Gadgetory

hey everyone I am at an Intel event joined by it's the chief core architect Roenick single and we're gonna be talking about a couple of topics say one of the most important is the new sunny Cove CPUs and just to give you some background Roenick here worked on halen previously so you led the performance focus of that yep and he's been working with Intel for a while now 22 years 22 years and the Hale um I think is probably close to a lot of you so anyone who who got their start with Nehalem like I did with overclocking this is the interview you'll want to see before that this video is brought to you by thermal grizzly and their high end thermal compounds thermal grizzly makes cryo not paced for high thermal performance and conductivity without being electrically conductive so you don't have to worry about shorting components cryo knot is particularly good for replacing stock GPU pastes as cryo knot is a non curing compound learn more at the link in the description below so let's start with a really top level thing here because no one at this point knows what sunny Cove is outside of this event so what what is sunny Cove meant will drill into it yeah thanks to you know sunny Cove is our next new microarchitecture so we've had the long line of products that we've built on the sky like microarchitecture at sky lake a B Lake and coffee Lake what we talked about today for the first time was detailing what we're gonna build on our next new microarchitecture like you said codenamed sunny Cove it's gonna go into a wide range of products standard PCs laptops desktops all the way to hours eons on the datacenter side and so what we talked about today is what are the big steps that we're taking with sunny Cove in terms of general purpose performance make the Machine more powerful as well as new instructions and new capabilities that we're adding to the processor itself to enhance algorithms like machine learning or cryptography right there's been a lot of advancement on the graphics front to you which is perhaps a different topic but some of the big ones you were talking on in your presentation so you had spectrum meltdown mitigations in hardware now we were talking about 14 nanometer improvements and how that's actually it gets brushed off a lot we've brushed it off because you see plus plus plus plus plus it loses the magic of it but to intel's credit 14 animator has had significant performance uplift right forget know and that's a great point so you know with Kove the first products with sunny Cove will be on 10 nanometer but in the meantime we do continue to advance what we can do on 14 nanometer both through what we do on the design and architecture side as well as on the manufacturing side so we've increased the frequency you know from the first skylights from you know 4.2 ish gigahertz now we're up to 5km oh yeah you know with some of the products on the data center side we're enhancing the architecture with new instructions for machine learning for both inference and training that will give us a really nice speed-up as well there so what it what is your biggest challenge when you're talking about before we even get into sunny Cove in in any detail what's your biggest challenge when you're trying to work with 14 animator and continue to push it yeah there's a thermal envelope is it what's going on yeah you know it's actually in some ways easier to work with amateur process technology like 40 nanometer because you have a lot of information about it you understand all about it and what its nuances are so really finding way to just continue to push the envelope on the performance especially through frequency is something we've been able to successfully take advantage of over the last year or so right and then on sunny Koh so let's let's touch on this one for you can't talk frequency today right so just get that out of the way right now can't talk overclocking today what can we talk i we can talk about a yes start you know where we'd start today is to say you know what is this is the base foundation the microarchitecture for the products that will come later on right and like I said those products you'll see desktop products you'll see PC products and so forth but the base capabilities does give us the fundamental throughput advantages to give you higher performance in addition to wherever the frequency ends up with these products so how do I increase the smartness of the machine building new algorithms how do I increase the number of operations I can do every clock right just that helps any of the applications that are out there whether it's games or multimedia applications or you know name your favorite application really so our focus is speed those all on with with IPC what do you look at to improve IPC at this point yeah you know direct you to ways to answer that first of all when we're doing our analysis what kind of workloads do I look at to say how do you know how do I tell if I'm actually increasing performance so we look at a lot of games we look at a lot of multimedia applications we'll look at a lot of things all on the server side of things because we're building a core to span that wide range and we want to choose advancements that help everybody across that but there will be things that are more specific to say a laptop part or a desktop right right that we do what are the kinds of things we do we do things like expanding the ability to find more instructions operate internal we do things like making our caches bigger one of the things we talked about today is we're increasing the first level data cache size by 50 percent yeah yeah that was a big one so what let me let me interject here for for people watching what if they leave a comment say well why didn't you do that before yeah and you know the first level data cache is one of the most fundamental pieces of your microarchitecture and it's always a really tricky balance between capacity and latency and bandwidth right and so we've really been focused on trying to maintain the latency as much as possible and really expanding the bandwidth over the last few generations and if you go back to what we were doing you