"Turn down (I/O) for what?!? - Lil Jon & TTP remix 2021 (*~*~*~ Axel F vs. A-ha ~*~*~*)
@cristianr9168
3 жыл бұрын
Beautiful
@Peter.H.A.Petersen
3 жыл бұрын
TTP Remix ? - I like the TDP Remix better ...
@KubeSquared
3 жыл бұрын
You really like videos exactly 13:37 in lenght, don't you Ian?
@TechTechPotato
3 жыл бұрын
This one was 13:40 after editing, and there's always room to massage that number :)
@handlevognhandlevognsen3128
3 жыл бұрын
L77T
@nycameleon
3 жыл бұрын
pretty pro if you ask me
@TechTechPotato
3 жыл бұрын
The problem is whether to go for 13:37 on the dot, or 13:36 and 29 frames, which rounds up. Looks like different browsers round it differently - either as floor(), ceil(), or round(). In my browser the thumbnail says 13:38, even though it's a 13:37 exactly.
@mihumono
3 жыл бұрын
@@TechTechPotato even KZitem app says 13:38 in thumbnail and 13:37 when You play the video.
@countach27
3 жыл бұрын
Undervolting the SoC on my 5800X shaved off about 7W or nearly 50% of its power consumption, according to HWinfo64, in a power limited scenario that’s a tangible difference in power budget
@Reinhardovich
3 жыл бұрын
Is it fully stable though?
@countach27
3 жыл бұрын
@@Reinhardovich Running 1900MHz IF at 975mV, have had no BSODs, no WHEAs or memory errors.
@tomlemanh5506
3 жыл бұрын
@@countach27 VDDP, VDDG @ 0.975?
@countach27
3 жыл бұрын
@@tomlemanh5506 I left them on Auto, so I suppose they should be at 975mV as well
@PanduPoluan
3 жыл бұрын
The fact that _even_ with power throttling due to I/O, EPYC is still kicking Xeon's arse front & back, and inside-out front & back again (BH6 reference). And that fact makes me even more in awe at the raw brute performance of the EPYC Zen 3 cores.
@Morkvonork
3 жыл бұрын
Is the power consumption of the I/O-die always the same regardless of load?
@TechTechPotato
3 жыл бұрын
It's within a small margin based on testing. See the 5950X numbers that pop up for difference between idle and load.
@Cooe.
2 жыл бұрын
11:20 With the GlobalFoundries wafer agreement AMD has, I actually don't see this happening ANYTIME soon. I think the I/O dies on Zen 4 products will use the newer, semi-recent 12nm update GlobalFoundries released (which they claim is almost competitive w/ 7nm solutions).
@theigpugamer
3 жыл бұрын
Amd releases milan and crushes intel Intel : the torture will be over when we launch sapphire rapids Amd gets 10% more perf, leaks 128 cores Bergamo plus sapphire rapid delays Intel : we are f***ed
@shadow7037932
3 жыл бұрын
And the most idotic thing is, Intel doesn't seem to have utilized Jim Keller's time and knowledge while he was there due to organization/management failures.
@BNOVA
3 жыл бұрын
@@shadow7037932 Keller's stuff won't come on board until after alder lake.
@Zadagu
3 жыл бұрын
I realy like the site notes.
@MarkRose1337
3 жыл бұрын
Where can I get a coffee mug with the TTP logo?
@steffeneilers8530
3 жыл бұрын
could this also have something to do with Daytona being 2 DIMM per channel and the Gigabyte one only having 1 DPC?
@VioletPrism
3 жыл бұрын
Great video :)
@ObiWanCannabi
3 жыл бұрын
the thing you dont take into account with chiplets is 4 cores is plenty for many tasks, and these have what 8 cores per chip? in reality it allows chips to focus on different things without being tied down by other operations, 8 cores does ya fine for most things, opening programs loading the data to memory etc is all done pretty fast with the minimum requirements, its only when you load real big files into memory and then have to seek thru that memory, i dont imagine it splitting the process over 4 chiplets when it only needs 4 cores for a process, so its not wasting that much power by having those chiplets, until it really needs to boost, and lets be real how many people using their home desktops for 3d rendering or anything that really requires 32 or 64 cores, this is all shit we wont be using for 25 years.
