[RUMOR] AMD EPYC 7004 Genoa Zen 4 CPU Allegedly has 96 cores and 192 threads with 12-channel DDR5-5200 memory support

[leadeater]

10 minutes ago, CarlBar said:

Memory channel count. Nothing says that every board has to support more than 8 channels. People have been saying current epic can hit memory bottlenecks with just 64 cores in some scenarios. Building extra memory channels into the IO die from the start may be an attempt along with the DDR5 to give those with such workloads an option.

Even standard 2U servers don't really have the area for 12 per CPU, single socket systems yes there is enough space but dual socket then it gets hard. Would have to put CPUs length ways rather than side by side and the consequence of that would be wasted area each side of the CPU and an increase in server depth.

 

It's not that it isn't possible it's just not all that necessary with the move to DDR5. A lot of HPC deployments also use lower core count CPUs on purpose due to the memory bandwidth. Super high core counts isn't always that useful for everything, great for VM hosting though or other similar shared resource deployments like Containers or a crap ton of websites.

[GDRRiley]

4 hours ago, leadeater said:

Just remember, if something sounds too good to be true... it's probably fake or inaccurate.

 

• 12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.
• 12 DDR5 Memory Channels: Possible, unnecessary and also physically and electrically problematic
• 128/160 PCIe 5.0: Likely correct
• 350W-400W TDP: Unlikely, lower TDP parts are already proprietary water cooling only and this is 100W+ more than those

if its 50% more pins that will give them the space to do it but I agree 12 core CCD would be more likely

not really, you've got 50% more cores to feed in some workload its already bandwidth limited

 

idk 280w parts are common to air cool with a big die and IHS 350 wouldn't be that much harder.

 

[GDRRiley]

13 minutes ago, leadeater said:

Yes, the HPC only 7H12 is 280W and deployed only with water cooling. Intel's Xeon 9200 series are also BGA only and come in water cooled systems. Once you go above 250W on a socket air cooling in servers become problematic. You can but it rules out a lot of configurations so no many want CPU with that much power usage. Then you also start encountering power density problems in racks, not ever hosting provider and business has 20kw+ per rack capability

I know places where its air cooled

HPC places do. we have racks at 60KW-70KW if my math for 192 nodes at 300W given 215W chip TDP+ ram or 2 135W chips

the racks are water cooled between them not right on CPU

[leadeater]

1 minute ago, GDRRiley said:

not really, you've got 50% more cores to feed in some workload its already bandwidth limited

Isn't DDR5 double the bandwdith? With the first generation DDR5 coming out at around 50% more.

 

3 minutes ago, GDRRiley said:

idk 280w parts are common to air cool with a big die and IHS 350 wouldn't be that much harder.

Not in servers, neither Intel nor AMD have generally available server CPUs above 250W. The reason for that is 1U servers and 1U nodes for 2U4N systems. Dell lets you configure a standard 2U server with the 7H12 but HPE does not.

[leadeater]

2 minutes ago, GDRRiley said:

We have racks at 60KW-70KW

Yea I know you do but mere mortals don't  I don't think there are any hosting providers in my country that offer that much per rack, not that we need that here. Small country problems, or lack of problem.

[GDRRiley]

19 minutes ago, leadeater said:

It's not that it isn't possible it's just not all that necessary when the move to DDR5. A lot of HPC deployments also use lower core count CPUs on purpose due to the memory bandwidth. Super high core counts isn't always that useful for everything, great for VM hosting though or other similar shared resource deployments like Containers or a crap ton of websites.

we run xeon phi and I think our Milan are top core count (not sure we don't have them in yet)

8 minutes ago, leadeater said:

Isn't DDR5 double the bandwdith? With the first generation DDR5 coming out at around 50% more.

 

Not in servers, neither Intel nor AMD have generally available server CPUs above 250W. The reason for that is 1U servers and 1U nodes for 2U4N systems. Dell lets you configure a standard 2U server with the 7H12 but HPE does not.

I see roughly 40% more at DDR5 4800 vs DDR4 32000 but with 50% more cores your going to starve them more

We do in 2U 4N

6 minutes ago, leadeater said:

Yea I know you do but mere mortals don't  I don't think there are any hosting providers in my country that offer that much per rack, not that we need that here. Small country problems, or lack of problem.

you don't need it because my lab allows your researchers to use our system.

