As a CG artist it's wild to think that a single-socket server with one of the top-end Turins in it would absolutely wipe the floor with the 200-socket render farms used on the movies I worked on 15-18 years ago comprised of 2 and 4 core Opterons!
I believe Blackwell gpus do all the legwork now. Unless you’re running a software that only allow for cpu rendering only which is probably true in some engineering and medical science fields.
@@Aggrofool oh yeah, gpus still have so low memory capacity for big simulations, or huge scenes, even with out of core tech. But they are catching up, i thing some high end gpus came with a decent 98 gb.
Not to ruin the obvious joke but LTT tried it (very badly and drew wrong conclusions) and windows only allocated like 64 logical cores (threads) on the 96core. So this 192c/384t cpu would be limited by windows unless you run 6x separate instances.
@@aidanbyrne8267 sometimes it's important to reiterate what words like "LTT" mean, in case readers were born yesterday. The world is unfortunately very forgetful of corporate incompetence and trust violations.
@@evalangley3985 We have it on fairly reliable authority that the absurd delays of Luna Lake and Arrow lake were getting OEMs pissed... and then the 13th/14th gen fiasco started rolling where Intel initially pretty much told the OEMs to just suck it up. If you look at the number of AMD systems Lenovo offers now, compared to a year ago, the shift has happened... AMD just need to beg TSMC for more capacity.
@@ktfjulien Yeah many server cycles are around 8 to 10 years so the next batch of server upgrades will be the next couple of years (2018-2028) and everyone's seen how good they've been. 6 years of steady improvement and field testing means that there is no longer that question mark of what to expect from epyc etc. Also techies talk to each other at the pub. Managers also.
I can immediately think of at least two clients where this configuration could reduce an entire farm from 4 units to 1. When we talk about how much power gets used, we have to also take into effect if you can reduce rack space and literal running racks. Reducing down several separated? Oh yeah. I can see CTOs drooling. The huge benefit I'm seeing in regards to mass virtualization is that this changes the entire outlook for possibilities on virtualized remote workstations instead of committing to Remote Desktop platforms. We've been experimenting with this as a way to reduce cost with a few clients, but you make it where you wipe out some of the double licensing cost, provide high level speed and performance, and you make it work.. I am incredibly eager to put a few of these into the field!
Yeah, the fact that we have hit a point where we aren't all demanding more power, or more does nothing for the end user. I can see offices running virtual machine only dealing with office products, seeing a huge benefit from these systems. More processing power isn't going to make a virtual machine running office products or data entry software run faster. However being able to decrease the number of racks does reduce cooling and running costs. Maybe allows smaller racks to placed in other site locations that might not have been possible due to needs of a full rack to run the hardware needs.
Lisa Su is a great Leader. She took this dying company and made some big right decisions. Now AMD competes with intel and 5 years ago noone could even dream of it. Dr. Lisa Su is a great leader, mutch respect.
They aren't even wrong... based on what Intel dropped today about arrow lake. Zen5 x3d is going to dominate Intel on the desktop as well. Also ZEN5 desktop systems, should be able to slot Zen6 when it comes. Intels mobo is going to be a one and done affair as always. Having seen now what Intel is cooking for consumer mobile and desktop. Combined with the full on stomping they are going to continue to take in big iron sales. I think at $21 their stock might be over values. :/ Once the fab is split... someone is going to scoop up Intel design, the vultures are circling.
x3d cache for all of the ccd chiplets would be fenomal in performance. combine that with some hbm memory onboard the same carrier board and coupled with dd5 rams. and relase this with reasonable pricing
@@chaddesrosiers1107they proved they can pivot to a 3rd party fab and got billions from the US government. It's a blue chip stock that will continue to be a blue chip stock
500 watt / 16 chiplets ~~ 30 watts per chiplet, with the IO node cost amortized. And amd has "juiced the volts" up a bit this time, since intel actually managed to produce a somewhat competitive chip with a process advantage. But if amd is strongly competitive as it is, they can merely play the process boost at a later convenient point.
EPYC Genoa has a new IOD and IF. EPYC Turin has again improved the IOD and IF. Meanwhile, Ryzen 9000 reuses the slow IF from Ryzen 7000 which was already worse than that of Zen 4 Threadripper / EPYC.
@@kevinerbs2778 It could, but the 7000 series TR already had much better IF and IOD than the AM5 Ryzen 9000. It's just a shame that AMD keeps reusing the same parts that are already the bottleneck on Ryzen 7000.