@pentachronic
3 жыл бұрын
I/O has always been the bottleneck for both power and speed (they go hand in hand). The only solution is integration on die but that affects yield big time. The tradeoff is why you have caches but with coherency on multiple CPU dies that's a big issue too.
@ПётрБ-с2ц
3 жыл бұрын
painful incompetence tbh i also recall how AMD approved laptop with 2x4 chips of DDR4 and without mux chip as review samples for RDNA2 mobile GPUs.
@chriswright8074
3 жыл бұрын
Chip shortage they use what they had that's been stated
@ПётрБ-с2ц
3 жыл бұрын
@@chriswright8074 i bet AMD had a choice there
@rostyloco1
3 жыл бұрын
+
@saultube44
3 жыл бұрын
Blah Blah Blah Blah
@tommihommi1
3 жыл бұрын
Zen 4 will probably put some of the power saving experience AMD has made on the APU side to use, along with going to a 7nm-class node. But at the same time they have to move to DDR5 and PCIe 5.0, so maybe uncore power consumption won't go down significantly after all.
@C.Q.Q
3 жыл бұрын
Isn't DDR5 supposed to lower power consumption?
@tommihommi1
3 жыл бұрын
@@C.Q.Q the DDR5 modules are lower power at the same speeds. The requirements and thus power for the memory controller go up.
@LunarLaker
3 жыл бұрын
@James Smith I'm not as optimistic about a smaller process IOD. The PHYs are something like half the die and mostly analogue. If they used some new interposer tech like LSI or any kind of silicon, that would be pretty rad. Albeit probably difficult on server processors given their size
@LunarLaker
3 жыл бұрын
@James Smith Analog scaling is typically -15% at best, which would give ≈72% area over 2 nodes. Of course a lot of the die is controllers and tag directory, which I understand as having logic and SRAM so that would scale decently. I've read about μbumps being an issue for Zen 2 CCDs that was closely avoided by a shrink from 150μm to 130μm that "only 2 vendors in the world could do," which doesn't lend confidence to an IOD at 6nm. Also, AMD still has $1.6B of wafers to buy from GloFo in the next 3 years, which of course will include embedded and potentially Athlons, but I could see them keeping Ryzen IODs on GloFo since IO needs haven't increased much and thermal budgets are generally looser. Obviously all/most of the 'leaks' point to 6nm so it is probably on that, but there's a case to be made that the benefits might not be so big.
@BNOVA
3 жыл бұрын
@James Smith I think AMD has completed thier deal with Gloflo as they will be making zen3 4 core chips for laptops using Gloflo 12nm and has bought on advance more I/O die chips to Mai tain contracts for existing server chips.
@khealer
3 жыл бұрын
l33t video
@leax_Flame
3 жыл бұрын
Neat. I’m really interested in an AMD I/O node change.
@Psychx_
3 жыл бұрын
Analog circruitry (PHYs) usually doesn't scale well. It gets like 10%-15% smaller when moving from 12nm to 7nm with power savings being in the same region at best.
@predabot__6778
3 жыл бұрын
@@Psychx_ That is what's been said, yes - one has to take into account the move to TSMC itself though -their process-nodes are not just more advanced, but as I understand it, the actual skill and knowledge of the operators and engineers over there, is far greater than at Global Foundries. That, in itself, will make the IO-die's significantly improved, so there's certainly ample room for improvements there. I wouldn't be suprised if we see power-savings on the IO-die, in the region of about 20% or so.
@leax_Flame
3 жыл бұрын
@@predabot__6778 So effectively what I’m hearing is, Great potential for improvement, but temper expectations. Either way 10% or 20% an improvement is still an improvement.
@nathanlowery1141
3 жыл бұрын
@@leax_Flame a quite good improvement... and thats just in the i/o die....
@jaymacpherson8167
3 жыл бұрын
Ian, very nice graphic on the CPU at about 1:40 if you added that to the AMD image. It makes your point very clear.
