Check Out the Delidded AMD Rome CPU

I'd love to know what type of matrix is wiring those chiplets to the I/O die. Will the outer dies have more latency to the RAM controller as compared to the inner dies?
Early rumors about the 8+1 showed a different layout. One central die with 2dies on every side. Which seems like logically the shortest path, but that doesn't really account for the connectivity layout of each chiplet.
 
mbelue said:
I'd love to know what type of matrix is wiring those chiplets to the I/O die. Will the outer dies have more latency to the RAM controller as compared to the inner dies?
Early rumors about the 8+1 showed a different layout. One central die with 2dies on every side. Which seems like logically the shortest path, but that doesn't really account for the connectivity layout of each chiplet.
Click to expand...

From the diagrams I saw, all the CPU dies are directly connected to the I/O die, which means each CPU die should have the same latency to the I/O die as all the rest. The positioning on the interposer doesn't mean that the latency will be any different. Otherwise, memory DIMMs would have varying latency between the chips on the ends from the ones in the center, but they don't. I expect that AMD would arrange the leads such that varying latency would not be a problem.

I'm mainly curious about the PCIe controllers. The diagram of the I/O die shows "I/O" as two parts, but it doesn't say specifically PCIe, and some have assumed that means the PCIe controllers are on the CPU dies. I'm not so sure about that, but I can't say for sure. That certainly would cause varying latency issues.
 
izWYIWR.jpg
 
mbelue said:
I'd love to know what type of matrix is wiring those chiplets to the I/O die. Will the outer dies have more latency to the RAM controller as compared to the inner dies?
Early rumors about the 8+1 showed a different layout. One central die with 2dies on every side. Which seems like logically the shortest path, but that doesn't really account for the connectivity layout of each chiplet.
Click to expand...
dgingeri said:
From the diagrams I saw, all the CPU dies are directly connected to the I/O die, which means each CPU die should have the same latency to the I/O die as all the rest. The positioning on the interposer doesn't mean that the latency will be any different. Otherwise, memory DIMMs would have varying latency between the chips on the ends from the ones in the center, but they don't. I expect that AMD would arrange the leads such that varying latency would not be a problem.

I'm mainly curious about the PCIe controllers. The diagram of the I/O die shows "I/O" as two parts, but it doesn't say specifically PCIe, and some have assumed that means the PCIe controllers are on the CPU dies. I'm not so sure about that, but I can't say for sure. That certainly would cause varying latency issues.
Click to expand...

Physical distance does increase latency but there are ways to reduce or mitigate this. Some boards can show this as not all vendors take the time to work on this issue or even care about it.

On the cpu side it will probably be somewhat buffered and appear as less of an issue due to the imc and or nature of the work being done. Also it may be that to help with this cores closer to the imc have preference to use memory further away while cores further from the imc prefer memory that is closer.
 
Its actually larger than that, the other half of it you just cant see as it exist in a quantum state about 2 inches to the right :sneaky:
 
That is about $640k SQL Enterprise license due to roughly $1000 per core pricing. I hope they pressure Microsoft on this (for non-VM needed environments). Microsoft are dweebs since people are pushed into free open source products from their own greed. Open up a support case explaining this or an enhancement request and it will be closed "buy our cloud services its great..." and you will then be spammed with cloud promotions.
 
My sister-in-law Brenda has an old stank-ho Nutburst P4 Gateway that she refuses to get rid of because 'she only uses it for word processing'. And that's fine. It runs Windows 2000, it has 3GB of RAM, it has an R300 graphics card and its original male mouse. I know this because I emailed notes to myself from the last time I worked on it. I hate it, it is an atrocity in beige, I hate working on it, I have no idea why the thing is still alive. Thankfully, it doesn't connect to the internet.

So today I'm surfing on the TV and I'm watching this video, and my mom comes in behind me and asks, "What is that?"

And I tell her, "It's a new EPYC processor from AMD, it's huge, it's probably the largest commercial processor in the world. It has 64 cores and can run 128 threads."

And she asks, "Would something like that fix Brenda's computer?"

