Gadgetory

hello I'm Gary Sims from Andros Thoreau's now this is the second video that I'm doing this week last time I covered dynamic the new processor architecture from arm and today we're going to be looking at the new high performance processor for arms called the cortex a 75 so what is the cortex a 75 and what does it bring to the table well let me explain now the court in a 75 is a processor that can be used inside of a new dynamic cluster and that's what the last video was about and it could be mixed and matched with the new cortex a 55 which will be in my next video now the a 75 is a high-performance processor and it's inspired by and take a lot from the cortex a 73 but it's more than just a tweak there's a whole bunch of new stuff in here that have really made this a great processor so let's get down to performance what do we see well according to armors at 22 percent performance increase so compared to the a 73 and will you translate that also just things like floating-point numbers when you look at things like memory bandwidth overall in protein their benchmark like Geekbench you may even see up to a 34% increase in the performance scores that you get from the benchmarks but the key thing about the a 75 it borrows the ideology from the a 73 and then it can do sustained performance maybe remember the a 73 was a great processor because unlike previous processes it was capable of making a peak performance but its peak performance was actually the same as it's sustained performance so what that meant is that when we get previous SOC s usually made me run a benchmark on it and it gives you this great score but if you ran that benchmark for a long time then actually the processor performance would start to dive because it would start to heat up and to maintain that heat dissipation specification so the phone doesn't burn your hand and so that the process they can get rid of that heat it would have to lower its performance now the a 73 and now the a 75 has this idea that the peak performers and the sustain performance are almost exactly the same thing which means that although we might just get good scores in the benchmark it means as you're using the phone for long periods playing games or your web browsing using all through the day you're actually gonna find that performance remains the same and that's one of the great things about the 873 and it's also a great thing about the a 75 now as I mentioned in the dynamic video because we've now got this greater single straight performance there's also the opportunity for SOC makers to put in a single core tech 875 core amongst a whole bunch of cortex 855 cause maybe even in a seven plus one configuration or in a two plus six configuration that will still hack see have a very good silicon footprint which means the costs will be good but actually you'll get that great single thread performance which really helps you eyes and certain types of applications still much more responsive here look at some of the micro architectural changes that we find in the a75 compared to the a 73 the first thing to note is that we've gone from a two way superscalar process with what the a 73 was to a 3-way superscalar processor of course now we've got the l2 cache is per core just because that's what is defined now as part of dynamic and that also be true in the cortex 855 of course it can be used in dynamic and of course it comes with access to the l3 cache which is across the whole cluster for those who are into the l1 cache is now a 64k 4-way associative cache and there is a very good low latency on getting data from the l2 cache the eight mt5 also had a state-of-the-art branch predictor which enables it to sustain the performance and keep those instructions flowing down the pipeline even when there are tight loop situations now branch bridges are a very glamorous part of SOC design but basically if you think about it like this when a program is running the CPU is fetching the instructions from the memory is the coding and the mem is executing them and at some point if the program jumps off to different completely part all those bits that have been fetched and started to be decode previously are now invalid because a program is now running somewhere else and our branch predictor works out where the programs jump to next and make sure the instructions are coming down the pipeline ready to execute even when the program has jumped off now the a75 has got a new Brock predictor it's an improvement over one that's in the a 73 and that helps improve performance and keeps those instructions flowing down the pipeline now the a 75 has got seven independent issue cues now what that basically means is that different types and instructions can be run in parallel so if you're loading something from one part of memory and the next instructions are loads of meals from another part of memory well actually you can do those at the same time because you're not going to use both of them until maybe the third instruction so they get happen at the same time and there are several different cues from what can be executed at the same time again that improves performance now one of the things we hear a lot about today of course is AI and neural networks now a lot of these things are not using long complicated floating-point numbers they're actually finding that you can run these neural networks using much less lower precision 16-bit floating-point numbers and in fact there's even some types of dot products as vector multiplications that can happen in just an 8-bit integer and so armored than a lot of work to improve the performance of 16-bit floating-point numbers and all these integer 8-bit integer dot products so that as we move more and more into using your networks in our software as we use more and more seeing AI assistants in our mobile phones the underlying hardware is able to run the instructions needed to make those more efficient and so there you have it the cortex page 75 is a new high-performance core from arm that will be used inside of a dynamic cluster along with the cortex a 55 you've got a great improvement in the overall performance maybe as much as 34-percent in a benchmark like Geekbench you've got a great single thread performance increase which means we might see this use in different types of combination no longer just like four plus four combination but maybe a one plus seven or we might see it in a different kind of configuration maybe a five plus one we'll have to see that lots of options for CPU makers to use we of course we've got the new the continued sustained performance along with that increase in performance and cause there have been improvements throughout the microarchitecture including the stuff we need for neural networks and for artificial intelligence I'm garrison reminder authority I hope you enjoyed this video this is the second of four videos I'm doing this week the next video will be about the cortex a 55 so make sure you subscribe to our Channel hit that Bell icon so that you get a notification whenever we publish a new video and last but not least do go over to Andrew AFRICOM because we are your source for all things Android