@tuxjob
3 жыл бұрын
Hi Ian! My daughter (5 years) is asking, why you call your channel 'TechTechPotato'... :-) Looking forward to your answer... :-D
@TechTechPotato
3 жыл бұрын
I have a habit of biting processors, or chips - I have a nickname at Intel as Dr Wafer Eater. Potatos are chips. Also it was a bad auto-subtitle in one of my first videos with Wendell; I posted the screenshot in the community tab if you want to scroll back 😀
@tuxjob
3 жыл бұрын
@@TechTechPotato Wonderful! Thanx a lot!!! x,D
@Irthex
3 жыл бұрын
I imagine that AMD did a minimum effort system with Daytona for Milan because they thought they could spend the engineering time more efficiently elsewhere. They want their partners to be selling cool systems. Milan didn't really need more performance from the reference system. Though I'm certain they're a bit frustrated that they didn't send a partner system to the reviewers instead.
@lifeinthaboot
3 жыл бұрын
Could've sworn I sub'd when "Creator of the Week" thing happened. Sub'd now at least 😘
@Psychx_
3 жыл бұрын
Do you know how AMD managed to reduce memory latency by ~10ns, despite keeping the same IO die when comparing AM4 Ryzen 5000 and Ryzen 3000?
@WereCatStudio
3 жыл бұрын
Ryzen 3000 has each die (CCD) split to 2 parts (CCX). So if the first 4 cores need to share data with the other 4 cores on the same die, then you need to send it back to I/O die and back to the die. On Ryzen 5000 each CCD is also a CCX so you don't have to send the data back and forth. The same applies to the L3 cache.
@GameRep101
3 жыл бұрын
@@WereCatStudio Your not wrong, but that doesnt answer the question. Becasue of what you said, they were able to decrerase the physical distance electrons have to travel which is why it takes 10ns less. They could have made these changes to the CCX complex and still not have reduced the latency. But that was a target for them.
@IM2awsme
3 жыл бұрын
All your videos are great and informative, your sub count dosn't do you justice.
@porina_pew
3 жыл бұрын
I'm really curious how they will handle DDR5. Ideally I'd like to see full bandwidth potential from each CCD to ram which implies a significant upgrade to Infinity Fabric and the IOD. My fear is they wont and instead try to mitigate it by throwing in more L3 cache, or we see a shift in single CCD consumer desktop CPUs moving towards APU-like architecture. Especially when running to fixed power limits (applies to Intel also), it will be interesting to see how the power budget is spread between components of a CPU going forwards.
@dermothoyne2393
3 жыл бұрын
1337... Good timing, for this Milan chip / Gigabyte Mainboard combo x299 chips in wheezing laughter fit, over no longer being Intels nuclear radiation leak
@Mohamed-iw2ow
3 жыл бұрын
Why is there a mia khalifa ad on this video lol
@TechTechPotato
3 жыл бұрын
Google ads work based on your previous personal searches.
@Mohamed-iw2ow
3 жыл бұрын
@@TechTechPotato not at all, i opted out of everything plus im not a coomer.
@Ts6451
3 жыл бұрын
@@Mohamed-iw2ow You got that ad because the Google AIs figure that this specific stimuli given at this time would be the optimal in furthering your neural rewiring in a way favorable for their great plan... Or perhaps it's just that not all advertisements are hyper-specifically targeted, and so might target some general demographic with the hope of enticing new people. I would expect that people that might fall under your idea of a "coomer" probably already have their favorites, and so it might be more effective to target a broader audience.
@Mireaze
3 жыл бұрын
1kw CPUs coming soon? Needs an ln2 pot to run without going thermonuclear
@erlienfrommars
3 жыл бұрын
That Nokia 3210 on your background is
@tomstech4390
3 жыл бұрын
Doubt it, it probably needs a charge :) then it'll be fine. Dropped one off a roof once, saw it fall in slow mo then shatter into its various parts, put it back together and worked fine.
@TheReferrer72
3 жыл бұрын
I thought the power budget was just an arbitrary figure for how much the chip needed cooling.
@TechTechPotato
3 жыл бұрын
AMD Enterprise processors it's the peak. AMD Consumer you want to look at Package Power Tracking.