 

I'm not sure what our A100+Milan nodes will run but they need 3 power wips. 6000 A100+1500 Milain in 12 racks
HPE/ Cray Shasta is insane

[leadeater]

1 minute ago, GDRRiley said:

you don't need it because my lab allows your researchers to use our system.

 

2 minutes ago, GDRRiley said:

We do in 2U 4N

You have 7H12/250W+ CPUs in 2U4N systems using the 1U height nodes? How fast do the fans "idle" at?

[GDRRiley]

1 minute ago, leadeater said:

You have 7H12/250W+ CPUs in 2U4N systems using the 1U height nodes? How fast do the fans "idle" at?

 3 chassis; each chassis has 16 compute blades with 4 nodes per blade.

ether 1 xeon phi 7250 or 2 E5-2698 v3 per node

not as loud as are old system but still ear protection needed

[leadeater]

20 minutes ago, GDRRiley said:

 3 chassis; each chassis has 16 compute blades with 4 nodes per blade.

ether 1 xeon phi 7250 or 2 E5-2698 v3 per node

not as loud as are old system but still ear protection needed

Neither of those are that high TDP though, 215W and 135W respectively.

 

Our 1U nodes with 2x 6242R (205W) have idle CPU temps of 54C with current active load of 8% avg with fan RPM at 9k RPM. New cluster with basically nothing on it. Hot aisle containment, water cooled with cold aisle temperature target of 27C.

 

Incidentally those E5-2698v3's are one of the SKUs I mentioned with less cores than the maximum for that architecture, 16 instead of 18. It was one of the ones recommended to us by HPE. We don't have any though, not that I know of anyway. Could be a department out there with them as there is some stuff not run by ITS, including a super old SGI cluster.

[GDRRiley]

Just now, leadeater said:

Neither of those are that high TDP though, 215W and 135W respectively.

 

Our 1U nodes with 2x 6242R (205W) have idle CPU temps of 54C with current active load of 8% avg with fan RPM at 9k RPM. new cluster with basically nothing on it.

 

Incidentally those E5-2698v3's are one of the SKUs I mentioned with less cores than the maximum for that architecture, 16 instead of 18. It was one of the ones recommended to us by HPE. We don't have any though, not that I know of anyway. Could be a department out there with them as there is some stuff not run by ITS, including a super old SGI clusters.

right but thats our old system its still over 1kw per RU of CPU alone.

we haven't released our new systems specs publicly yet so I can't talk about it

 

our last system was E5-2695 v2, the phis are likely the only ones they could get in 10k quality its minor cut down from the 72 core and 74 die

 

[leadeater]

1 hour ago, GDRRiley said:

right but thats our old system its still over 1kw per RU of CPU alone.

True, though the reason more has to do with the CPU cooling itself rather than power density per U. Getting 250W+ from a tiny 1U height cooler ain't fun, I've seen it on 2U, that Dell R7525 for example.

 

I guess I also should have been more clear too, 250W+ dual socket configurations. There are single socket 1U systems that support air cooling of CPUs up to 280W but not many dual. Supermicro for example only supports 225W CPUs in their 1U nodes (dual socket) for 2U4N systems, 280W single socket blade, 280W standard 1U dual socket Ultra series but only if you meet specific conditions (dunno what they are but it'll be high airflow upgrade kit and ambient inlet temperature range, not much else it could be).

 

Generally vendor and OEM server systems for the standard market top out at 225W or there about per socket, higher than that is the exception and comes with requirements like the fan upgrades etc. This is why the mass market Xeons and EPYCs don't go above those TDPs, ones that do are marketed as specialty cases. That's why the TDP claim for this rumor to me is the most sus, it's far too high.

 

Cray for example, who are part of the more specialist hardware group, have two variants of their current XC50, the XC50 and the XC50-AC with the later being Air Cooled. NIWA/NeSI here has both variants in use, they also use Xeon 6148 CPUs.

[HenrySalayne]

8 hours ago, leadeater said:

Just remember, if something sounds too good to be true... it's probably fake or inaccurate.

