MilanX


Also;
Looks like the rumors of a Dense version of Zen4 for servers where not crazy. Zen4c. So ya AMD is going to have stupid high core count versions of Zen4c 128 core server chips anyway. The other rumor about them being the e cores for Zen 5... who knows. :)
 
  • Like
Reactions: noko
like this
I thought it was weird that I was getting so many notifications about the RTX 4000 series being 2-3x faster than the RTX 3000 series, Now I get it. AMD had a press day so RTX 4000 series info gets "leaked" makes sense now.

I'll have to watch this later, AMD has been woefully silent in the Datacenter GPU department for a long while so I am interested to see what they bring to the table.
 
Yea using Cache stacking on an MCM design video card. Really has me interested what those new cards would be like.

Maybe Nvidia is really worry if they leak 4k series info. With chip shortages it won't matter anyway lol
 
Brackle said:
Yea using Cache stacking on an MCM design video card. Really has me interested what those new cards would be like.

Maybe Nvidia is really worry if they leak 4k series info. With chip shortages it won't matter anyway lol
Click to expand...
Not likely, NVidia's next chips is also MCM, and based on their numbers I'm not terribly thrilled with what I am seeing.

If you compare the stats from the existing MI100, and compare them to the announced MI250 it's going from one compute tile, to 2 compute tiles, basically doubling the number of cores on an updated process and getting a clock speed increase. So if you are putting down 2x as much silicon you should hope for at least a 2x performance increase, the updated blah blah blah should also get some performance too. But the card is going from 300W to 500W, it's gonna be big and hot.

Still, this is good progress all the same, better than 2x the performance at less than 2x the power increase is a win no matter how you look at it, I just somehow hoped for better, which really is just unrealistic on my part.
I just hope that TSMC has found some way to get AMD more fab time because now they just have another market they are trying to get back into, and these are big silicon parts which means if their TSMC supply remains the same they are just decreasing from the rest, and the stuff they are cutting back on sure as hell won't be the consoles.
 
The MilanX performance increase with V-Cache for those workloads over Milan should transfer over to the desktop when available. Did AMD throw a curve ball with the 15% gaming performance increase for VCache on the desktop, wording to think it only accelerates gaming for the most part? Now it appears it is only 15% on games since it pushes the tests to GPU limiting conditions and does not indicate the overall improvement the V-Cache will push in normal type computer programs. Going have to sell that Intel stock recently bought, bad call on my part :(.
 
noko said:
The MilanX performance increase with V-Cache for those workloads over Milan should transfer over to the desktop when available. Did AMD throw a curve ball with the 15% gaming performance increase for VCache on the desktop, wording to think it only accelerates gaming for the most part? Now it appears it is only 15% on games since it pushes the tests to GPU limiting conditions and does not indicate the overall improvement the V-Cache will push in normal type computer programs. Going have to sell that Intel stock recently bought, bad call on my part :(.
Click to expand...
Your Intel stock is fine… expect that to go up especially with them expected to get some of those lucrative military contracts that previously all went to Texas Instruments.

AMD makes great stuff but they play in too many fields to become a serious contender in any of them, they are severely hampered by TSMC’s limited output capacity.
 
noko said:
The MilanX performance increase with V-Cache for those workloads over Milan should transfer over to the desktop when available. Did AMD throw a curve ball with the 15% gaming performance increase for VCache on the desktop, wording to think it only accelerates gaming for the most part? Now it appears it is only 15% on games since it pushes the tests to GPU limiting conditions and does not indicate the overall improvement the V-Cache will push in normal type computer programs. Going have to sell that Intel stock recently bought, bad call on my part :(.
Click to expand...
Intel is still a good bet as a stock. They are working over a bunch of Western governments for handouts to jump start their Fab business rehabilitation. What happens over the next couple years probably isn't all that important in the grand scheme due to silicon shortages. Even if AMDs Zen 3 refresh stomps on them. (which I expect it will within reason... I'm not expecting more then 15-25% depending on the work load) It will be what happens in the 3-5 years after that that really matter. Investors will keep Intels stock from feeling too much second place pain... as long as countries continue agreeing to 10s of billions in handouts for Intel to build new fabs, and shortages keep Intel CPUs moving anyway.