@GameRep101
3 жыл бұрын
You are right and wrong. It isnt entirely arbitrary, tech tech potato doesnt know what he is talking about. The silicon can take way more or less power than TDP, becasue yes, TDP is for thermals, literally stands for thermal design power. Intel and AMD use different formulas, the actual power the CPU uses depends on the voltage response curve of the silicon, which changes based on the manufacturer and the surface area of the silicon(Super simplification). For Ryzen 5000 there is on avg a 12% delta above TDP for actual power. Poor intel right now is at 40% avg for their 11series. Idle power wont have an effect on power allocation to the cores, it will only effect performance in that if the software has the silicon pulling excess current while Idle you will thermally saturate your cooling solution sooner causing thermal throttling sooner. Imagine you have your foot on the break and on the gas pedal of a car. You wont move but will use alot of gas. Your engine is still doing the same thing than if you hadn't put your foot on the break. Once you let your foot off the break its going to allocate all of that power to movement, where else is it going to allocate that power? Same thing in a cpu, theres no idle Piece of the silicon saying Im using this power. Tech Tech potatos lack of understanding is hilarious, he has all these theories based off high level overviewes that are so far from reality from and engineering perpective. It's like watching a kid theorize about how santa can travel to everyones house in one night, not knowing that he cant't figure it out becasue he doesnt have a real understanding of what santa actually is.
@lordofthecats6397
3 жыл бұрын
@@GameRep101 Yeah, TTP should totally read what an actual professional wrote. The two articles below explain it pretty well ;) www.anandtech.com/show/13544/why-intel-processors-draw-more-power-than-expected-tdp-turbo www.anandtech.com/show/14873/reaching-for-turbo-aligning-perception-with-amds-frequency-metrics-
@djnorth2020
3 жыл бұрын
I'm always interested in lower power usage first and then the performance increase.
@TheKb117
3 жыл бұрын
yup, lower idle power and lower power usage overall, then performance next... as a normal consumer like me.
@c.m.7692
3 жыл бұрын
Hey Dr.Cutress! Thanks for the awesome informative content that helps me relax after stressful days... Can I very discreetly ask you a little question under this pile of comments?
@c.m.7692
3 жыл бұрын
Alright, here it is: if reducing power on IO allows such a massive gain in perfs, what would be the impact of enhancing the power consumed on the infinity fabric? Would reducing the temperature of the whole package or the bottom of the package (through the socket for example) have a meaningful impact on performance? I really want to stick that TEC cooler on some socket and I am starting to think this might be it....
@TechTechPotato
3 жыл бұрын
@@c.m.7692 in general, I'd avoid using a TEC. They effectively double power consumption, have hard limits, and require lots of management
@c.m.7692
3 жыл бұрын
@@TechTechPotato Thanks for taking the time! More generally would you see backplate/socket cooling being a thing at some point?
@mark11145
3 жыл бұрын
Why? Look at it from a business decision. Perhaps it was because AMD won the performance crown even without it so…. 1) Why disappoint those wanting to upgrade at launch. Instead you have a clean launch and upgrade Story rather than; now if you want an extra 10% upgrade the motherboard as well. Easier to swallow than the old motherboard doesn’t fully support Milan at launch. 2) Two gut punches to Intel rather than one. Two waves of promotion rather than one. 3) Jebaiting Intel. Intel thought they knew what Milan was. Perhaps even locked down the specifications in their product road maps and moved forward. Surprise 😈 Or maybe AMD was still working it out with Motherboard vendors and at launch wanted to wait and see if they could pull it off before things started shipping. I can’t wait to read the book about AMDs comeback 10 years from now and get the real story behind this.
@johnscaramis2515
3 жыл бұрын
I'm not sure how often only the CPUs in servers are exchanged and not the whole server including board. And if you put a new CPU in an old board, usually you can assume one or the other drawback.
@TheSilviu8x
3 жыл бұрын
If anything, I'll probably go with Intel, for ddr5 platform, Amd is slowly creaping up the prices a bit too much.
@miyagiryota9238
3 жыл бұрын
Not at all compared to $20000 xeons! ROFL!
@earthtaurus5515
3 жыл бұрын
@@miyagiryota9238 Precisely, it seems folks have forgotten how much Intel was charging for even less performance. A few years ago the amount of processing power and capabilities offered by a 2700X for example would have set you back thousands with Intels HEDT platform as opposed to a mere £200. I paid £193.97 for mine (the R5 3600 was £192.68 at time of purchase - Sept 2019). It's aggressively undervolted, All core O/C to 4075 with a Arctic Esports 34 coupled with 32 gigs of ram at 2800Mhz (4 x 8 - two exact same kits of 16 gb bought at different times = £142) at 1.236 volts 16-15-15-15-36.I can run VM with 12 threads, 7zip ultra compression with 16 threads, still run a game like Tekken 7 at 1080p, firefox with hellova lot of tabs, steam and all the usual cumulated background tasks that a person acrues over time - _at the same time_ - The PC is completely responsive.