 

• 12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.

Maybe we'll see something like Foveros for this new generation. Stacking the CCDs on top of the IO die would give some more room to work with.

[porina]

8 hours ago, leadeater said:

12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.

Is this going 5nm? If they use that primarily to make smaller chiplets, as opposed to doing much more in the same area, that could help out in the total area concerns.

 

6 hours ago, leadeater said:

I'm not even sure it's possible to connect them to the IOD, there's actual traces in the sub-straight and suffice to say shit gets complicated as you increase the number of die you need to interconnect.

Time for Infinity Fabric 2: double the bandwidth per wire, make it half as wide, or something like that.

 

4 hours ago, leadeater said:

Isn't DDR5 double the bandwdith? With the first generation DDR5 coming out at around 50% more.

One module is basically double the bandwidth, and you can (eventually) buy speed grades according to your needs. Do note that with DDR5, one module has two channels at half the width each relative to DDR4. If we take the 12 channels at face value, that's only 6 modules.

 

 

 

And the reason I found this thread, seems too close for me to start a new thread: claimed additional instruction/data format support:

https://wccftech.com/amd-epyc-genoa-zen-4-cpus-rumored-to-feature-avx3-512-bfloat16-instructions/

Quote

The Zen 4 core architecture will also feature two new ISA's & those include AVX3-512, BFLOAT16, and a few other extensions that have not been stated.

This is very interesting to me. Per core, AMD has generally lagged Intel in FP64 compute and remains so today. AMD's best is slightly better than Intel consumer, but falls far behind 2 unit AVX-512 enabled Intel CPUs. Are AMD finally going to catch up and/or pass there? There is a big question in their implementation though. Even though 1st 2 generations of Ryzen supported FMA3, it was without sufficient hardware backing so its performance was more like Sandy Bridge. They didn't resolve that until Zen 2. There's been an ongoing question if AMD will ever implement AVX-512. And... now they might be. Firstly, what the hell is AVX3-512? Give or take some variation in spacing and hyphen usage, Intel calls their implementation AVX-512. Is AVX3 even a thing? Are AMD going to do something similar but different? And, going back to my earlier point, supporting the instruction in itself is no guarantee they will provide additional hardware support to make it faster. Note Intel are not so great in this area either offering 1 and 2 unit AVX-512 offerings, only the latter provides the greater speedup.

 

I don't care about machine learning stuff but someone might care about BFLOAT16 too.

[leadeater]

13 minutes ago, porina said:

One module is basically double the bandwidth, and you can (eventually) buy speed grades according to your needs. Do note that with DDR5, one module has two channels at half the width each relative to DDR4. If we take the 12 channels at face value, that's only 6 modules.

Hmm that's a really interesting configuration possibility. Mind you in that case why not 16? Just double what we have now. I wonder if going with a doubled up 6 (12 total in 6 slots) that might actually result in less memory pins and traces compared to now and also physically smaller total socket and memory area on the motherboard. SP3 and 8 memory slots is really tight on narrow boards.

[leadeater]

15 minutes ago, porina said:

Intel calls their implementation AVX-512

Maybe Intel trademarked AVX-512?

[porina]

20 minutes ago, leadeater said:

Hmm that's a really interesting configuration possibility. Mind you in that case why not 16? Just double what we have now. I wonder if going with a doubled up 6 (12 total in 6 slots) that might actually result in less memory pins and traces compared to now and also physically smaller total socket and memory area on the motherboard. SP3 and 8 memory slots is really tight on narrow boards.

I'm not really familiar with server space considerations so your earlier comments were interesting in how this might impact that. Beyond that, we'll just have to see how this rumour goes.

 

Ram bandwidth has always been a tricky one to balance. Some workloads can't get enough of it, others hardly care. IMO the balance has shifted far more towards the cores in recent years, and ram bandwidth has not kept up. Increasing large caches help mitigate that to a degree but at some point data still has to move in and out. 6 module DDR5 would be slightly faster than 8 module DDR4, depending on exact speed grade chosen. Maybe they decide that is still generally sufficient and will provide even bigger caches to help mitigate against needing to increase that.

 

17 minutes ago, leadeater said:

Maybe Intel trademarked AVX-512?

