Gadgetory

the g.skill ripjaws km 570 gaming keyboard features genuine Cherry MX mechanical switches customizable perky lighting effects and macro support full and key rollover and more click the link in the description for more details excellent so last week I attended an AMD press event as many of you deduced by my tweets and/or a particularly insightful conspiracy thread on the AMD subreddit while much of the information shared there still remains under embargo today you'll be hearing about Radeon instinct a new suite of hardware and open-source software offerings designed to dramatically increase performance efficiency and ease of implementation of deep learning workloads and although there is likely an issue of my audience who drools over news about machine learning hardware and what it means for our future robot overlords the book of you probably want to know what the hell this has to do with Vega the code name for Radeon technology groups long-awaited next-gen GPU family to explain that let's first figure out what the heck Radeon instinct is so basically there's too much data it's generated everyday by everything from social media posts to traffic sensors and coming through it is impossible without networks of big fast computers and E's new solution for this is Radeon instinct represented by a hardware hardware software combination of their new Zen based naples server cpus radeon gpus and ro CM Radeon Open Compute for deep learning neural networks going into any depth about this type of thing is beyond me so I will be linking to several articles in this video's description for anyone who wants to read a little bit further about that PC perspectives write-up is also linked down there because I Orion shroud a shroud out TM for inspiring the wild theories I'm about to share so since GPUs can process massive quantities of parallel operations they're ideal for handling tasks that provide huge amounts of data for example analyzing global financial markets helping the AI on an autonomous vehicle deal with sensor input or hyper realistic 3d real-time renders of ray-traced environments aim D introduced three GPUs today that they have renamed accelerators to align with their intended tasks and side note there all passively cooled because they're made to live in server racks that have their own loud-ass fans the Radeon instinct mi6 has 5.7 teraflops of compute performance 16 gigs of memory 224 gigabytes per second of memory bandwidth a 150 watt TDP and is Polaris based so you could equate it to NR x 480 on the gaming GPU side apart from it having twice the memory and a bit less memory bandwidth the Radeon instinct mi-8 has eight point two teraflops of compute performance four gigabytes of HBM memory a 175 watt TDP and AMD tells us quite directly that it's based on their VG GPU which is the same one that's used in the r9 fury and nano finally there is the Radeon instinct mi-25 with a 300 watt TDP but no compute performance listed or any other juicy specs really apart from one huge one it will use and these next generation high performance Vega GPU architecture so let me be very clear here that we're now moving into the theory crafting and speculation part of this video and to begin let's assume a couple things one is that there's probably only one actual Vega GPU the chip itself the Radeon is almost ready to launch and to is that the rack mount solutions that AMD has also showed us today have somewhat accurate compute performance listed my suspicion which mr. Shrout further validated at breakfast the other day was that if these two things are correct then it's a simple matter of division to get a ballpark idea of Vegas compute performance these racks are truly stacked with variants in the 100 and 400 teraflop range as well as the ridiculous invent tech rack with three petaflop of GPU compute by way of 120 mi 25 GPUs so let's do math and remember that Nvidia is $1200 Titan X Pascal card boasts floating-point computer formats of 11 teraflops if the invent tech rack can push three petaflop s-- that's 3000 teraflops divided by 120 mi-25s that's 25 teraflops that's that can't be right what about the other X the hundred teraflop one has four mi-25s that makes 25 teraflops per card the Falcon which has 16 mi-25s for 400 teraflops also 25 per card I mean it's called the mi 25 the 25 is right there the mi6 and mi Linus okay let's not rush to conclusions I present you also this shaky footage of Roger Cordrey at the press event actually showing us an MI 25 card and it looks big and also possibly dual slot if we zoom and enhance there whereas the mi6 and mi maybe be a dual GPU card with two of the Vega GPUs in it that would still be 12.5 teraflops per GPU though which would still seem like more than the Titan XP until you realize that the instinct cards are referencing FP 16 values whereas the Titan XP is measuring FP 32 computations so basically you cut the numbers in half for the instinct cards and you get roughly the equivalent so my dual GPU Theory probably goes out the window in that case 12.5 teraflops tho and FP 32 mode is still more than the type next P again this is speculation floating point rock compute performance doesn't always mean good frame rates in games we don't have launch dates for anything we don't have full specs and I could just be flat-out wrong but anyway one last thing to note AMD says that Radeon instinct products are expected to ship in the first half of 2017 so if the mi-25 is among them we could assume vega GPUs will be ready and available in that time frame as well but who really knows for sure anyway guys that is all for this video I hope you've enjoyed it I don't often do the speculation things so let me know if you thought this was useful for you at all of course there was more juicy information from that event but I can't share it with you yet so hit the like button and subscribe if you want to see those videos when they come out also check out my store link down there if you want to buy some merchandise and as always thanks for watching