@truevy
3 жыл бұрын
I’m liking this video because of the t-shirt :) and oh yeah the AMD stuff.
@Speak_Out_and_Remove_All_Doubt
3 жыл бұрын
A little bit of semantics but I would say we have readed the limit of 'energy density' rather than the power limit of CPUs. So if they made the surface area twice as big they could use twice as much power as long as they design the CPU to spread out the heat spots. This would naturally bring other compromises but might be worth it at some point. There are also a lot of other technologies out there that we haven't tapped into yet, micro-fins within the CPU to channel cooling around the chip or carbon nanotubes to act as insanely efficient heat pipes. So we could still see massive power consumption increases, we just need to be more creative with our heat management.
@tripzero0
3 жыл бұрын
How much do customers of these platforms care about idle power? The fact that Intel uses half the idle power seems like a big deal.
@alexmills1329
2 жыл бұрын
If you have a lot of servers sitting idle often why do you have those cpus?
@tripzero0
2 жыл бұрын
@@alexmills1329 because you never need CPUs... until you do. 😂
@nycameleon
3 жыл бұрын
12:48 What's your minimum specification!!!
@TheKb117
3 жыл бұрын
sooo, Daytona is like training weights that limit the potential of Milan... Seems like Piccolo with his heavy turban and cape... Miian be like: This is not yet my final form... too much DBZ reference, hahaha
@ancient1der
3 жыл бұрын
With ddr5 modules being able to be 2 channels on one stick, I was wondering why they weren't going to 4 channel memory on any of the leaked spec for Zen 4. Maybe this answers that question. My brain thinking since 2 channels increase performance wouldn't 4?
@paulmaydaynight9925
3 жыл бұрын
^_~ so... min spec =10x lightpeak @ 70% less power interconnects + external IO lightpeak @ $1 a port x16
@treyquattro
3 жыл бұрын
this is really interesting, and pretty unexpected. My expectation is that performance and efficiency improvements are transmitted more or less equally throughout the processor, but apparently this is a wrong assumption (as nearly all assumptions turn out to be!) I want improved IO throughput, but not at the expense of thread utilization. That said, maybe you have to more closely examine your load mix to determine what's the best direction to go in. I wouldn't have expected to have to do those kinds of tests, instead treating the processor as a black box, and the memory subsystem pretty much as another black box.
@ole7736
3 жыл бұрын
So was it voltage? Power saving options such as turning off PCIe lanes? If the latter, the power consumption might come back up in certain scenarios.
@SuperCape
3 жыл бұрын
A 350 watts space heater will heat a small sized room in minutes. How TF can you pump all that energy through a piece of silicon the size of a toe nail and not bust it?
@ropersonline
3 жыл бұрын
I'm watching this video, and I keep mishearing "for the cores" as "for the cause".
@mariodrv
3 жыл бұрын
Where leves that Global foundries if they don't use their 14nm for IO anymore?
@trjozsef
3 жыл бұрын
Dual side cooling with side pins, extreme 3D chiplet placement for a larger area, we can push the envelope. At the end of the day if you can cool Cerebras, you can cool these puny CPUs.
@zzco
3 жыл бұрын
lol nice cut-out of Dr. Su :p
@cj09beira
3 жыл бұрын
they really need to move away from using the substrate for inter chiplet io if they want better efficiency, packaging tech is where innovation needs to be right now
@shadow7037932
3 жыл бұрын
Seems like AMD is doing exactly that with their/TSMCs TSV tech.
@kazedcat
3 жыл бұрын
Substrate traces are cheap they will go interposer when there is no other way. Moving to 7nm IO die saves a lot of power so that is what they are going to do first. The chiplet GPU is going interposer to connect them because it is necessary to get the required bandwidth between chiplet.
