Understanding XB1's internal memory bandwidth *spawn

[zupallinere]

You're doubting the possibility? I don't see why'd they bother lying to X1 devs first. People who have the hardware.

It's not a lie if it can happen once in a while but that doesn't mean it is a reliable metric. With no context for the number it seems like a classic PR fact that looks good on a bulleted list but means little for practical purposes. Who knows when those kinds of numbers can actually occur ?

By the way if people are adding up bandwidths can one say that their computer with a gtx 680 now has 260 GB/sec of bandwidth ( 68+192) ? What other kind of bandwidth additions can be used ??

 

[Scott_Arm]

It's not a lie if it can happen once in a while but that doesn't mean it is a reliable metric. With no context for the number it seems like a classic PR fact that looks good on a bulleted list but means little for practical purposes. Who knows when those kinds of numbers can actually occur ?

By the way if people are adding up bandwidths can one say that their computer with a gtx 680 now has 260 GB/sec of bandwidth ( 68+192) ? What other kind of bandwidth additions can be used ??

If you have a device that can read from two pools of memory at once, why wouldn't you say it has available bandwidth equal to the sum of the bandwidth to each pool? Why would that not be true? In your example of the computer with a gtx 680, it is not true because (as far as I know) the GPU cannot access the system RAM directly. It has to copy from system RAM to GPU RAM. In the case of the Xbox One, the GPU can access the ESRAM and the DDR3 at the same time, as far as we know. Yes, the ESRAM is very small, so you are unlikely to sustain peak bandwidth from reads, but the bandwidth is available.

 

[zupallinere]

If you have a device that can read from two pools of memory at once, why wouldn't you say it has available bandwidth equal to the sum of the bandwidth to each pool? Why would that not be true? In your example of the computer with a gtx 680, it is not true because (as far as I know) the GPU cannot access the system RAM directly. It has to copy from system RAM to GPU RAM. In the case of the Xbox One, the GPU can access the ESRAM and the DDR3 at the same time, as far as we know. Yes, the ESRAM is very small, so you are unlikely to sustain peak bandwidth from reads, but the bandwidth is available.

Well the Maxwell chipset will be able to access system memory so the 680 isn't a good example. So with a graphics card that can access system memory we are talking 16 GB/sec ( pcie gated ) + say 192 for 208. Seems that would be a good bullet point for PC makers to use of. Of course the differential between systems would just be measuring the GPU bandwidth.

 

[pjbliverpool]

In your example of the computer with a gtx 680, it is not true because (as far as I know) the GPU cannot access the system RAM directly. It has to copy from system RAM to GPU RAM. In the case of the Xbox One, the GPU can access the ESRAM and the DDR3 at the same time, as far as we know. Yes, the ESRAM is very small, so you are unlikely to sustain peak bandwidth from reads, but the bandwidth is available.

He asked if the computer could be claimed to have that much bandwidth as opposed to the GPU alone. In that context I'd say its perfectly valid to add them together. Apart from the fact that no current PC can achieve 68GB/s of system memory bandwidth of course.

 

[pjbliverpool]

Well the Maxwell chipset will be able to access system memory so the 680 isn't a good example. So with a graphics card that can access system memory we are talking 16 GB/sec ( pcie gated ) + say 192 for 208. Seems that would be a good bullet point for PC makers to use of. Of course the differential between systems would just be measuring the GPU bandwidth.

16GB/s each way so in esram terms that's 32GB/s

 

[taisui]

I had a theory on the ESRAM being async in design with all numbers matched up, see #212

 

[zupallinere]

16GB/s each way so in esram terms that's 32GB/s

Yeah I was trying not to go down that road but yeah.

 

[dobwal]

It's not a lie if it can happen once in a while but that doesn't mean it is a reliable metric. With no context for the number it seems like a classic PR fact that looks good on a bulleted list but means little for practical purposes. Who knows when those kinds of numbers can actually occur ?

By the way if people are adding up bandwidths can one say that their computer with a gtx 680 now has 260 GB/sec of bandwidth ( 68+192) ? What other kind of bandwidth additions can be used ??

That assertion would apply to all max bandwidth figures. Relatively nothing pulls max bandwidth 100% of the time and context is rarely provided.

