AMD EPYC 9655 Benchmarks Show The Terrific Generational Gains With 5th Gen EPYC

[sophisticles]

The fact that vendors tend to update RAM speed to match what their new cpu's can adequately use should be obvious, not require any kind of analysis.

Otherwise we'd still be running 133mhz memory, and that would be bottlenecked. Too much RAM performance available for the cpu to use, and it just wastes money. So they attempt to strike a balance, but they have to provide enough that future factory overclocked releases still have adequate room to grow. If RAM speed isn't keeping up with what the cpu can do, they add more on-chip cache, to balance things out again.

If you want any proof, just go to any phoronix article where Michael tests out RAM speeds. There should be a few of them to look through. Any cpu vendor. You'll get an occasional outlier test where it matters, and a bunch where you get slight improvements, no matter what cpu you look at.

You have completed ignored what I said, I pointed out that the performance increase, when normalized for core/thread count, between the 9655 and the 9654, 5th Gen vs 4th Gen, is almost entirely the result of increased clock speed and faster ram speed support and not due to any significant IPC improvements.

AMD's current processors have very fat cores that quickly get bottlenecked in most tasks as core count goes up

Math doesn't lie, my statements are correct.

[smitty3268]

How do you explain the fact that my analysis matches the observed across the board increase with 93% accuracy?

The fact that vendors tend to update RAM speed to match what their new cpu's can adequately use should be obvious, not require any kind of analysis.

Otherwise we'd still be running 133mhz memory, and that would be bottlenecked. Too much RAM performance available for the cpu to use, and it just wastes money. So they attempt to strike a balance, but they have to provide enough that future factory overclocked releases still have adequate room to grow. If RAM speed isn't keeping up with what the cpu can do, they add more on-chip cache, to balance things out again.

If you want any proof, just go to any phoronix article where Michael tests out RAM speeds. There should be a few of them to look through. Any cpu vendor. You'll get an occasional outlier test where it matters, and a bunch where you get slight improvements, no matter what cpu you look at.

[helf]

I have to chime in here:

He said he had an AMD /duron/ and moved from PC133 to PC2100 (DDR266).

I can *totally* believe a 50% speed up. The Duron had 128K L1 and 64K L2 and it was *intensely* memory bandwidth hungry. Much like the Celeron Netburst chips.

Still I insist that mp3 conversion is not such a memory-bound task that it would benefit so much from increased memory bandwidth. Or the encoder you were using was so poorly written that used so much memory for a task that is CPU bound to make that happen.



As said, it depends on the workload. Some tasks will run slightly slower, some memory bound tasks will run significantly slower.




You can be even a NASA scientist, but you misunderstood I was speaking about some technicalities beyond memory bandwidth vs. real world performance.

[sophisticles]

If you are entirely bottlenecked by memory speed, then sure. That's almost never the case, though, and certainly never the case in any kind of large benchmark suite that's testing a lot of things and averaging them.

In most cases, boosting memory speed by 25% is probably going to give a 5% or smaller overall performance boost. Expecting it to give 25% and then saying it's a failure otherwise is crazy talk unless you have a particular memory-bound test in mind.

Edit: and the fact that you're talking about 133Mhz ram just proves what a non-issue this has been in current systems.

How do you explain the fact that my analysis matches the observed across the board increase with 93% accuracy?

[smitty3268]

Actually, many times, it does, if memory speed is a bottleneck.

If you are entirely bottlenecked by memory speed, then sure. That's almost never the case, though, and certainly never the case in any kind of large benchmark suite that's testing a lot of things and averaging them.

In most cases, boosting memory speed by 25% is probably going to give a 5% or smaller overall performance boost. Expecting it to give 25% and then saying it's a failure otherwise is crazy talk unless you have a particular memory-bound test in mind.

Edit: and the fact that you're talking about 133Mhz ram just proves what a non-issue this has been in current systems.

[blackshard]

The doubling of performance was obviously due to the 133Mhz SDR bottlenecking the system, the switch to DDR 266 eliminated the bottleneck and aloud the Duron to be all it could be, as evidenced by the boot times cut in half.

Still I insist that mp3 conversion is not such a memory-bound task that it would benefit so much from increased memory bandwidth. Or the encoder you were using was so poorly written that used so much memory for a task that is CPU bound to make that happen.

Think about your computer right now, regardless of what it is. If you swap out the ram with ram at half the speed, what do you think will happen?

These EPYCs have 96 fat cores, with lots of execution units per core, capable of handling two threads simultaneously and they need to be fed and clearluy they are not being fed fast enough.

As said, it depends on the workload. Some tasks will run slightly slower, some memory bound tasks will run significantly slower.

"Useless simplifications" are how data analysis is done and how forecasting is accomplished.
...
cut
...

You can be even a NASA scientist, but you misunderstood I was speaking about some technicalities beyond memory bandwidth vs. real world performance.

[sophisticles]

I really doubt mp3 encoding would benefit so much from memory bandwidth increase, since it is not a memory bound task but rather a heavy CPU bound task. If an increase in performance could be seen with more memory throughput, indeed it would not be totally depending to it.