@cj09beira
3 жыл бұрын
@@kazedcat problem with just moving to 7nm io die is that they can't actually shrink it, amd is already at the top end of organic interposer pin densities, unless they reduce the number of pins (=lower performance), the choice isn't a easy one that's for sure, maybe they can use some emib like embedded interposer, we will see
@kazedcat
3 жыл бұрын
Bruno Cardoso AMD are quite happy to add dark silicon to the design to keep the pin out they need. They have done this when they switch from 14nm Ryzen into 12nm Ryzen. Dark silicon have minimal impact to cost. A giant interposer on the other hand is a lot more expensive even if you use older nodes you need to pay for the development of mask. Then there is reticle limit problem. The reason AMD give as to why they did not use interposer for Rome is that it exceed the size limit of the interposer. There are ways to solve it like having multiple interposer but that adds more cost.
@Kilohercas
3 жыл бұрын
you need more sleep :(
@animeintensifies
3 жыл бұрын
Heyy first time seeing a sponsor! Congrats man you’re growing fast :D
@TechTechPotato
3 жыл бұрын
This is actually number four :) I have some exciting sponsored stuff in the works, stay tuned
@IM2awsme
3 жыл бұрын
@@TechTechPotato how long until wafer scale chips are used for MMO servers? I need my full no compromise VR MMOs.
@awesommee333
3 жыл бұрын
Hm. I wonder if we’ll start like seeing tiered io dies or something with higher core counts. Cause with a lot of cross talk and like 2nm that feels like that’s going to be a LOT of power, so lol maybe like have several io dies for different clusters of cores. No idea but I really wanna see how this chipset scaling could be taken further, as epyc seems to show that they might be close to reaching the limit?
@Dysphoricsmile
3 жыл бұрын
*I can't stop staring at what looks like an RTX 3080/90 series card with a REALLY weird fan in the background.* Every damn video now. Dr. Ian Cutress, what is that thing's minimum specification?
@TechTechPotato
3 жыл бұрын
I made a video about it... :) kzitem.info/news/bejne/rm2hqGScgKGeZJw
@davidgunther8428
3 жыл бұрын
For both CPU and GPU 300W for package power is starting to seem like the point where cooling and electrical power delivery stop being simple. Might be a W/mm² limitation with heatpipes, at least with the standard 6mm size? Supplying large amounts of power at 1V and then extracting it at low temperatures, all while relying on copper wiring and copper plate heat spreaders. We need to invent a new metal! 😆
@shadow7037932
3 жыл бұрын
Watercooling exists ;)
@davidgunther8428
3 жыл бұрын
@@shadow7037932 still has to go through a copper base plate
@namibjDerEchte
3 жыл бұрын
@@davidgunther8428 yes, it does, but that can handle about 100W per Zen3 CCD, if evenly loaded. SP3 has the heat sufficiently spread out to make it scale basically just as good. The cooling looking for SP3 is more like 900W, including the 100W IO die. I wouldn't be surprised about LGA pin melting, though, at these currents. Power delivery may need to move to the back of the board to fit closer to the pins without being in the way of cooling.
@calaphos
3 жыл бұрын
Nvidia is selling 400W A100 GPU modules and pack anywhere from 4 to 16 of them into a single system. I don't see why there isnt room for 350W CPUs in a 2 socket server.
@Winnetou17
3 жыл бұрын
It's about the die size.
@davidhalliday7776
3 жыл бұрын
The Anand graph was comparing EPUC 7763 vs EPYC 75F3 according to the graph legend (not the title above the graph) these are different CPU!
@TechTechPotato
3 жыл бұрын
Check again. It had both CPUs, comparing each in Daytona vs each in GBT. Source: I'm Anandtech's senior editor 😉
@marlls1989
3 жыл бұрын
One huge power budget savings would be eliminating the x86 frontend. Decoding and executing the virtual x86 instructions has been a waste of power for a long time! Now that we are hitting the ceiling of things we can keep powered on at a time in a chip, it might be better to rethink using this horrendous ISA that was kept for business reasons.
@marlls1989
3 жыл бұрын
Also, the interconnection is probably been implemented using some form of asynchronous logic, QDI templates excel as interconnects due their robustness against delay variability and voltage scaling.
@bp56789
3 жыл бұрын
We're nearly at 50k subs
@gfeie2
3 жыл бұрын
So what about those TechTechPotato plushies?
@ciaduck
3 жыл бұрын
I thought TDP was NOT actually a measure of CPU Wattage. Manufacturer TDP is just a made up number. What is the actual power draw? Can we trust that TDP number?
