AGP question

[pigdos]

I learned in my basic digital logic and design courses that the clock rate
determined the rate at which flip-flops/registers could read in data. My
question is how can a GPU running at say, 500Mhz, clock-in AGP data at a
rate of 2+GB/sec? I suppose if the registers clock in data on the rising and
falling edges it wouldn't be an issue. Is this what happens?

 

[Trent]

I learned in my basic digital logic and design courses that the clock rate
determined the rate at which flip-flops/registers could read in data. My
question is how can a GPU running at say, 500Mhz, clock-in AGP data at a
rate of 2+GB/sec? I suppose if the registers clock in data on the rising and
falling edges it wouldn't be an issue. Is this what happens?

You're assuming that the bus width is 8 bits/ 1 byte.

 

[The little lost angel]

I learned in my basic digital logic and design courses that the clock rate
determined the rate at which flip-flops/registers could read in data. My
question is how can a GPU running at say, 500Mhz, clock-in AGP data at a
rate of 2+GB/sec? I suppose if the registers clock in data on the rising and
falling edges it wouldn't be an issue. Is this what happens?

Hmm AGP bus is 32bit. So 500Mhz x 32 bit at single data rate is just
nice 2GB/sec. pPp

 

[Robert Redelmeier]

Hmm AGP bus is 32bit. So 500Mhz x 32 bit at single data
rate is just nice 2GB/sec. pPp

True enough, but actually there's more headroom: IIRC, plain
AGP is 64bit * 66 MHz = 528 MB/s [peak, outside of longish setup]
AGP does go to 8x = 4.2 GB/s bus limit. If GPU is 64+bit, then
500 MHz GPU limit is 4.0 GB/s, assuming one clock per xfr.

The GPU's bus is mostly busy hammering on vid.ram. DMA [busmaster]
transfer to system RAM are usually infrequent, which is one reason
why plain PCI vid.cards are sold (The other reason is for expansion
of PCIe- & AGP-less econoboxes). When transfer between system &
vid.RAM _is_ important (software MPEG2 decoding or other attempts
to use the CPU for GPU work), then systems often cannot keep up.


-- Robert

 

[George Macdonald]

Hmm AGP bus is 32bit. So 500Mhz x 32 bit at single data
rate is just nice 2GB/sec. pPp

True enough, but actually there's more headroom: IIRC, plain
AGP is 64bit * 66 MHz = 528 MB/s [peak, outside of longish setup]
AGP does go to 8x = 4.2 GB/s bus limit. If GPU is 64+bit, then
500 MHz GPU limit is 4.0 GB/s, assuming one clock per xfr.

I think you're thinking of PCI-X. AGP is 32-bits address & data + 8-bits
SBA. For AGP 3.x 8x, the common clock is 66MHz, the source synchronous
strobe clocks are 4x that with DDR giving 533MT/s for a peak bandwidth for
AGP 8x of 2.1GB/s.

 

[pigdos]

I see. So obviously any modern GPU can more than keep up w/AGP. Is this one
reason PCI express is the superior tech?