The doubling of performance was obviously due to the 133Mhz SDR bottlenecking the system, the switch to DDR 266 eliminated the bottleneck and aloud the Duron to be all it could be, as evidenced by the boot times cut in half.

Back in the day motherboards were much more forgiving in the CPU used as many CPU's used the same socket so if the CPU fit the motherboard the system would work.

I don't remember what what CPU I had before the Duron but I remember finding a Duron at a great price so I pulled the existing CPU that only supported SDR and popped in the Duron; performance increased but not like i had hoped. The new motherboard and ram eliminated the bottleneck.

Think about your computer right now, regardless of what it is. If you swap out the ram with ram at half the speed, what do you think will happen?

These EPYCs have 96 fat cores, with lots of execution units per core, capable of handling two threads simultaneously and they need to be fed and clearluy they are not being fed fast enough.

What you are telling about is a useless simplification of the real world, because you don't consider the timings, the access patterns, the barriers and, in general, the whole memory hierarchy.

"Useless simplifications" are how data analysis is done and how forecasting is accomplished.

As i have stated I spent 10 years working for medical labs, many of those years analyzing data and using the results to make predictions.

Many times you have numerous interdependent and independent variables to consider and it is impossible to create a math model that models the data 100 percent.

What you do is you simplify, the same way you were taught to simplify equations in grammar school, so that you can create a model that approximates the data to a high degree and can be used to forecast future trends.

You saw an example of it right here, I took the base and max clock speed of the two compared processors, did the same for the two ram speeds and created a simple model to predict the performance increase.

Since we have the actual data we can determine how accurate the model is, my model predicts a 37% performance increase and the observed increase was about 40%, ergo my model is accurate to a high degree.

There is no need to over-complicate things, it's not rocket science.

[geerge]

Actually, many times, it does, if memory speed is a bottleneck.

Back in the Duron days I had one with 133Mhz SDR and I bought a new motherboard that supported 266Mhz DDR, performance doubled, mp3 encoding was twice as fast as was VCD and SVCD encoding, boot times were cut in half.

In this case, clearly these CPUs are bottlenecked by ram speed, because my calculated speed increase matches Michael's measured speed increase.

If you think about it, it makes sense, you have lots of big fat cores that need to be fed with data, and obviously they are not being fed fast enough.

AMD can probably keep increasing performance at a good clip by just increasing the ram speed that's supported.

You're right memory bandwidth is important for a chunk of workloads. It's always a balance between cores and bandwidth, and despite jumping to ddr5 the cores have gotten better faster so the set of workloads that are bandwidth starved has increased relative to say a decade ago. Some workloads remain fed thanks to cache, some are alleviated somewhat by cache. AMD (and presumably intel) is prepping for 17600 MR-DIMM's. intel has done some shenanigan's with 8800 MCR-DIMM's that for some reason they're calling MR-DIMM's, naming bullshit as usual probably. If I had to guess why they did the 8800 half-measure, probably some of their accelerators were bandwidth-starved so one couldn't be implemented without the other, and they had to do something to be competitive sometimes, as a stop-gap.

[blackshard]

Actually, many times, it does, if memory speed is a bottleneck.

Back in the Duron days I had one with 133Mhz SDR and I bought a new motherboard that supported 266Mhz DDR, performance doubled, mp3 encoding was twice as fast as was VCD and SVCD encoding, boot times were cut in half.

In this case, clearly these CPUs are bottlenecked by ram speed, because my calculated speed increase matches Michael's measured speed increase.

If you think about it, it makes sense, you have lots of big fat cores that need to be fed with data, and obviously they are not being fed fast enough.

AMD can probably keep increasing performance at a good clip by just increasing the ram speed that's supported.

I really doubt mp3 encoding would benefit so much from memory bandwidth increase, since it is not a memory bound task but rather a heavy CPU bound task. If an increase in performance could be seen with more memory throughput, indeed it would not be totally depending to it.

Also dramatic increase in performance with higher memory bandwidth means that the workload data does not fit into CPU caches, so it is a matter of the "size" of the workload. And caches are definitely there to avoid stalling the CPU and let it wait for data coming from DRAM. Only when your CPU stalls costantly for data from DRAM, you get such dramatic performance increases.

What you are telling about is a useless simplification of the real world, because you don't consider the timings, the access patterns, the barriers and, in general, the whole memory hierarchy.

[sophisticles]

That's not how things work. Doubling your RAM speed won't double performance.

Actually, many times, it does, if memory speed is a bottleneck.

Back in the Duron days I had one with 133Mhz SDR and I bought a new motherboard that supported 266Mhz DDR, performance doubled, mp3 encoding was twice as fast as was VCD and SVCD encoding, boot times were cut in half.

In this case, clearly these CPUs are bottlenecked by ram speed, because my calculated speed increase matches Michael's measured speed increase.

If you think about it, it makes sense, you have lots of big fat cores that need to be fed with data, and obviously they are not being fed fast enough.

AMD can probably keep increasing performance at a good clip by just increasing the ram speed that's supported.