3 ms·

So I've observed something from many of the tests/reviews posted so far, as well as Apple's specs: Apple claims 200GBps memory bandwidth for the M1 Pro, and 400

by trebor 5y ago

So I've observed something from many of the tests/reviews posted so far, as well as Apple's specs: Apple claims 200GBps memory bandwidth for the M1 Pro, and 400GBps for the M1 Max. Some tests show the M1 Max saturating the memory bandwidth between 200 - 240 GBps rates. Maybe someone who understands better than me can explain these observations: 1. DDR5 RAM is supposedly 36-38GBps bandwidth, DDR4 3200 is about 25.6GBps bandwidth. (Supposedly Alder Lake is the first chipset to support DDR5.) 2. Maybe the 200-400 GBps rate advertised is the entire bus bandwidth for the SoC? If this is actual memory bandwidth, then Apple increased that bus speed 5-15x! 3. Maybe Apple divided the RAM into additional banks with more lanes to increase the bandwidth? I've been very irritated by the pricing, but if Apple figured out how to optimize the memory bandwidth this much then maybe it's worth the extra cost per GB. Did I catch on to something majorly different here?

jackpeterfletch 5y ago
Anandtechs article covers this in a lot of detail. But yeah it is supposed that 400GBps is peak bandwidth for the entire SoC. That’s CPU incl pow power cores, GPU, NPU and whatever else firing on all cylinders, in an ideal access pattern. https://www.anandtech.com/show/17024/apple-m1-max-performance-review https://www.anandtech.com/show/17024/apple-m1-max-performanc...
Tagbert 5y ago
In the tests I had seen about the bandwidth question, the findings suggest that the CPU can only push the memory at a little over 200GBps but if you also have the GPU active it will use the additional bandwidth. This is constant with the advantage of the M1 Max being the added GPU cores.
ksec 5y ago
M1 Pro and Max uses LPDDR5, not DDR5. >If this is actual memory bandwidth, then Apple increased that bus speed 5-15x! Yes because GPU requires lots of bandwidth. Top of the line Nvidia GPU has ~ 1TB/s memory bandwidth. >I've been very irritated by the pricing, but if Apple figured out how to optimize the memory bandwidth this much then maybe it's worth the extra cost per GB. You are not only paying for memory aka LPDDR5 chip, you are also paying for the additional memory controller. Which has an implication on yield and die size cost. Whether you think it is worth or still irritated by that pricing is of course subjective.
- trebor 5y ago
  So then this is part of why it seems so fast, because the CPU is now approaching GPU memory bandwidths? Hmm, I really wonder what it'll look like in a year or so with the M2/M2Pro/M2MAX then. Thanks for clarifying the extra memory controller, that's helpful too.
  - ksec 5y ago
    >because the CPU is now approaching GPU memory bandwidths? No. The single core CPU performance has zero relation to memory bandwidth.[0] i.e You do not gain or loss any performance going from M1 Pro 10 Core with ~200GB/s Memory bandwidth to M1 Max 10 Core with ~400GB/s bandwidth. The additional memory bandwidth are there only to support GPU performance. The reason why CPU performance is fast is slightly complicated. But to over-simplify it to two point, Apple has access to all the common code path on their platform and can optimise of it. e.g NSObject release is 5 times faster than on x86. Second being Apple managed to do something the whole industry thought ( at the time ) was impossible, having a wide decode and kept them well fed. You can read more about it here [1] on HN. > what it'll look like in a year or so with the M2/M2Pro/M2MAX then. You can take a look at [2], which is a chart from Nuvia, started by ex-Apple CPU architects. Given the amount of lead time in CPU design ( 3-4 years ), you can pretty much bet Apple will follow the path as shown by Nuvia to reach a Geekbench Score of 2000 within next 2-3 years. Hope this answer your questions. [0] CPU Performance is always latency sensitive. Hence the amount of cache matter much more than memory bandwidth. And Apple's SoC has always had large amount of cache compared to rest of the industry, even when it is a Smartphone SoC compared to Desktop part. [1] https://news.ycombinator.com/item?id=25257932 https://news.ycombinator.com/item?id=25257932 [2] https://www.anandtech.com/show/15967/nuvia-phoenix-targets-50-st-performance-over-zen-2-for-only-33-power https://www.anandtech.com/show/15967/nuvia-phoenix-targets-5...

by trebor 5y ago

jackpeterfletch 5y ago

Anandtechs article covers this in a lot of detail. But yeah it is supposed that 400GBps is peak bandwidth for the entire SoC. That’s CPU incl pow power cores, GPU, NPU and whatever else firing on all cylinders, in an ideal access pattern. https://www.anandtech.com/show/17024/apple-m1-max-performance-review https://www.anandtech.com/show/17024/apple-m1-max-performanc...

Tagbert 5y ago

In the tests I had seen about the bandwidth question, the findings suggest that the CPU can only push the memory at a little over 200GBps but if you also have the GPU active it will use the additional bandwidth. This is constant with the advantage of the M1 Max being the added GPU cores.

ksec 5y ago

M1 Pro and Max uses LPDDR5, not DDR5. >If this is actual memory bandwidth, then Apple increased that bus speed 5-15x! Yes because GPU requires lots of bandwidth. Top of the line Nvidia GPU has ~ 1TB/s memory bandwidth. >I've been very irritated by the pricing, but if Apple figured out how to optimize the memory bandwidth this much then maybe it's worth the extra cost per GB. You are not only paying for memory aka LPDDR5 chip, you are also paying for the additional memory controller. Which has an implication on yield and die size cost. Whether you think it is worth or still irritated by that pricing is of course subjective.

trebor 5y ago
So then this is part of why it seems so fast, because the CPU is now approaching GPU memory bandwidths? Hmm, I really wonder what it'll look like in a year or so with the M2/M2Pro/M2MAX then. Thanks for clarifying the extra memory controller, that's helpful too.
- ksec 5y ago
  >because the CPU is now approaching GPU memory bandwidths? No. The single core CPU performance has zero relation to memory bandwidth.[0] i.e You do not gain or loss any performance going from M1 Pro 10 Core with ~200GB/s Memory bandwidth to M1 Max 10 Core with ~400GB/s bandwidth. The additional memory bandwidth are there only to support GPU performance. The reason why CPU performance is fast is slightly complicated. But to over-simplify it to two point, Apple has access to all the common code path on their platform and can optimise of it. e.g NSObject release is 5 times faster than on x86. Second being Apple managed to do something the whole industry thought ( at the time ) was impossible, having a wide decode and kept them well fed. You can read more about it here [1] on HN. > what it'll look like in a year or so with the M2/M2Pro/M2MAX then. You can take a look at [2], which is a chart from Nuvia, started by ex-Apple CPU architects. Given the amount of lead time in CPU design ( 3-4 years ), you can pretty much bet Apple will follow the path as shown by Nuvia to reach a Geekbench Score of 2000 within next 2-3 years. Hope this answer your questions. [0] CPU Performance is always latency sensitive. Hence the amount of cache matter much more than memory bandwidth. And Apple's SoC has always had large amount of cache compared to rest of the industry, even when it is a Smartphone SoC compared to Desktop part. [1] https://news.ycombinator.com/item?id=25257932 https://news.ycombinator.com/item?id=25257932 [2] https://www.anandtech.com/show/15967/nuvia-phoenix-targets-50-st-performance-over-zen-2-for-only-33-power https://www.anandtech.com/show/15967/nuvia-phoenix-targets-5...