially people from Highpoint and ACARD).

On this Thursday we want to publish the article in an updated version. And there we want to write one page about your patch with benchmarks and Timings from our Logic-Analyzer. With your patch, the burst-rate increased on Via-based mainboards from 78 to 93 MByte/s (with KT133A, on a MVP3-Mainboard from 64 to 84 MByte/s). Our Logic-Analzyer showed us, that the VIA-PCI-Bus transfers now 32 data-packages in one burst, if your patch is installed (without patch 24).

Can you explain us, what your PCI-Latency-Patch exactly does? By the way, from VIA we received no answer until now about the PCI-problem....

Many thanks & greetings from Germany

Best regards