Interesting point. I didn't do a deep search by any means but I haven't been able to determine that either way. I did see "BunnyPeople" is listed as an Intel trademark, if the older members around here remember that one. One feature about trademark law is that you get protection from something that is very similar but not identical. Lots of "ifs" coming up: if AVX-512 is trademarked, and if "AVX3-512" were an attempt to get around that, in my non-lawyer opinion that would be too similar and not be allowed.

 

My suspicion is that assuming AVX3-512 will be a thing, AMD may be getting big enough to not just copy what Intel introduced nearly 4 years ago, but do their own spin on it. It may not be directly compatible even if doing something similar when viewed from afar. Could it be implemented better overall, and/or for the Zen family architecture in general?

[bcredeur97]

9 hours ago, leadeater said:

Just remember, if something sounds too good to be true... it's probably fake or inaccurate.

 

• 12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.
• 12 DDR5 Memory Channels: Possible, unnecessary and also physically and electrically problematic
• 128/160 PCIe 5.0: Likely correct
• 350W-400W TDP: Unlikely, lower TDP parts are already proprietary water cooling only and this is 100W+ more than those

Random meaningless credibility score from random dude on the internet: 3/10

 

I'll take that bet and 10x it, $6 is no fun.

+ the model/SKU number is “7004” 

 

which doesn’t make sense, if it’s a 7000 series then I’d expect the high tier model (which is what this sounds like) to be 79xx 

[leadeater]

4 minutes ago, bcredeur97 said:

which doesn’t make sense, if it’s a 7000 series then I’d expect the high tier model (which is what this sounds like) to be 79xx 

That's actually series generation name, EPYC naming is crap lol. 7000 = product line, xxx4 = generation. 7004 refers to the entire product line of EPYC 4.

 

7741 - Zen

7742 - Zen 2

7743 - Zen 3  (or will be when exits)

7744 - Zen 4  (or will be when exits)

 

[PianoPlayer88Key]

 

I was kind-of hoping for something like this for Zen 5. (I'd prefer to wait for 2nd-gen DDR5 to let the kinks get worked out.) .... I could use the uplift in multi-threaded performance over my i7-4790K.  

 

 

Spoiler

To get that picture...

 

First I sourced a die shot photo of an Epyc or Threadripper.

Then, I measured the chiplets, did some calculations, and "shrunk" them from 7nm to 3nm.  Also did the same with the I/O die, "shrinking" it from 14nm to 5nm.

 

That does assume, however, that the number of cores per die and the number of transistors per core stays the same, the only difference being a manufacturing process shrink (and therefore room to add more chiplet dies.)

That's basically my concept of a "die-shrunk" (7nm -> 3nm chiplets) Threadripper/Epyc Zen 5 CPU.  It assumes staying at 8 cores per chiplet, among other things, but is there any reason AMD couldn't go to like 12 or 16 (or more) cores per chiplet?  (And why is my spellchecker still tagging chiplet, Epyc, Threadripper, etc?)

 

Or, if it could have as many cores as a 2015-2016 era mid-to-high-end GPU....

 

 

 

Or even if I don't go all-out with a high-end Epyc build ... I'm pretty sure I'll need a LOT more RAM than what Ryzen / AM5 will offer.

Spoiler

 

 

 

I'd like that workload, on my next system, to be not much more than little blip at the bottom of the task manager graph.... and that INCLUDES the "Committed" part.  

 

I'd also like plenty of expansion slots, and NO bandwidth sharing.  ("M.2 slot 2 shares bandwidth with 2 SATA ports, when the M.2 slot is populated 2 SATA ports are disabled" would NOT be allowed, for example.)  I think that precludes AM4 and would practically necessitate Threadripper / Epyc for me.

 

 

 

Now if price-to-performance could make a big jump .....

 

 

For example... I was hoping that, for example, about 81 months after I bought my i7-4790K (January 2015 to October 2021), we would have had about a 225-375x or so improvement in price/performance.

Why that much?  Well it's not like it's not been done before....

Spoiler

 

In January 1989, my dad got a 286-10 CPU.  The invoice doesn't say what he paid for the CPU itself (it groups the CPU, motherboard, case, 640k RAM, PSU, FDD, keyboard together at $939), but other places I've looked up give me the idea that it might have been somewhere between about $275-500 or so.