mentioned a handle me use that as a reference point I could do one load one store at a time and there were 16 bytes wide now I can do today in skylake two loads one store and I can do them up to 64 bytes wide what we talked about today is now we're doing two loads and two stores and we're increasing the capacity and we're maintaining the latency so it's really how do you balance those three vectors of latency capacity and bandwidth right that makes sense speaking of balancing things with these security mitigations in hardware suspect or meltdown big topic last year so now where do you how do you deal with with spectrum meltdown getting thrown into the mix of things I would imagine the the threat there is you might lose some performance in exchange for preventing those attacks so how do you write vent them and so it's a great question I mean Spectre in meltdown was you know something that over the past year has been a lot of focus and attention from us both for our existing products as well as for our future products so you know we now have products shipping today that have Hardware mitigations for some of the cases like the Spector v3 or sorry they're all done right you know and that's what happens in whisky Lake we'll have coming soon on the data center side and later on on the client side mitigations for what was called Spectre v2 the branch target injection case all of these are things where we now have a chance to harden the microarchitecture put in hooks in the heart and the microarchitecture and like you said provide the protection while bringing back the performance so what how how long did you personally have to work on tackling this product problem I think they the researchers exposed it to Intel AMD all the other affected parties at least six months ahead of public announcement right that's exactly right so you know I was working on that starting late last summer basically I mean even when I took my summer vacation last year while on summer vacation every day was violent that's a great thing that you know it was something that it was that important to us and it was something we had to invest in and we we understood that you know the landscape had changed in some ways because these were new vulnerabilities that had these researchers did a really good job in identifying right so working with them working with others in the ecosystem that's one of the really good things that's come out of this is working with all of our software partners to understand what can be done across the ecosystem and then let's see so for sunny Cove you had your three topics there are security at the beginning you talked about right deeper you talked about a wider yep a smarter smarter right so we went through a little bit of those already but when you're talking about wider what are you looking at for sunny wider you know part of it is I can I now have more functional units you know we already talked about for instance I can now do two loads and two stores right at a time we also introduced for instance on the vector side the ability for certain operations where we can do one per cycle now we can do two per cycle in some cases so if those are the kinds of examples of making things wider just finding opportunities to do more work at the same time does this increase the package size in a meaningful way or do you just cram more into the same well so remember that the first products with sunny Cove will be on ten nanometer so we get the advantage of the shrink factor by moving to the new process technology so again the overall package size will be a function of the product in terms of number of cores and so forth so right we'll have more of those details right let me throw two here that are kind of general for you so let's let's talk about a AVX a little bit I came up briefly today first question for you this is not specific to sunny coat but just Intel in general and and talking thermals okay so for a BX everyone pretty much knows prime blender as a real world example they really burn the CPU why is a VX so much hotter than any other application okay yeah and so a VX is simply processing in general is interesting right if you're able to get take good advantage of these instructions it's actually very energy efficient because there's less instructions going through the machine to do the same amount of work however at any given point in time instead of doing two multiplies or four multiplies I'm doing more of them at the same time so at any given point in time you actually are doing more work in the processor and more work takes more power right right so there are cases like you're saying where you will see power go up as a result of a V hmm yeah and we saw as a recent example we retested the 2600 K Sandy Bridge and the performance difference there in blender versus modern CPUs was nonlinear compared to for example a 7700 k7r k99 RK which part of that is of course cores than part of it I think is the AVX capabilities yeah so there's there's a significant performance benefit you know and that's one of the things we always tell people is as we add new capabilities and new instructions there's a lot of opportunity for performance that people are tuning their software right in AV X is a great example of that right right and anything you want to plug here for sunny Cove at the end I mean that's what's really interesting you well to me you know each of these projects that we work on whether it's an inhale amor a Sandy Bridge or a sky like it's a long journey and to get to the point where it you know we're about to get it into you know your your customers hand right you know is an exciting time just to see what people think of it right because we're very excited about all the work were poured into it how long does this process take it's you know it's a it's longer than you know then building a cake or building a bed years right yeah it's even just taping out yeah well it's um it's a several year development cycle on these that's right well for more information check the description below we'll have a link to one of our articles or something like that and thank you for joining me thank you we'll see you all next time