From a PR perspective yes the next few years will belong to AMD. The first US Exoscale will be 100% AMD... and before the shine wears on that win they will probably have a Zen4 based systems to talk about. The Zen4c stuff is the thing of nightmares for Nvidia. I know their ARM purchase is looking iffy still Nvidia is clearly banking on a ARM future... and AMD has just shown the industry a road map clearly aimed at ARM servers. Zen4c is a huge announcement. They are planning a more ARM like x86 server core that can scale... 128 core Zen4c servers are going to compete very well vs ARM.

The rumors are out that Zen4c will be the efficient cores included in Zen 5, who knows on that one Zen5 is still a good ways away. Its an exciting thought though... with chiplet tech and 3D stacking. A Zen 5 could have one chiplet with a Zen 5 core that is big and power hungry and perhaps have even fewer cores onbaord. Imagine a 4 core chiplet with cores that are 20% larger and more powerful with a much expanded thermal envelope. (Just imagine Zen cores with a IPC bump but also an actual clock speed bump AMD could finally go full ham and have actual high clock speed zen cores) While a second 16 core chiplet of Zen4c cores could be put on the same package.... with both sharing 3D cache space. I'm sure Zen 4 will be very cool... and bring all the DDR5/PCIe5 stuff and perhaps an even better use of the 3D stacking tech. IMO its Zen5 where things will get interesting. Seeing where AMD is going I would say Intels next real die shrink with meteor lake is going to have to deliver, on performance and delivery. Intel can't afford to have their "4nm" slip by even a few months this time. Its supposed to be in production by the end of 2022 and it better be. If AMD delivers a Zen5 that I believe may be one of those rare true generational leaps... Intel can't afford to have their fab keep pushing their engineers to backport things to old processes.
 
  • Like
Reactions: noko
like this
ChadD said:
Intel is still a good bet as a stock. They are working over a bunch of Western governments for handouts to jump start their Fab business rehabilitation. What happens over the next couple years probably isn't all that important in the grand scheme due to silicon shortages. Even if AMDs Zen 3 refresh stomps on them. (which I expect it will within reason... I'm not expecting more then 15-25% depending on the work load) It will be what happens in the 3-5 years after that that really matter. Investors will keep Intels stock from feeling too much second place pain... as long as countries continue agreeing to 10s of billions in handouts for Intel to build new fabs, and shortages keep Intel CPUs moving anyway.

From a PR perspective yes the next few years will belong to AMD. The first US Exoscale will be 100% AMD... and before the shine wears on that win they will probably have a Zen4 based systems to talk about. The Zen4c stuff is the thing of nightmares for Nvidia. I know their ARM purchase is looking iffy still Nvidia is clearly banking on a ARM future... and AMD has just shown the industry a road map clearly aimed at ARM servers. Zen4c is a huge announcement. They are planning a more ARM like x86 server core that can scale... 128 core Zen4c servers are going to compete very well vs ARM.

The rumors are out that Zen4c will be the efficient cores included in Zen 5, who knows on that one Zen5 is still a good ways away. Its an exciting thought though... with chiplet tech and 3D stacking. A Zen 5 could have one chiplet with a Zen 5 core that is big and power hungry and perhaps have even fewer cores onbaord. Imagine a 4 core chiplet with cores that are 20% larger and more powerful with a much expanded thermal envelope. (Just imagine Zen cores with a IPC bump but also an actual clock speed bump AMD could finally go full ham and have actual high clock speed zen cores) While a second 16 core chiplet of Zen4c cores could be put on the same package.... with both sharing 3D cache space. I'm sure Zen 4 will be very cool... and bring all the DDR5/PCIe5 stuff and perhaps an even better use of the 3D stacking tech. IMO its Zen5 where things will get interesting. Seeing where AMD is going I would say Intels next real die shrink with meteor lake is going to have to deliver, on performance and delivery. Intel can't afford to have their "4nm" slip by even a few months this time. Its supposed to be in production by the end of 2022 and it better be. If AMD delivers a Zen5 that I believe may be one of those rare true generational leaps... Intel can't afford to have their fab keep pushing their engineers to backport things to old processes.
Click to expand...
Intel/IBM 2nm is still on track for a mid-2024 launch date and it is a process that is working as expected inside of lab environments. TSMC's 3nm is also expected around the same time to be going live, the announcements for Zen 4 and 5 are great but production end of 2022 means availability for Q2 in 2023, and keeping to an 18-month launch cycle puts Zen 5 cleanly into 2025, there is a lot of time there for things to swing wildly. Those chips are also dependent on TSMC's 5nm process which is even more constrained than the current 7nm process, which very well could mean few parts released to the consumer space.