Wikipedia's "instructions per second" page lists the 286-12 at 1.28 MIPS, so my calculations say the 286-10 would have been about 1.0667 MIPS.

 

In October 1995 (81 months after January 1989), he bought an "AMD DX4-120" CPU.  That invoice lists the CPU at $102.

According to WIkipedia, the DX4-100 (Intel) was 70 MIPS.  Assuming equal IPC, the 120 MHz version, I guess, would have been 84 MIPS.  (That assumes equal IPC.)

 

I haven't been able to find other online benchmarks, like older versions of Cinebench, 7-Zip, 3DMark, Blender, etc. comparing the two CPUs.

 

 

 

 

 

Also I'm probably going to wait until at least November 2022 to buy (to hopefully give time for 2nd-gen DDR5 stuff to come out) ... I may even wait a bit longer, like into 2024 or so.  (I don't want to wait TOO long, though, I'd like to be able to keep my same platform for a long time.  I want to be able to upgrade my system ONE part at a time - maybe 1 part every 6 to 48 months or so.)  I could anticipate replacing my future Tier A+ 80+ Titanium PSU twice (due to its death) before I replace the motherboard, for example.)

 

 

For a more real-world example.... I'd like my next CPU, that's the same price (or less) as the 4790K was ($329 I think when I bought it)....

to be as fast at transcoding video to 4K H.265 at max settings in Handbrake, as the 4790K is at transcoding audio with a multi-threaded GUI version of LAME.

 

Yesterday it took about a minute to transcode an hour of audio to mp3, at 320kbps.  (The audio was extracted from several dozen or so videos, ranging from 2 seconds to 6 minutes long.)

 

The videos are currently in progress.  (The queue is in ascending order of playtime / original file size.  All are sourced from my Panasonic FZ1000, and are 4K H.264 videos.)

A couple of the most recent ones took about an hour to transcode, for a minute of footage.

 

That's faster than I remember, though - I thought it took about 4 days to transcode a 4-minute video when I tested it about a year or 2 ago or so, but I didn't properly document / screenshot its progress at the time.  (Right now I have a camera trained on the monitor in time lapse mode - 1 shot every 60 seconds, to later be compiled into a 60fps video and added to a 1080p60 video I shot of the PC doing the audio transcode.)
(I think that same video is later in the queue, we'll see what happens this time.)

 

Spoiler

Maybe it's a different version of Handbrake, or maybe I'm using different settings?  (I'm using Q / RF = 0, keyint=1, lossless=1, placebo, 10-bit H.265, no QuickSync (idk if that was available for H.265; also I don't have my GTX 1060 3GB installed right now so it's just running on the CPU), with resolution and framerate at 3840x2160 and source, respectively.)  I forget what settings I was using the previous time, but I thought it was something like what I'm using now - basically maxing everything out for a "worst-case" scenario.

 

In the real world, I probably wouldn't go quite so crazy with the encoding settings, but I would still want my new system to be that much faster, taking the settings into account.

Spoiler

For example....

 

With the settings I'm using right now, it seems to be taking about an hour to transcode a minute of footage.

If more "reasonable" settings would have it being real-time (for example a minute to transcode a minute of footage)....

 

Then, if a new CPU takes, say, a minute to transcode an hour of footage at max / insane settings,

Then it should take about 1 second to transcode 1 hour of footage at more "reasonable" settings.

 

(As I mentioned earlier, I'd want the new CPU to transcode video as fast as the old one transcodes audio.
Old CPU in yesterday's test took a minute to transcode an hour of audio, or about 60x realtime.

Old CPU in the current ongoing test took about an hour or so to transcode a 1-minute video, or about 1/60th realtime.

 

Therefore, I'd like the new CPU to be (60 / 1/60 , if I got that right) ... 3600x faster.

 

 

Better yet, it would use the same ratio based on what I thought I remembered from the Handbrake run a year or 2 ago....

If it took 4 days to transcode a 4-minute video, then that would be, if my calculations are correct, about 1/1440th realtime.