Zen 5 for desktop is all about making Intel's power consumption stupid. More and more corporations and institutions have implemented green policies. Heck, EU funds are being allocated to help state entities improve their computing infrastructure. AMD is fine with catching flak from gamers and enthusiasts if it means they can get performance parity with Arrow Lake at half the power or less. Because that's what's going to sway the purchasing decision for millions of boring office beaters.
The IF link speed has increased one server but is fundamentally the same beyond that between desktop and server. Going faster on desktop may not have made much difference if it is not a bottleneck in the system.
On desktop, my understanding of Zen 6 is getting away from needing an "infinity fabric" because of moving to direct connects between die. That ALONE should be about 10% IPC over heavily threaded apps or systems that are running many processes at the same time. I haven't considered Zen 4 - 5 because I already heard Zen 6 should be a serious MCM overhaul using direct connects and that's what I want to upgrade to.
As i had just to my wife this is a fucking nuke was only a few years ago 92 cores was insane then 120 then this jump comes along and jumps the fucking entire market with nuclear ferocity....
What absolutely baffles me, 192 cores, 500 watts. That's about 2.5 watts per core, running at 3-4ghz. About. I'm not sure how much other components are sucking down, interconnects, dram controller, etc. That's mind blowing to me. I remember a single core sucking down 100 watts and overheating.
Yeah that's where the real money is (and AI) . Consumer desktop customers might be extremely vocal and demanding but reality is, AMD doesn't earn much from consumer desktop sales .
FWIW, "Turn, Turn, Turn" was written and originally performed by Pete Seeger. But the Byrds version was not only more melodic than yours, it was more melodic than his, too.
Great video Wendell! Nice to see your enthusiasm. The love letter metaphor is fitting. Might have to get me some of these new rippers for my long postponed upgrades. Thanks for the updates. While watching the video on my porch, i was thinking to myself... There is a small percentage of genpop who would understand the avalanche of info disseminated... quite literally riffed out in this video. Add to it the clear passion, expertise and excitement... Makes me feel a little smug to be in the small percentage group 😆 . Wendell, you really eloquently packed and unpacked way more info than is otherwise humanly possible for a video of this duration. Did anyone notice there were moments when Wendell's brain had a challenge buffering, queueing and context switching the information he was processing and then also divulging to the audience? Made me smile. Engaging content. Kudos.
I was listening to your conclusion. You are correct that these CPUs are meant for composable systems where you have a high core count, lots of memory, and then use the local disk for spill. Then you use high speed networking and create an NVMe fabric married w slower denser (cheaper) SSD/Disk for a tiered storage solution. Makes a lot of sense and this is why AMD is the front runner for server consolidation and potentially replacing Intel during the refresh cycle. Looking at current workloads... if you have 2TB of addressable memory.... you could in theory support 256 cores.
I’d expect it to perform rather terrible in many games as they achieve such a high L3 cache amount by only having 1 core active per ccd. You’d be constantly hitting the penalty for ccd to ccd latency if the game uses multiple threads. Entirely speculative so it’d still be interesting to see.
Why woukd it that? Only windows Vista did that ot would never pin a main thread f or a game ro a core/thread. It used to run the main thread through ever single core you had regaurdlesa of being a core or hyper thread. it was waful coding.
We will know when 3D V-Cache hits for Zen 5. Zen 4 saw a ~15% uplift from V-Cache. If Zen 5 sees a 25% uplift we know it's memory starved more than anything.
@@andersjjensenyes. I am believing that Desktop Zen5 parts are IOD bottlenecked since launch day. With so many architectural advances on the cores can't exist another explanation. Fortunately 8000 MT/s memory is a viable option now and should help to mitigate that memory starvation
Slightly off the topic, I thought I share the love. my Ryzen 7800X3D was hitting high temperatures and I tried many thermal pastes but not liquid metal. The BSFF from Nuomi Chemical, the non-silicon, non-conductivity thermal paste is amazing stuff. I was using the highly rated Noctua paste was running 5-6 C higher. I don't know what it's made from and they even claim it outperforms liquid metal. It would be great if we can have some proper tests on BSFF.
I 100% need 192 cores. Is that necessary for my music production? Nope. But is that so absurdly what I am going to put into my computer? Also no. But I need it anyways
@@righteousone8454 well thats most probably a software problem. i mean that $h!t still supports 4th gen icore with a minimum of 8Gb ram. when was the last time they updated there software and how it runs on newer hardware?
It's almost like adding more cores = more performance on multithreaded workloads. Honestly AMD should just make a 1024 core chip; that'd be better. I don't know why they don't.
Then the cpu size will be like half of A4 paper. I'm not complaining tho, just imagining. Also, maybe the latency will be an issue there, because the chiplets will be spreaded far and wide. Unless they can create something like 128cores per chip, with reasonable size.