I snorted so hard that I farted at the same time.
 
Last edited:
oROEchimaru said:
That is about $640k SQL Enterprise license due to roughly $1000 per core pricing. I hope they pressure Microsoft on this (for non-VM needed environments). Microsoft are dweebs since people are pushed into free open source products from their own greed. Open up a support case explaining this or an enhancement request and it will be closed "buy our cloud services its great..." and you will then be spammed with cloud promotions.
Click to expand...

Back in my old company, I scratched my head as well when MS switched to per core licensing, I do hope they bring down their prices, honestly they are still cheaper than Oracle.
 
"Latency" on a PCB is about 160ps per inch, just for reference.

It can be as high as 190ps on inner layers; that's roughly 5x the clock frequency, at 100Mhz; or 5 inches per clock tic.
 
Forget the size of the processor, look at the size of that Diamond on her finger!

Congratulations AMD for an Industry first!
 
The real implications of AMD's chiplet strategy will bear fruit when it comes to the GPU side. I'm expecting AMD to follow a similar principle with Navi and to scale up performance, they just need to increase the number of dies in a package. The base line consumer system is looking more like it'll be a CPU die, an IO die and a GPU die, all interconnected via Infinity Fabric. The difference between a 16 core consumer chip and an 8 CPU + GPU consumer product would be simply swapping that one die out in the design. This radically simplifies development costs (though at higher manufacturing costs) at a time when AMD was heavily constrained.
 
Some of you knock this chip but watch this thing fucuck Intel with its tiny 7nm giant schlong. I'm excited for the 7nm process
 
Venice was the last time I had an AMD CPU in my PC. Looking forward to Rome
 
mbelue said:
I'd love to know what type of matrix is wiring those chiplets to the I/O die. Will the outer dies have more latency to the RAM controller as compared to the inner dies?
Early rumors about the 8+1 showed a different layout. One central die with 2dies on every side. Which seems like logically the shortest path, but that doesn't really account for the connectivity layout of each chiplet.
Click to expand...
There are a few ways they can get the latency about the same, aside from doing zig-zag traces like you see in larger pcbs. They could rotate one of the pair of chiplets 180° and have all the traces come out between the chiplets with the traces on different layers, or they could have them come out the top and bottom of each chiplet with the traces on the same layer (and just go to a different side of the central chip). It's also possible they don't terminate at the edge of the chip....
 
seanreisk said:
My sister-in-law Brenda has an old stank-ho Nutburst P4 Gateway that she refuses to get rid of because 'she only uses it for word processing'. And that's fine. It runs Windows 2000, it has 3GB of RAM, it has an R300 graphics card and its original male mouse. I know this because I emailed notes to myself from the last time I worked on it. I hate it, it is an atrocity in beige, I hate working on it, I have no idea why the thing is still alive. Thankfully, it doesn't connect to the internet.

So today I'm surfing on the TV and I'm watching this video, and my mom comes in behind me and asks, "What is that?"

And I tell her, "It's a new EPYC processor from AMD, it's huge, it's probably the largest commercial processor in the world. It has 64 cores and can run 128 threads."

And she asks, "Would something like that fix Brenda's computer?"

I snorted so hard that I farted at the same time.
Click to expand...

Holy shit I had to log into my account after not having done so in a while to reply to this.

Fucking hilarious.
 
seanreisk said:
My sister-in-law Brenda has an old stank-ho Nutburst P4 Gateway that she refuses to get rid of because 'she only uses it for word processing'. And that's fine. It runs Windows 2000, it has 3GB of RAM, it has an R300 graphics card and its original male mouse. I know this because I emailed notes to myself from the last time I worked on it. I hate it, it is an atrocity in beige, I hate working on it, I have no idea why the thing is still alive. Thankfully, it doesn't connect to the internet.

So today I'm surfing on the TV and I'm watching this video, and my mom comes in behind me and asks, "What is that?"

And I tell her, "It's a new EPYC processor from AMD, it's huge, it's probably the largest commercial processor in the world. It has 64 cores and can run 128 threads."