Combine that with the 60x realtime for the old CPU to transcode audio ....

And that would mean the new CPU would be, if I'm calculating right .... about 86,400 times faster.

 

Of course that would mean it'd be more expensive than the $330 I paid for my 4790K.  (But hopefully not like several thousand bucks or even upwards of $10K....  ) 

 

 

 

By "reasonable" settings - I don't mean something intended for download / streaming, or casual watching, I'm still talking about archival / editing quality - lossless, all I-frames, etc, just not going super crazy with H.264/265 with RF=0 and placebo, and possibly using a different codec entirely, but idk what one I'd want to use.  I'd prefer a codec with wide compatibility, not something that's specific to one or 2 vendors over 1 or 2 generations.

Also I'd hope more reasonable settings would give a more reasonable file size, even if the transcoded file is still bigger than the original due to making sure it's "archival / editing" quality.  (One of the 1-minute videos that it's completed transcoding is about 16 GB, from about a 750-800 MB or so source file.)

 

 

 

 

 

Also on a "smaller CPU" scale .... when it's time for me to go to my next convention (maybe TwitchCon 2022? or hopefully I can someday go to a PAX?) ... I'd like to get a pocketable PC with a full-desktop OS (like a GPD Pocket type device) that is capable of running a bunch of VMs, like, a LOT more than this.

Spoiler

The 5th currently-empty Discord is there so it's ready to go when I hit the 100-server-per account limit on my 4th Discord account, which should be quite soon - I thought I had already hit it but apparently got dropped from a few servers or they went offline, idk.

Also I'd need to up the per-VM performance a bit -- since I updated Xubuntu to 20.04 or whatever it was (from 17.10) it's now very sluggish trying to play video on, say, youtube.  (Each VM is assigned 2.75 GB RAM with 1 CPU thread from my 6700K ... it was fine when I originally set it up.  )

Also hopefully that GPD Pocket like device would be somewhere around $700-800 or so.... I'd imagine it using a 5nm or 3nm low-TDP AMD CPU  

Basically I want to be able to keep track of all my discords "on the go" without having to log in & out one at a time like I'd have to do on my phone.  (The only way I know how to do that is running DIscord in multiple VMs, and the only way I know how to run multiple VMs effectively is with an actual desktop UI OS.)

 

 

(I wish there was a way to "preview" posts so I'd get a better idea of how long it is.... the spoilers being expanded in the edit window kinda throws me off....)

[Loote]

10 hours ago, leadeater said:

Just remember, if something sounds too good to be true... it's probably fake or inaccurate.

 

• 12 Chiplets: Possible but also physically and electrically problematic [Edit again] (Read the source, actually claiming 12 chiplets. I should read sources first lol) 12 core CCD makes far more sense than 8 core CCD with Zen 4 if this many cores.
• 12 DDR5 Memory Channels: Possible, unnecessary and also physically and electrically problematic
• 128/160 PCIe 5.0: Likely correct
• 350W-400W TDP: Unlikely, lower TDP parts are already proprietary water cooling only and this is 100W+ more than those

Random meaningless credibility score from random dude on the internet: 3/10

 

I'll take that bet and 10x it, $6 is no fun.

It may be a true leak... of things they considered. Maybe even things they will make available, but possibly not everything in 1 CPU model. Perhaps special editions that cost a hefty premium because if they are able to fit 12 channels in there and there is a client ready to pay extra for everything that comes with such feature, it can be a useful for the future too. If they keep on increasing the core count sooner or later those channels will make more and more of a difference, so training in this gen could lead to it becoming a standard in, say, 128 core Zen 5.

[CTR640]

19 hours ago, HelpfulTechWizard said:

can't wait for a 192core server video from ltt.

But will he drop it?

[jagdtigger]

15 hours ago, James Evens said:

That's like 3 VM ware licences per CPU.

Time to look for something better then....

[Helpful Tech Witch]

1 hour ago, CTR640 said:

But will he drop it?

yeah. he will kill one of the ccxes this time tho.

[Murasaki]

Neat! Can't wait to have one!

 

*phone rings*

*answers*

 

Oh? I can't? Oh. Okay.

[RoseLuck462]

Can't wait to see the Mini-ITX builds!