Nice CPUs you got there, but how does it overcome the biggest hurdle? Multi-year Xeon contracts for single percent discounts for corporate purchasers !!!
9:11 The dual GMI3 interface is something AMD has had on Zen 4 and Zen 4c chiplets but never fully leveraged in an actual product. The previous IO die had 12 GMI3 interfaces arranged in groups of 3 and a CCD could not have different GMI3 link split across an IO die group, thus limiting any benefit to four CCD and below models (32 core max using Zen 4). This group rule likely still exists but it does not really matter now that Turin's IO die supports four groups of 4 GMI3 links and thus able to get to 128 cores (eight Zen 5C dies). Per the recent die shots of a Zen 5 chiplet, it too has dual GMI3 links, though unused on the desktop.
I've been on the client side & SCCM side of things for too long. I'm curious about which of these new chips would be the best fit for a non-clustered database server , assuming it was paired with the appropriate network and storage subsystems? Great vid! Love the Encounter at Farpoint reference at the end. 🖖🏾
Man I can't wait till these get 'cheap' enough secondhand for me to throw one of these in the homelab. It will probably be a while before I get to play with these CPU's but I'm so looking forward to it.
Since most of us use OLTP, I keep wondering what a PostgreSQL benchmark might look like (I miss AnandTech and their database benchmarking). As for hosted LLMs, I don't see the reason to run GPUs on server for inference when there's services like Groq that are super cheap. Would love to hear other use cases.
That's the default background to the Essential Phone!!! I had one of those, very sad that they cancelled them. I still use that background on my phones to this day, it's really nice
Currently, vendors known for ISV-certified workstations such as Dell Precision, Z by HP, and Lenovo Thinkstation do not offer models powered by Epyc, only Threadripper Pro.
One thing missing and why I generally distrust Phoronix benchmarks for HPC results is that compiler optimizations are rarely or improperly used. Can you maybe try comparing the Granite Rapids using the intel oneAPI compilers against the AMD Turins using the GNU compilers? In particular using -xhost and -mprefer-vector-width=512 or alternatively -qopt-zmm-usage=high for an older "classic" compiler.
Most of the customers I support, don't need 128 cores. It's more like 8 - 16, but A LOT of Ghz and IPC, since some of their applications are ancient and run on like 1 or 2 cores. I mean 9175F at 4.2 / 5.0 is impressive, but even in a VM Environment your typical "small to medium business server" - 12 out of 16 cores are are going to sleep most of the time no matter the OS as Host, because the base of Software like Microsoft Office and "Your Average Business Solution" was coded in like 1999 and just got more features put on top of existing "legacy stuff that is still in there because of compatibility" and I bet "small to medium business servers" are probably 75% of the market at least in "numbers sold".
I liked the poem. Comic shot with star wars score in the end. That heatsink duo is cool. And I don't mean just its ability to cool how much wattage? 500W Cpu that 's crazy... 3x more than my PC while gaming. But that many cores doing alot of usefull job. I would like to see that powerdraw per core on desktop cpus.
About the Desktop vs. Server memory: At least in HPC, this feels like it's been a long time coming: Check out Fugaku's setup from 2020, with HBM2 as main memory!
One loasd I'd love to see is some hardware simulation stuff with something like vcs, but afaik they don't really do benchmarks and while you could totally do it on real designs, the licensing for this get pretty crazy. When we run regressions it typically takes multiple days for some projects, running just a few ms of the big SoC
I don't watch Level1Tech all that often compared to the Steves and opening the video almost gave me a heart attack since I'm used to an intro card. We call that "mit der Tür ins Haus fallen" in German (to fall with the door into the house). How about a greeting first?
That 9175F with the 16cores and 512MB of L3... really stands out from the lineup. I wonder what industry wanted such a massive L3 for so few cores. High Frequency Traders maybe?
Long time watcher but, I don't ask for much. Any chance your escapades cooinside with Revvit? My friend is an architect working on large scale projects but, of course he (they) are warry of going deep on tech.
Hi Wendell, I am building a 9950x precisely to use simulation though I can't afford what you got there. I am a user of IRONCAD with AMPS Technologies and venturing to Flowsquare but trying to make a work around on meshing at it only supports Inventor. Almost all hardware is in, just waiting for the processor itself due to stock delays.
Any info regarding Zen5 based threadripper for workstations? I was thinking on buying a Zen4 7000 series TR but with Zen5 release I opted for a 9950x now and wait for whatever AMD will release on Threadripper or 3DCache versions. Maybe the 3D cache can accelerate my mostly Java based workloads as the JVM is very cache size bound i believe
All it would take is premium SFX motherboards to complete the workstation dream with this generation. Though a WX competitor for the 5090 would also be welcome.
