3 ms·

Less Physical Distance should create less latency, plus an optimized profile for one list of model of banks vs the diversity in the modula industry with differe

by javchz 4y ago

Less Physical Distance should create less latency, plus an optimized profile for one list of model of banks vs the diversity in the modula industry with different clock speeds and latencies.

mtlmtlmtlmtl 4y ago
Is there any actual evidence that this marginally lower latency makes any real life difference? And why would the laptops suddenly ship with a variety of modules if they're replaceable? You can still ship it with the same modules in every laptop and get those benefits. And if someone upgrades it, that's still an improvement over no upgrade path, so this makes no sense to me.
- kube-system 4y ago
  > Is there any actual evidence that this marginally lower latency makes any real life difference? Yes, compare the specs of DDR4 and DDR5 to LPDDR4X and LPDDR5X. The latter are significantly higher performance. This is also the reason that Dell recently introduced CAMM memory modules -- it is an attempt to address the packaging bottleneck that is limiting the speed of DIMMs currently. The amount of time that it takes a signal to go down a wire has been relevant for DRAM for a while now. If you look at the traces on the board of any relatively modern computer, you'll see some that take circuitous routes, for the purpose of having the signals arrive at the CPU at the same time. You can see this even on relatively low performance devices like a Raspberry Pi. https://www.cnx-software.com/wp-content/uploads/2019/06/Raspberry-Pi-4-Model-B-Large.jpg https://www.cnx-software.com/wp-content/uploads/2019/06/Rasp...
  - wtallis 4y ago
    > If you look at the traces on the board of any relatively modern computer, you'll see some that take circuitous routes, for the purpose of having the signals arrive at the CPU at the same time. Having all the signals arrive at the same time isn't the same thing as having the signals arrive soon.
    - kube-system 4y ago
      Correct. My point was to simply illustrate that the fractions of time involved are small enough that trace length matters.
- Wowfunhappy 4y ago
  High bandwidth, low latency unified memory is a central component of the M series architecture and a key reason those chips perform so well at their power profile. I'm not sure what "evidence" I could provide that would convince you since we don't have high latency Apple chips to benchmark against. However, there's a reason VRAM is soldered onto GPUs.
wtallis 4y ago
Soldered RAM doesn't really help much with latency; most of the DRAM latency ocurrs within the chip itself. Where soldering RAM does help is with reaching higher clock speeds with lower power (in phones and laptops), or with reaching clock speeds that are impossible to shove through a DIMM connector (GPUs). That higher frequency helps phones save on pin count by using a narrow memory bus, and allows laptops to have lots of memory bandwidth to feed the integrated graphics when using typical laptop/desktop bus widths.

by javchz 4y ago

Less Physical Distance should create less latency, plus an optimized profile for one list of model of banks vs the diversity in the modula industry with different clock speeds and latencies.

mtlmtlmtlmtl 4y ago

Is there any actual evidence that this marginally lower latency makes any real life difference? And why would the laptops suddenly ship with a variety of modules if they're replaceable? You can still ship it with the same modules in every laptop and get those benefits. And if someone upgrades it, that's still an improvement over no upgrade path, so this makes no sense to me.

kube-system 4y ago
> Is there any actual evidence that this marginally lower latency makes any real life difference? Yes, compare the specs of DDR4 and DDR5 to LPDDR4X and LPDDR5X. The latter are significantly higher performance. This is also the reason that Dell recently introduced CAMM memory modules -- it is an attempt to address the packaging bottleneck that is limiting the speed of DIMMs currently. The amount of time that it takes a signal to go down a wire has been relevant for DRAM for a while now. If you look at the traces on the board of any relatively modern computer, you'll see some that take circuitous routes, for the purpose of having the signals arrive at the CPU at the same time. You can see this even on relatively low performance devices like a Raspberry Pi. https://www.cnx-software.com/wp-content/uploads/2019/06/Raspberry-Pi-4-Model-B-Large.jpg https://www.cnx-software.com/wp-content/uploads/2019/06/Rasp...
- wtallis 4y ago
  > If you look at the traces on the board of any relatively modern computer, you'll see some that take circuitous routes, for the purpose of having the signals arrive at the CPU at the same time. Having all the signals arrive at the same time isn't the same thing as having the signals arrive soon.
  - kube-system 4y ago
    Correct. My point was to simply illustrate that the fractions of time involved are small enough that trace length matters.
Wowfunhappy 4y ago
High bandwidth, low latency unified memory is a central component of the M series architecture and a key reason those chips perform so well at their power profile. I'm not sure what "evidence" I could provide that would convince you since we don't have high latency Apple chips to benchmark against. However, there's a reason VRAM is soldered onto GPUs.

wtallis 4y ago

Soldered RAM doesn't really help much with latency; most of the DRAM latency ocurrs within the chip itself. Where soldering RAM does help is with reaching higher clock speeds with lower power (in phones and laptops), or with reaching clock speeds that are impossible to shove through a DIMM connector (GPUs). That higher frequency helps phones save on pin count by using a narrow memory bus, and allows laptops to have lots of memory bandwidth to feed the integrated graphics when using typical laptop/desktop bus widths.