And she asks, "Would something like that fix Brenda's computer?"

I snorted so hard that I farted at the same time.
Click to expand...

God fucking damn you. I'm laughing my ass off in work and getting strange looks

:D
 
dgingeri said:
I'm mainly curious about the PCIe controllers. The diagram of the I/O die shows "I/O" as two parts, but it doesn't say specifically PCIe, and some have assumed that means the PCIe controllers are on the CPU dies. I'm not so sure about that, but I can't say for sure. That certainly would cause varying latency issues.
Click to expand...

"The chipsets contain no memory controller or PCIe, but are instead tied to a central IO chip built on 14nm. The IO chip has 8 channels of DDR4 plus support for PCIe 4.0. The company connects each chiplet to the IO chip via a 2nd-generation reduced latency Infinity Fabric."

https://www.pcworld.com/article/331...e-rome-server-cpu-tells-us-about-ryzen-2.html
 
dgz said:
Venice was the last time I had an AMD CPU in my PC. Looking forward to Rome
Click to expand...

I just built a Ryzen 2700x system...MY last AMD system was a K5! The new Ryzen chips are DAMN fast...you will be amazed if you are upgrading anything older than a year or two, REGARDLESS OF BRAND! The "Intel is better" crowd will tout 5Ghz and an extra 10-15fps at the most common resolutions used for gaming, but for anyone who isn't an uber-gamer, the AMD chips perform quite nicely, ESPECIALLY for non-gaming applications.

The BEST thing about Ryzen, though, is that I can upgrade to the new 3rd gen chips, and the 4th gen after that, without replacing my motherboard and RAM! I can probably get a couple hundred dollars for my current CPU on eBay when I upgrade too...so upgrading with the latest tech is pretty damn cheap!

What AMD showed with "Rome" is going to be their new "Epyc" server and workstation chips, but we should see the technology "trickle down" into Threadripper and Ryzen as well. Those chips are probably going to be what you see in your next PC.
 
Can we talk about the video quality here and how I could barely hear anything they said? Lol
 
I have to hand it to AMD...they now how to make a cheap processor.....That's not necessarily an all good thing.
 
seanreisk said:
And she asks, "Would something like that fix Brenda's computer?"

I snorted so hard that I farted at the same time.
Click to expand...

Tell your mom the internet says thank you. Needed a good laugh :ROFLMAO:

Note to self: Said something nice about someone's mom, on the internet today. Man how things have changed!
 
ecktt said:
I have to hand it to AMD...they now how to make a cheap processor.....That's not necessarily an all good thing.
Click to expand...

So you like Intel's option?
Oh wait.............
How's it not a good thing? cheaper to make = more profits for them (to fund R&D for future stuff) and cheaper price for us... (except for Apple)
 
jnemesh said:
I just built a Ryzen 2700x system...MY last AMD system was a K5! The new Ryzen chips are DAMN fast...you will be amazed if you are upgrading anything older than a year or two, REGARDLESS OF BRAND! The "Intel is better" crowd will tout 5Ghz and an extra 10-15fps at the most common resolutions used for gaming, but for anyone who isn't an uber-gamer, the AMD chips perform quite nicely, ESPECIALLY for non-gaming applications.

The BEST thing about Ryzen, though, is that I can upgrade to the new 3rd gen chips, and the 4th gen after that, without replacing my motherboard and RAM! I can probably get a couple hundred dollars for my current CPU on eBay when I upgrade too...so upgrading with the latest tech is pretty damn cheap!

What AMD showed with "Rome" is going to be their new "Epyc" server and workstation chips, but we should see the technology "trickle down" into Threadripper and Ryzen as well. Those chips are probably going to be what you see in your next PC.
Click to expand...

I have a 13 inch Ryzen laptop. This little thing is fucking amazing. Sadly, it doesn't run Linux properly. It doesn't even spin its fan until I run something really demanding. Yes, I mean actual games. 1080/60 fps video is nothing to this thing
 
You must log in or register to reply here.
Back
Top