I would love to see the Epyc 9575F paired with a 4090 (or 5090) in action on cyberpunk 2077, final fantasy 16 or black myth wukong. I would ask please to test on cpu-z too if possible.
We have so many benchmarks!
forum.level1techs.com/t/amd-epyc-turin-is-out-zen5-9000-benchmarking-w-comsol-phoronix-test-suite/218260
Thanks for all your hard work, Wendell! Great content.
this dude could stand to lose 150 lbs lol
How about the 6400DDR? possible?
As a CG artist it's wild to think that a single-socket server with one of the top-end Turins in it would absolutely wipe the floor with the 200-socket render farms used on the movies I worked on 15-18 years ago comprised of 2 and 4 core Opterons!
Do render farms still use CPU rendering?
I believe Blackwell gpus do all the legwork now. Unless you’re running a software that only allow for cpu rendering only which is probably true in some engineering and medical science fields.
Hey, bro, anyone ever tell you that you look like Nick Diaz's hypothetical older brother? That picture had me double take until I zoomed in
@@Aggrofool oh yeah, gpus still have so low memory capacity for big simulations, or huge scenes, even with out of core tech. But they are catching up, i thing some high end gpus came with a decent 98 gb.
@@Aggrofool We need both sir. GPU and CPU.
depend on how much memory it need to render or how complex of the scene and shader.
192 cores? We need Cities Skylines 2 benchmark on that thing.
Not to ruin the obvious joke but LTT tried it (very badly and drew wrong conclusions) and windows only allocated like 64 logical cores (threads) on the 96core.
So this 192c/384t cpu would be limited by windows unless you run 6x separate instances.
@@tomstech4390You don't need the "(very badly)" caveat.
Saying it was LTT is enough 😅
@@aidanbyrne8267 sometimes it's important to reiterate what words like "LTT" mean, in case readers were born yesterday. The world is unfortunately very forgetful of corporate incompetence and trust violations.
@@tomstech4390 Cities Skylines 2 runs on Linux too and quite well.
@@benjaminreynolds3659 Paradox has good Linux support for all their games. But I doubt Skylines 2 can be salvaged even with 2x 192 cores.
amd has become a execution monster. gj lisa from 2% 2018 to 34% in 6 years is craaazy
After Genoa they should be >50%. Let's just wait for the next round of server refreshes to come in...
Honestly they outdone themselves several times. Turin is a masterpiece.
@@ktfjulien Yeah, if the competition was not using monopolistic strategies to prevent AMD from gaining that market share...
@@evalangley3985 We have it on fairly reliable authority that the absurd delays of Luna Lake and Arrow lake were getting OEMs pissed... and then the 13th/14th gen fiasco started rolling where Intel initially pretty much told the OEMs to just suck it up. If you look at the number of AMD systems Lenovo offers now, compared to a year ago, the shift has happened... AMD just need to beg TSMC for more capacity.
@@ktfjulien Yeah many server cycles are around 8 to 10 years so the next batch of server upgrades will be the next couple of years (2018-2028) and everyone's seen how good they've been. 6 years of steady improvement and field testing means that there is no longer that question mark of what to expect from epyc etc. Also techies talk to each other at the pub. Managers also.
I can immediately think of at least two clients where this configuration could reduce an entire farm from 4 units to 1. When we talk about how much power gets used, we have to also take into effect if you can reduce rack space and literal running racks. Reducing down several separated? Oh yeah. I can see CTOs drooling. The huge benefit I'm seeing in regards to mass virtualization is that this changes the entire outlook for possibilities on virtualized remote workstations instead of committing to Remote Desktop platforms. We've been experimenting with this as a way to reduce cost with a few clients, but you make it where you wipe out some of the double licensing cost, provide high level speed and performance, and you make it work.. I am incredibly eager to put a few of these into the field!
Yeah, the fact that we have hit a point where we aren't all demanding more power, or more does nothing for the end user. I can see offices running virtual machine only dealing with office products, seeing a huge benefit from these systems. More processing power isn't going to make a virtual machine running office products or data entry software run faster. However being able to decrease the number of racks does reduce cooling and running costs.
Maybe allows smaller racks to placed in other site locations that might not have been possible due to needs of a full rack to run the hardware needs.
Did Wendell just hit us with slam poetry ??
Never change Wendell
Wendel be Turin into a poet!
Agreed, never change Wendell. However, maybe don't ever do that again!!! 😂
You guys realize he was quoting a song, right?
