3 ms·
>In the SPECfp suite, the M1 Max is in its own category of silicon with no comparison in the market. How much of that performance is due to the M1 Pro/Max havi
by ChuckNorris89 5y ago
>In the SPECfp suite, the M1 Max is in its own category of silicon with no comparison in the market.
How much of that performance is due to the M1 Pro/Max having way more memory bandwidth than the Intel/AMD chips, and also being specifically designed from the ground up to make use of all that bandwidth? AFAIK the RAM used by the M1 Pro/Max is more similar in performance to the GDDR used in graphics cards vs the slow-ish ageing DDR4 used in Intel/AMD systems that are designed to prioritize compatibility with RAM of varying quality, speeds and latencies instead of raw performance at a specific high speed.
So I'm curious to know how an X64 chip would perform if we even the playing field not just in node size but also if Intel and AMD would adapt their X64 designs from the ground up with a memory controller, cache architecture and instruction pipeline tuned to feed the CPU with data from such fast RAM.
I'm asking this since AFAIK, Ryzen is very sensitive to memory bandwidth, the more you give it the better it performs to the point where if you take two identical laptops with the same Ryzen chip but one has 33% faster RAM, then that laptop will perform nearly 33% better in most CPU/GPU intensive benchmarks, all things being equal.
- namibj 5y agoShould be fine to compare to an EPYC 72F3 or 73F3 (Zen3, 8/16 cores, 8 channel DDR4-3200 RDIMM (204.8 GB/s theoretical ceiling), $2500/$3500). If memory latency is that important, one would AFAIK have to compare to Zen3 Threadripper, because EPYC (and afaik Threadripper Pro, the other 8-channel variant of Zen3) is rather locked down in regards to memory "overclock"s. (Note that this is more of a latency/efficiency tradeoff, an any moderate OC is trivial to cool.)
- dragontamer 5y ago> 8 channel DDR4-3200 RDIMM (204.8 GB/s theoretical ceiling) But the M1 Max is 8-channel LPDDR5 400GBps, or literally twice the bandwidth of that EPYC
- rsynnott 5y agoPer the article, it can't use more than about 250GB/s for the CPU, though, so it probably isn't an unreasonable comparison.
- gigatexal 5y agoIs anyone else finding it hilarious people are comparing these mid-tier consumer chips to server grade EPYCs? :rofl:
- dragontamer 5y agoEPYC almost certainly has more compute power though. Honestly, the memory bandwidth is imbalanced. Its mostly there to support the GPU, but the CPU also gains benefits from it. Its hard enough to push an EPYC to use all 200GBps in practice. EDIT: For workstation tasks however, 64GB is huge for a GPU, while 400GBps is huge for a CPU. Seems like win/win for the CPU and GPU. Its a very intriguing combination. GPU devs usually have to work with much less VRAM, while CPU devs usually have to work with much less bandwidth. 64GB is small for CPU workstation tasks however. Its certainly a strange tradeoff.
- hajile 5y agoThe article in question shows that a mere 8 big cores and 2 little ones can use 243gb/s. I'm guessing Apple will go with HBM3/4 before too long due to the lower power consumption and great performance.
- gigatexal 5y agoI thought HBM was a power sucking tech?
- dragontamer 5y agoHBM is very low-clock speed and super efficient. HBM's downside is that it requires many, many, many pins. Each channel is 1024-pins of communications (and more pins for power). In practice, the only thing that can make HBM work are substrates. (Typical chips have 4x to 6x HBM stacks, for well over 4096 pins to communicate, plus more pins for power / other purposes) But HBM is among the lowest power technologies available. Turns out that clocking every pin at like 500MHz (while LPDDR5 is probably a 3200 MHz clock) saves a lot on power. Because DRAM has such high latency, the channel speed is more for parallelism more so than anything else. (DDR4 parallelizes RAM into 4-bank groups, each with 4-banks. All 16 can be accessed in parallel across the channel). HBM just does this parallel access thing at a lower clock rate, to save on power. But spends way more pins to do so.
- AnthonyMouse 5y ago> So I'm curious to know how an X64 chip would perform if we even the playing field not just in node size but also if Intel and AMD would adapt their X64 designs from the ground up with a memory controller, cache architecture and instruction pipeline tuned for that kind of fast RAM. We can get a pretty good idea about this by looking at Threadripper, which has more memory channels: https://www.anandtech.com/bench/product/2842?vs=2666 https://www.anandtech.com/bench/product/2842?vs=2666 (This is Zen 2 Ryzen vs. Zen 2 Threadripper because Zen 3 Threadripper isn't out yet.) In nearly everything it's about the same, because most workloads aren't memory bandwidth limited. But then you get into the multi-threaded SPEC tests where the M1 Max is doing really well, and Threadripper does really well. And this is still with less memory bandwidth than the M1 Max because it's using DDR4 instead of LPDDR5. The lesson here is that memory bandwidth limited workloads are limited by memory bandwidth.
- fomine3 5y agoFor Ryzen, its Infinity Fabric clock speed is linked to RAM clock speed, and Infinity Fabric speed is important for performance. It's not same thing as that RAM bandwidth is important.