@ciaduck
3 жыл бұрын
@@-szega Thanks. I also saw Ian reply to another comment with the same question. I found it after posting. But this makes more sense now. Thanks for taking the time to clear it up for me.
@robinvonnegut7987
3 жыл бұрын
Have we really reached maximum power usage for single CPUs? I remember there was a Xeon W-3175X running at 1kW with a compressor as its cooling or something🤣
@chesshooligan1282
3 жыл бұрын
Put a couple of wheels on that Xeon and you have a scooter.
@robinvonnegut7987
3 жыл бұрын
@@chesshooligan1282 LOL
@asitisrequiredasitisrequir3411
3 жыл бұрын
lol video length
@kusumayogi7956
3 жыл бұрын
So how can exactly you get extra 10% performance?
@johnscaramis2515
3 жыл бұрын
That's quite simple: in power limited scenarios, every Watt that goes into another part of the processor (be it Ram controller, PCI PHYs, USB PHYs etc.) ist not available to the cores. With less power necessary for other components in the CPU, more power is available for the cores, hence higher turbo. Let's say Power limit is 200W and we only count IOD and 16 cores. If the IOD consumes 100W, 100/16= 6.25W are available per core. If the IOD consumes 50W, 150/16= 9.38W are available per core. And turbo frequence relies on available power...
@velo1337
3 жыл бұрын
is there any chance of doing an podcast with an actual amd or intel engineer? or are they all under heavy nda?
@TechTechPotato
3 жыл бұрын
Senior Engineer or VP maybe, anyone lower than that unlikely. PR won't allow it, legal probably wouldn't. They're all working on stuff two generations out anyway
@PhoeniXfromNL
3 жыл бұрын
nice, besides a very interesting video, that length is just humor lol
@bfish9700
3 жыл бұрын
Dr ian cuttress bringing it.
@Alan_Skywalker
3 жыл бұрын
Perhaps Gigabyte motherboards' power reading isn't accurate. As they aren't in their LGA3647 motherboards. There are little things motherboards can do as AMD CPUs are just huge SOCs and motherboard is nothing but a huge wireboard with slots. Also Agesa FW is universal.
@johnscaramis2515
3 жыл бұрын
A wrong reporting does not increase the benchmark results.
@Alan_Skywalker
3 жыл бұрын
@@johnscaramis2515 It does. If the reporting is lower, the CPU will not think it reaches the power wall, thus will maintain at higher frequency. Actually WC621D8A-2T has a setting in its bios to reduce the power reporting by half, so the CPU could boost higher.
@Alan_Skywalker
3 жыл бұрын
@@johnscaramis2515 BTW power hacking is not a new thing. It happened so much that HWinfo had an update about "power reporting deviation" back then. www.hwinfo.com/forum/threads/explaining-the-amd-ryzen-power-reporting-deviation-metric-in-hwinfo.6456/
@G-u-z-i-o
3 жыл бұрын
Cat Tax was the best part!
@wile123456
3 жыл бұрын
We are also dealing with the problem of lack of sleep in tech journalism, hope youre alreight and your dark rings under your eyes are just an anomoly
@TechTechPotato
3 жыл бұрын
That's just me normally; the light and what I wear makes it look worse than it is
@wile123456
3 жыл бұрын
@@TechTechPotato ah okay, no worries then. But I could imagine you're the type of guy staying up all night to do a marathon interview haha
@marcin_karwinski
3 жыл бұрын
They'll decrease the I/O die manufacturing node to 7nm or less but they'll "compensate" with the L4/L3 cache as shown by Lisa Su. More on-package cache = more power needed, but reduced power and latencies for the smaller number of roundtrips to RAM for data/instructions may result in roughly similar overall package power...
@GameRep101
3 жыл бұрын
Any Electrical Engineer will tell you this guy has no idea what he is talking about. He literally has no idea how impedance works and generates heat in silicon. The Idle power is almost completely software-driven. The cores are getting the same power on both platforms. The infinity fabric won't generate any practical difference in thermal dissipation and power usage while Idle. Even at load, it will be negligible in comparison to active cores by about one order of magnitude. I could go on and on, but it is clear this guy knows only high-level jargon about this field but has no education on the physics, math, and engineering nuance required to understand why these products behave the way they do.