He rewrote the text to a Byrds song
I feel so out of place watching this videos knowing that I don't even work in IT much less next to a datacenter, but I love this!😂
The Zen5, 16 core 9175F @ 6Ghz and 512 MB Cache looks amazing. I would virtualise Davinci Resolve, and render out on that type of server.
*16 chiplets, 128 mouth watering cores.
I was looking at that too, it looks fantastic value performance wise.
I am so curious about its core configuration. Is it 1 core per CCD, or something with 3D cache?
Ia that the one that says 768mb of total L3 cache?
Lisa Su is a great Leader. She took this dying company and made some big right decisions. Now AMD competes with intel and 5 years ago noone could even dream of it. Dr. Lisa Su is a great leader, mutch respect.
That's what Zen5/5c was made for, servers.
AMD will just x3D their chips for gamers and thats good enough (until intel move).
Exactly, this is what Zen 5 is all about, server dominance.
They aren't even wrong... based on what Intel dropped today about arrow lake. Zen5 x3d is going to dominate Intel on the desktop as well. Also ZEN5 desktop systems, should be able to slot Zen6 when it comes. Intels mobo is going to be a one and done affair as always.
Having seen now what Intel is cooking for consumer mobile and desktop. Combined with the full on stomping they are going to continue to take in big iron sales. I think at $21 their stock might be over values. :/ Once the fab is split... someone is going to scoop up Intel design, the vultures are circling.
X3d makes sense for some server workloads too but Zen 5/5c really shining for servers I totally agree with.
x3d cache for all of the ccd chiplets would be fenomal in performance. combine that with some hbm memory onboard the same carrier board and coupled with dd5 rams. and relase this with reasonable pricing
@@chaddesrosiers1107they proved they can pivot to a 3rd party fab and got billions from the US government. It's a blue chip stock that will continue to be a blue chip stock
500 watt / 16 chiplets ~~ 30 watts per chiplet, with the IO node cost amortized. And amd has "juiced the volts" up a bit this time, since intel actually managed to produce a somewhat competitive chip with a process advantage. But if amd is strongly competitive as it is, they can merely play the process boost at a later convenient point.
You underestimate how much power the IO die uses, it's around 80~100 watts under load.
@@gatocochino5594 That's why I said it is amortized across the cpu.
EPYC Genoa has a new IOD and IF. EPYC Turin has again improved the IOD and IF.
Meanwhile, Ryzen 9000 reuses the slow IF from Ryzen 7000 which was already worse than that of Zen 4 Threadripper / EPYC.
x3D yer gaming!
Lol that means 9000 series threadripper could get the better i/o die.
@@kevinerbs2778 It could, but the 7000 series TR already had much better IF and IOD than the AM5 Ryzen 9000. It's just a shame that AMD keeps reusing the same parts that are already the bottleneck on Ryzen 7000.
Zen 5 for desktop is all about making Intel's power consumption stupid. More and more corporations and institutions have implemented green policies. Heck, EU funds are being allocated to help state entities improve their computing infrastructure. AMD is fine with catching flak from gamers and enthusiasts if it means they can get performance parity with Arrow Lake at half the power or less. Because that's what's going to sway the purchasing decision for millions of boring office beaters.
The IF link speed has increased one server but is fundamentally the same beyond that between desktop and server. Going faster on desktop may not have made much difference if it is not a bottleneck in the system.
On desktop, my understanding of Zen 6 is getting away from needing an "infinity fabric" because of moving to direct connects between die. That ALONE should be about 10% IPC over heavily threaded apps or systems that are running many processes at the same time.
I haven't considered Zen 4 - 5 because I already heard Zen 6 should be a serious MCM overhaul using direct connects and that's what I want to upgrade to.
LET'S F*CKIN GOOOOOOO!
EDIT: You sumbitch I can't believe there was an actual manual docking meme at the end. I'm laughing real hard at that one.
So, it just blows everything out of the fucking water.
As i had just to my wife this is a fucking nuke was only a few years ago 92 cores was insane then 120 then this jump comes along and jumps the fucking entire market with nuclear ferocity....
@@joshuagreer8046 AMD really came back with a vengeance with the zen lineup.
Essentially yes, this suddenly makes Xenons look like crap which will shake the markets violently
The stars in the eyes of Wendell 🤩 So much passion in that video is honey to my ears. 🔥
The opening and ending were *chef's kiss*
What absolutely baffles me, 192 cores, 500 watts. That's about 2.5 watts per core, running at 3-4ghz. About. I'm not sure how much other components are sucking down, interconnects, dram controller, etc. That's mind blowing to me. I remember a single core sucking down 100 watts and overheating.
