AMD CPU impacts of memory latency

[BrightCandle]

The PS4 is going to be equipped with GDDR5, which by my calculation looks to have about 16% more latency than typical DDR3. Given that I am interested in how AMD processors typically respond to such an increase in latency from RAM.

So lets say you had 11-11-11-28 on some 2133 Mhz RAM that 16% increase would coincide with 13-13-13-33. We can get to that just by multiplying each timing by 1.16 and then rounding the value up.

Anyone fancy giving me an idea of numbers with their CPU. If someone could do something basic like a SuperPI run with both timings or point me to a review that investigates this I would appreciate it.

 

[ShintaiDK]

The latency on a GTX680 is in the 200ns range. While less than 30ns on a CPU.

I assume the PS4 uses the 8x32bit access for memory.

However latency on the CPUs have been mostly irrelevant for the last many years due to caches.

 

[sefsefsefsef]

What calculations did you perform that led you to the conclusion that GDDR5 has higher latency than DDR3? GDDR5 devices themselves do not have higher latency for row activation, column read, or precharge commands (which most people typically associate with "latency"), so I'm wondering where you came up with your notion that it has higher latency.

 

[BrightCandle]

What calculations did you perform that led you to the conclusion that GDDR5 has higher latency than DDR3? GDDR5 devices themselves do not have higher latency for row activation, column read, or precharge commands (which most people typically associate with "latency"), so I'm wondering where you came up with your notion that it has higher latency.

From hynix's datasheets on its GDDR5 modules. They describe in detail all the timings and thus the latencies and show them in ns as well as some of them in clock cycles at various frequencies. It shows a clear trend of those timings being higher than an average set of DDR3 modules, by about 16%. A high end set of DDR3 however could outperform it more than that but I thought it unlikely that a) they would make any real performance difference and b) it unlikely any AMD user was using such high end modules for comparison.

Neitherless hynix's datasheet is pretty clear, the timings are quite noticeably higher than a normal DDR3 module, which is not a surprise considering the clock speed they are trying to achieve and the interface adjustments that were necessary to achieve the throughput they do.

 

[sefsefsefsef]

From hynix's datasheets on its GDDR5 modules. They describe in detail all the timings and thus the latencies and show them in ns as well as some of them in clock cycles at various frequencies. It shows a clear trend of those timings being higher than an average set of DDR3 modules, by about 16%. A high end set of DDR3 however could outperform it more than that but I thought it unlikely that a) they would make any real performance difference and b) it unlikely any AMD user was using such high end modules for comparison.

Neitherless hynix's datasheet is pretty clear, the timings are quite noticeably higher than a normal DDR3 module, which is not a surprise considering the clock speed they are trying to achieve and the interface adjustments that were necessary to achieve the throughput they do.

DDR3 timings, this PDF on page 48:
http://www.hynix.com/datasheet/pdf/dram/H5TQ1G4(8_6)3AFP(Rev0.1).pdf

GDDR5 timings, this PDF on page 133:
http://www.hynix.com/datasheet/pdf/graphics/H5GQ1H24AFR(Rev1.0).pdf

The timings (measured in nanoseconds), are actually better for GDDR5 than they are for DDR3. Do you have a specific source you can cite that contradicts this?

And by the way, the external IO frequency doesn't necessarily have anything to do with the internal latency of a chip. You can have a chip with very high speed IOs, and very low internal latency, like GDDR5. There is no conflict there.

 

[2is]

Timings alone don't tell you much, you have to look at the clock speed too. They work together.

 

[itsmydamnation]

People need to understand that GPU memory latency is largely caused by the GPU itself not GDDR5. its focused on throughput not latency. CPU's on the other hand its much more of a balance, intel has faster memory access then AMD yet both stop seeing performance gains at around the same memory speed.