@GameRep101
3 жыл бұрын
* Resistance, not impedence, my bad
@TechTechPotato
3 жыл бұрын
Nah mate, only spent 10 years explaining the stuff on AnandTech, have a PhD in a material science, lead media partner for all companies concerned. But, you know 🤷♂️
@GameRep101
3 жыл бұрын
@@TechTechPotato Nice power move. Material science will no doubt be the most important field in the future for allowing and realizing new technology. I was too negative in my message, I apoligize. This video has to relate to lay audiences, so of course it's not going to satisfy the people who know the chemistry, physics, circuit logic, and math principles already.
@allenshepard7992
3 жыл бұрын
If not 300 watts, step back to six cores and four sockets ? The overhead of 1,200 watts seems excessive. Water and Freon cooling are making inroads from the mainframe to servers to enthusiasts. For years people have put PC in mineral oil for better cooling. Is there a need for 7Ghz cpu with DDR6 memory ? Yes. Much like 288 Mhz was not enough.
@glasser2819
3 жыл бұрын
brilliant topic sir 👍 Its interesting to trace the overall CPU efficiency down to functional bottlenecks. Software has a role to play by matching the hardware strength area. I spent 20 year studying that very topic 🙂
@10lauset
3 жыл бұрын
Cheers
@rarrawer
3 жыл бұрын
Could the fact that they provided reviewers with a system capable of using all three generations be down to them wanting to be sure that reasonably fair comparisons could be made between the different generations? Only having to swap one component in the whole computer does look attractive for testing, benchmarking, and analysis.
@w0nd3rlu573r
3 жыл бұрын
E-ffi-cien-cy!
@TechLevelUpOfficial
3 жыл бұрын
will the next gen EPYC CPUs have even more power dedicated for cores when they have an IO die that's built on 5nm for example? since the I/O dies used in today's AMD CPUs are mostly built on GF 12nm i guess
@johnscaramis2515
3 жыл бұрын
There's no such thing as power "dedicated" to single parts of the CPU. The CPU remains within its dedicated power limit, if one component (in a power limited scenario) draws less power, there's more left for other components.
@Havox7
3 жыл бұрын
Very good stuff. It will be interesting to see how intel and amd go about balancing these systems. I do wonder if they could develop some software that would monitor usage and dynamically shift around power.
@pauljones9150
3 жыл бұрын
I have to ask. Is the intro a herd of avacadoes eating their chip bellies or a herd of potatoes eating chips or a herd of chips eating chips
@MihailSurkov
3 жыл бұрын
Why cant they use hybrid silicon photonic/electronic io die interconnectors for chiplets? Its seems to me like ideal solution.
@theigpugamer
3 жыл бұрын
Just put a potato in the server
@NaokiWatanabe
3 жыл бұрын
Very interesting.
@rafnavi4500
3 жыл бұрын
Aye
@johnathanwilko2922
3 жыл бұрын
Can it handle 2160p? My processor can't play 2160p properly so should I upgrade to this processor?
@Supcharged
3 жыл бұрын
You need dual socket 7763 to play 2160p videos.
@johnathanwilko2922
3 жыл бұрын
@@Supcharged Thanks. I don't know what that means but I will look for it. Can I add a socket to my motherboard? I know it has got expansion slots. So can I add a small motherboard with a socket for a 2160p CPU?
@carbonium1264
3 жыл бұрын
modern GPUs have decode accelerators that can handle 3840:2160@60fps just fine
@johnathanwilko2922
3 жыл бұрын
@@carbonium1264 OK. Thanks. Maybe I buy a modern GPU and leave my CPU alone. I'm worried. Can moderns GPUs handle 4K, 8K and 16k? Because everyone will be making videos in 16K soon. And I don't think my CPU can do AI. I was reading AI is going to become very popular soon.
@Supcharged
3 жыл бұрын
@@johnathanwilko2922 yes, AI is the current big thing. You should purchase nvidia’s DGX A100 system.
@prashanthb6521
3 жыл бұрын
AMD is becoming excellent in building IC engines when the industry is moving towards electric. The future belongs to chips like APPLE M1.
@UniqueBreakfastTaco
3 жыл бұрын
that future has a whopping 17% marketshare. ecosystem will always hold that back.
Пікірлер: 184