I was glued to every second of this video, great coverage Wendell!
Hope this bodes good things for the next Threadripper.
Finally a bench mark in COMSOL, I love to see it and can completely appreciate how blazing fast those sims are.
So good to see Wendell excited!
AMD never disappoints in the enterprise segment.
AMD came to collect what's left of Intel's shares.
They earned and deserved them in 2003/4.
What's that smell? Oh, it's just the AMD server market share heating up.
AMD is gunning for that server and workstation market. Sounds like a real beast of a CPU.
Yeah that's where the real money is (and AI) . Consumer desktop customers might be extremely vocal and demanding but reality is, AMD doesn't earn much from consumer desktop sales .
@@djlim4612 Yeah many homes only buy laptops. x86 is becoming a business asset. Servers is where its at. AMD won.
FWIW, "Turn, Turn, Turn" was written and originally performed by Pete Seeger. But the Byrds version was not only more melodic than yours, it was more melodic than his, too.
Great video Wendell! Nice to see your enthusiasm. The love letter metaphor is fitting. Might have to get me some of these new rippers for my long postponed upgrades.
Thanks for the updates.
While watching the video on my porch, i was thinking to myself... There is a small percentage of genpop who would understand the avalanche of info disseminated... quite literally riffed out in this video. Add to it the clear passion, expertise and excitement...
Makes me feel a little smug to be in the small percentage group 😆 .
Wendell, you really eloquently packed and unpacked way more info than is otherwise humanly possible for a video of this duration.
Did anyone notice there were moments when Wendell's brain had a challenge buffering, queueing and context switching the information he was processing and then also divulging to the audience? Made me smile.
Engaging content. Kudos.
AMD killing it. Well done!
Imagine a 16 core CCD trickling down to the desktop Ryzen line... 🤔
Been imagined and named the ryzen 9975X3D if I am not mistaken 😂 idea was one 8 core x3d and one 16 core 5c
We can only dream, that with ECC and low idle for non 24/7 workloads...
oh wait yh they talked about that idea before on this channel right? @annebokma4637
@annebokma4637 which solves scheduling as c cores are slower. Hype train here we come
I can see the memory bandwidth starvation already!
I was listening to your conclusion.
You are correct that these CPUs are meant for composable systems where you have a high core count, lots of memory, and then use the local disk for spill.
Then you use high speed networking and create an NVMe fabric married w slower denser (cheaper) SSD/Disk for a tiered storage solution.
Makes a lot of sense and this is why AMD is the front runner for server consolidation and potentially replacing Intel during the refresh cycle. Looking at current workloads... if you have 2TB of addressable memory.... you could in theory support 256 cores.
I secretly hope someone benchmarks that 7175F with 512 L3 cache in a gaming workload.
mmmm cfd .... mmmmmm
I’d expect it to perform rather terrible in many games as they achieve such a high L3 cache amount by only having 1 core active per ccd. You’d be constantly hitting the penalty for ccd to ccd latency if the game uses multiple threads. Entirely speculative so it’d still be interesting to see.
Why woukd it that?
Only windows Vista did that ot would never pin a main thread f or a game ro a core/thread.
It used to run the main thread through ever single core you had regaurdlesa of being a core or hyper thread. it was waful coding.
Your saying AMD could have gone full nuclear on desktop gaming but did not because they kept their zen4 IOD.
The same thought exactly
yep
We will know when 3D V-Cache hits for Zen 5. Zen 4 saw a ~15% uplift from V-Cache. If Zen 5 sees a 25% uplift we know it's memory starved more than anything.
@@andersjjensenyes. I am believing that Desktop Zen5 parts are IOD bottlenecked since launch day. With so many architectural advances on the cores can't exist another explanation. Fortunately 8000 MT/s memory is a viable option now and should help to mitigate that memory starvation
@@andersjjensen ty, good explanation!
Slightly off the topic, I thought I share the love. my Ryzen 7800X3D was hitting high temperatures and I tried many thermal pastes but not liquid metal. The BSFF from Nuomi Chemical, the non-silicon, non-conductivity thermal paste is amazing stuff. I was using the highly rated Noctua paste was running 5-6 C higher. I don't know what it's made from and they even claim it outperforms liquid metal. It would be great if we can have some proper tests on BSFF.
I am little bit jealous about your server kind of knowledge!
AMD is currently demolishing Intel in overall performance and efficiency.
OK. That last scene was hilarious.
I 100% need 192 cores. Is that necessary for my music production? Nope. But is that so absurdly what I am going to put into my computer? Also no.
But I need it anyways
When you get them do post a video of it running cinebench. I need to see how pedestrian my 7950x workstation is in comparison.