 

[ROG27]

That assertion would apply to all max bandwidth figures. Relatively nothing pulls max bandwidth 100% of the time and context is rarely provided.

I believe in Leadbetter's article, MS themselves purported that the best real world throughput in an ideal alpha blending scenario was ~133 GB/s. I believe they also said that achieving simultaneous read and writes was on a limited basis and not how the ESRAM normally functions most of the time. I could be misinterpreting, however.

 

[dobwal]

Looking at the leaked documentation on VGLeaks (http://www.vgleaks.com/durango-memory-system-overview/), I'm not sure how MS now arrives at 109 GB/s (102.4 GB/s prior to uptick) being the minimum bandwidth available for the ESRAM and 204 GB/s (192 GB/s prior to uptick) being the maximum, when diagrams included in their prior documentation seem to indicate that the bidirectional read and write bus had a 51.2 GB/s max throughput for read and 51.2 GB/s max throughput for write? How can both read and write's throughput seemingly double to now handle 102.4 GB/s each (factoring in the uptick)?

Actually thats when moving data within eSRAM.

Whats funny in that chart is that the vgleaks states that esram access to dram is a max bandwidth for both read or write at 68 GBs with a combined max bandwidth of 136 GBs which is a figure that makes absolutely no sense.

 

[ROG27]

Actually thats when moving data within eSRAM.

Whats funny in that chart is that the vgleaks states that esram access to dram is a max bandwidth for both read or write at 68 GBs with a combined max bandwidth of 136 GBs which is a figure that makes absolutely no sense.

I noticed this, as well. It seems like their own internal documentation is almost contradictory at times. System DRAM certainly is not capable of simultaneously reading and writing.

 

[taisui]

Actually thats when moving data within eSRAM.

Whats funny in that chart is that the vgleaks states that esram access to dram is a max bandwidth for both read or write at 68 GBs with a combined max bandwidth of 136 GBs which is a figure that makes absolutely no sense.

why not?
You read at 68GBps from the ESRAM, write to the DRAM at 68Gbps, the total BW consumes is 68+68 = 136?

 

[blakjedi]

why not?
You read at 68GBps from the ESRAM, write to the DRAM at 68Gbps, the total BW consumes is 68+68 = 136?

On the leak, you can read from the esram at 102gb/s while simultaneously reading from the RAM at up to 68 gb/s pre-upclock..

 

[taisui]

On the leak, you can read from the esram at 102gb/s while simultaneously reading from the RAM at up to 68 gb/s pre-upclock..

The table was specifically laid out as examples of moving data between memory. You can't write more than you can read, and nor can you read more than you can write? Note the source/destination legends on the table.

 

[dobwal]

why not?
You read at 68GBps from the ESRAM, write to the DRAM at 68Gbps, the total BW consumes is 68+68 = 136?

If you are reading from esram and writing that data to DRAM at 68 GBs. You shouldn't be able to read data from DRAM and write into ESRAM at 68 GBs at the same time. Even if eSRAM would allow it, the DRAM won't.

 

[ROG27]

why not?
You read at 68GBps from the ESRAM, write to the DRAM at 68Gbps, the total BW consumes is 68+68 = 136?

You are actually correct here because the table is illustrating bandwidth consumed, not throughput at anyone point in time for just the ESRAM I believe, no?

 

[taisui]

If you are reading from esram and writing that data to DRAM at 68 GBs. You shouldn't be able to read data from DRAM and write into ESRAM at 68 GBs at the same time. Even if eSRAM would allow it, the DRAM won't.

Do elaborate?

 

[dobwal]

Do elaborate?

DDR3 doesn't support concurrent reads and writes.

 

[blakjedi]

The table was specifically laid out as examples of moving data between memory. You can't write more than you can read, and nor can you read more than you can write? Note the source/destination legends on the table.

Yes. GPU Read bandwidth is 170gb. Write is 102.

http://www.vgleaks.com/durango-memory-system-overview/

 

[taisui]

DDR3 doesn't support concurrent reads and writes.

it's not doing that though. The table was describing combination of multiple scenarios, 2 source x 2 dst = 4 outcomes.