@@xYarbxin multi it's probably hitting 180-200k or breaking cinebench
Meanwhile in Cubase with dozens of plug-ins: struggling to use 6 cores, let alone 8
@@righteousone8454 well thats most probably a software problem. i mean that $h!t still supports 4th gen icore with a minimum of 8Gb ram. when was the last time they updated there software and how it runs on newer hardware?
i want more cache. 1Gb should be fine.
But also cores.
It's almost like adding more cores = more performance on multithreaded workloads. Honestly AMD should just make a 1024 core chip; that'd be better. I don't know why they don't.
Then the cpu size will be like half of A4 paper. I'm not complaining tho, just imagining.
Also, maybe the latency will be an issue there, because the chiplets will be spreaded far and wide. Unless they can create something like 128cores per chip, with reasonable size.
@@ivrgn1720 LoL
Nice CPUs you got there, but how does it overcome the biggest hurdle? Multi-year Xeon contracts for single percent discounts for corporate purchasers !!!
An insurance fire.
i audibly laughed out loud (weird saying that) to the docking at the end. thanks, Wendell.
Not the last "manual docking" Riker performed during his time on the Enterprise.
@@rhekman I wonder if he ever contracted a "computer virus" from this kind of "manual docking" (or maybe one from Minuet?) :P
Can’t wait to buy one of these used for a 20th of the msrp on eBay in about 5 years or so
That end was so good, I had to check if I didn't accidentally hit play on my ST:TNG episode opened in the background :P
9575F? That's really hot!
Epyc Intro!!!
9:11 The dual GMI3 interface is something AMD has had on Zen 4 and Zen 4c chiplets but never fully leveraged in an actual product. The previous IO die had 12 GMI3 interfaces arranged in groups of 3 and a CCD could not have different GMI3 link split across an IO die group, thus limiting any benefit to four CCD and below models (32 core max using Zen 4). This group rule likely still exists but it does not really matter now that Turin's IO die supports four groups of 4 GMI3 links and thus able to get to 128 cores (eight Zen 5C dies).
Per the recent die shots of a Zen 5 chiplet, it too has dual GMI3 links, though unused on the desktop.
Would love it if someone could use AutoTune to create a mix of the intro to sound like the original song ^_^
I've been on the client side & SCCM side of things for too long.
I'm curious about which of these new chips would be the best fit for a non-clustered database server , assuming it was paired with the appropriate network and storage subsystems?
Great vid! Love the Encounter at Farpoint reference at the end. 🖖🏾
F series CPUs bigger bar better for db workloads
1:41 I had to do a double take and you confirmed it again at 10:16. 🤯
Turin's new IO die is the bomb here.
I love you Wendell *provides unsolicited hug*
Sapphire Rapids's bloodbaths by Wendell and Patrick made my day. Looking for the Intel 6900's vengence later.
Man I can't wait till these get 'cheap' enough secondhand for me to throw one of these in the homelab. It will probably be a while before I get to play with these CPU's but I'm so looking forward to it.
9175F with all the cache?
Music to my ears!! LOL Nice! Great throwback to Airplane, :Surely!"
instant like for the ''Don't call me Shirley'' reference
Since most of us use OLTP, I keep wondering what a PostgreSQL benchmark might look like (I miss AnandTech and their database benchmarking). As for hosted LLMs, I don't see the reason to run GPUs on server for inference when there's services like Groq that are super cheap. Would love to hear other use cases.
😂 I was waiting for the don't call me surely.
That's the default background to the Essential Phone!!! I had one of those, very sad that they cancelled them. I still use that background on my phones to this day, it's really nice
This Byrdbrain loved those modified lyrics.
Currently, vendors known for ISV-certified workstations such as Dell Precision, Z by HP, and Lenovo Thinkstation do not offer models powered by Epyc, only Threadripper Pro.
Do you know where could I find / test some comparison for Comsol?
I'm mainly interested in electromagnetic models
Mary Hopkins thanks you !
I wish the low core count cpus were much cheaper. Perfect for a workstation .
A hundred million PCIe Gen 5 lanes don't come cheap :P
@@andersjjensen Very true . All i really need is memory bandwidth and some 30 pcie lanes
One thing missing and why I generally distrust Phoronix benchmarks for HPC results is that compiler optimizations are rarely or improperly used.
Can you maybe try comparing the Granite Rapids using the intel oneAPI compilers against the AMD Turins using the GNU compilers? In particular using -xhost and -mprefer-vector-width=512 or alternatively -qopt-zmm-usage=high for an older "classic" compiler.
Most of the customers I support, don't need 128 cores. It's more like 8 - 16, but A LOT of Ghz and IPC, since some of their applications are ancient and run on like 1 or 2 cores.
I mean 9175F at 4.2 / 5.0 is impressive, but even in a VM Environment your typical "small to medium business server" - 12 out of 16 cores are are going to sleep most of the time no matter the OS as Host, because the base of Software like Microsoft Office and "Your Average Business Solution" was coded in like 1999 and just got more features put on top of existing "legacy stuff that is still in there because of compatibility" and I bet "small to medium business servers" are probably 75% of the market at least in "numbers sold".
Imagine new Threadripper generation of these chips. My wallet is ready
Your wallet lied to you
@@henson2k he doesnt even have a wallet
10 years from now, I can't wait.
I went to Turin on holiday just a week ago.
I liked because of that ending.
I liked the poem. Comic shot with star wars score in the end.
That heatsink duo is cool. And I don't mean just its ability to cool how much wattage?
500W Cpu that 's crazy... 3x more than my PC while gaming. But that many cores doing alot of usefull job.
I would like to see that powerdraw per core on desktop cpus.
I can see the excitement in Wendell's face, and in my mirror too.
Would love to throw those 9575F at some ANSYS apps like CFX and Fluent with 100+ mil cells.
Leaving room for generational improvements.
Would love to go the RECC on everything, because why not
Love opening youtube to a new video lol
About the Desktop vs. Server memory:
At least in HPC, this feels like it's been a long time coming: Check out Fugaku's setup from 2020, with HBM2 as main memory!
He`s gone to the Byrds!
He was so pleased with himself 😂
@@steverussell7005 Spot the Bard! :D
You are EXACTLY my kinda nuts! Brilliant video. C'mon algorythm, tickle tickle.
This is exactly what we needed
Discord spinning 'something went wrong' in the background, we have all been there.
One loasd I'd love to see is some hardware simulation stuff with something like vcs, but afaik they don't really do benchmarks and while you could totally do it on real designs, the licensing for this get pretty crazy.
When we run regressions it typically takes multiple days for some projects, running just a few ms of the big SoC
Those prices sure are Epyc! $14k?!? Wowza. I'll take 2!
I don't watch Level1Tech all that often compared to the Steves and opening the video almost gave me a heart attack since I'm used to an intro card. We call that "mit der Tür ins Haus fallen" in German (to fall with the door into the house). How about a greeting first?
"Turin, Turin, Turin " LOL I see what you did there, the classic song The Byrds "Turn, Turn, Turn"
no windows task manager screenshot??
That 9175F with the 16cores and 512MB of L3... really stands out from the lineup. I wonder what industry wanted such a massive L3 for so few cores. High Frequency Traders maybe?
Wait how many chiplets is that?
@@kevinerbs27782
Probably more to do with licensing :/
Long time watcher but, I don't ask for much. Any chance your escapades cooinside with Revvit? My friend is an architect working on large scale projects but, of course he (they) are warry of going deep on tech.
Hey new subscriber here! What’s the standard “unit compute”? If there isn’t a metric, shouldn’t there be?
I have exactly zero use-cases for a server like that... but I want one... badly!
Hi Wendell, I am building a 9950x precisely to use simulation though I can't afford what you got there. I am a user of IRONCAD with AMPS Technologies and venturing to Flowsquare but trying to make a work around on meshing at it only supports Inventor. Almost all hardware is in, just waiting for the processor itself due to stock delays.
thanks for keeping ida of forums alive. Crucial part of internet tbh
Damm, they didn't even try with the desktop chips did they?
Any info regarding Zen5 based threadripper for workstations? I was thinking on buying a Zen4 7000 series TR but with Zen5 release I opted for a 9950x now and wait for whatever AMD will release on Threadripper or 3DCache versions. Maybe the 3D cache can accelerate my mostly Java based workloads as the JVM is very cache size bound i believe
We need galactic civilisations benchmarks. A truely multithreaded game not sure it can use that many cores but we can try
It would be interesting to see how llamafile runs some of the bigger LLMs on CPU, using a couple of the better chips in this generation.
ok the youtube algo is getting really good... last night it recommended the STTNG video that the end audio is from, so I guess it saw the future?
All it would take is premium SFX motherboards to complete the workstation dream with this generation. Though a WX competitor for the 5090 would also be welcome.
I've love to see a reality capture dataset of a few thousand high res photos on this.
I would love to see the Epyc 9575F paired with a 4090 (or 5090) in action on cyberpunk 2077, final fantasy 16 or black myth wukong. I would ask please to test on cpu-z too if possible.
I Would like to know how long the 192 core needs to Render the old final fantasy movie.