NVidia's only real goal for purchasing ARM is to phase out Mali and get their stuff in there as the new defacto mobile standard, but all their other plans are still possible with the existing licensing structure and in no way is dependent on their purchasing of ARM, their DoE contract for the Grace Supercomputer is still on-track to go live in 2023, and while not necessarily Exoscale it will still be the fastest AI cluster out there (publically, I am sure in some government bunker is one that dwarfs it).

I am happy though that AMD scored a supercomputer contract, CUDA has become too dominant in the data center and an alternative is needed, this could be the shot in the arm that GPUFort needs so it can one day actually stand against CUDA as a viable alternative.

In regards to Intel 4 (previously 7nm) it is already shipping custom commercial parts those projects are on track and nearing completion so there shouldn't be any significant delays there, Inte's 10nm process was a huge setback and overall a large outlier, the nanoribbon transistor design would have guaranteed them the lead in density and power efficiency, but it never worked outside of the test labs. And given Intel's former practice of designing their chips around their production roadmaps when that design failed to materialize, it set them back 3 generations of CPU designs so the 10nm failures don't really have any impact on their 7, 4, 3, and 20A roadmaps.

Intel's current leadership is also not above asking for help, they have already contracted out large parts of TSMC's 5nm and 3nm processes where they may be actually getting more of those wafers than AMD does at this point, what they are slotted for is still a bit of a mystery but GPU's and mobile CPU's are the current rumor mill favorites.

AMD has utterly embarrassed Intel (or rather Intel embarrassed themselves and AMD just put it under the spotlight), Intel isn't taking that lightly and they are removing the gloves, I really look forward to seeing what the next 3 years of actual competition brings, but I am sure that both Intel and NVidia are confident in their roadmaps to not disappoint.
 
Any word on next gen Threadrippers?

I'm still a little bummed they delayed my drop-in TRX40 upgrade.
 
  • Like
Reactions: mikeo
like this
Zarathustra[H] said:
Any word on next gen Threadrippers?

I'm still a little bummed they delayed my drop-in TRX40 upgrade.
Click to expand...
Supposedly the Threadripper 5000 series parts will be launching mid to late 2022 and will retain compatibility for the sTRX4 and sWRX8 boards, the release of those parts is supposedly going to line up with the Intel HEDT components. The rumor mill suggests that AMD pushed the launch back so they could get the 3D cache on them because they are expecting Intel's upcoming Sapphire Rapids parts to put up a good fight.
 
Last edited:
Lakados said:
Supposedly the Threadripper 5000 series parts will be launching mid to late 2022 and will retain compatibility for the sTRX4 and sWRX8 boards, the release of those parts is supposedly going to line up with the Intel HEDT components. The rumor mill suggests that AMD pushed the launch back so they could get the 3D cache on them because they are expecting Intel's upcoming Sapphire Rapids parts to put up a good fight.
Click to expand...

A just plain delay would be disappointing.

A delay due to adding something new and exciting I can live with.
 
Zarathustra[H] said:
A just plain delay would be disappointing.

A delay due to adding something new and exciting I can live with.
Click to expand...
I'm just happy they are keeping sTRX4, because I hate dealing with Lenovo, and having to deal with them for another Threadripper system would make me give up on the platform and just buy some Xeons instead.
 
Lakados said:
I'm just happy they are keeping sTRX4, because I hate dealing with Lenovo, and having to deal with them for another Threadripper system would make me give up on the platform and just buy some Xeons instead.
Click to expand...

Yeah, I'm pleased because this TRX40 motherboard wasn't exactly cheap. Not having to buy another one is a positive.

At the same time, sticking with TRX40 will preclude any mid cycle refreshes to move up to PCIe Gen 5, but that wasn't particularly likely to begin with, and really not necessary yet, IMHO.

You bought a Threadripper system from Lenovo?
 
Last edited:
Zarathustra[H] said:
Yeah, I'm pleased because this TRX40 motherboard wasn't exactly cheap. Not having to buy another one is a positive.

At the same time, sticking with TRX40 will preclude any mid cycle refreshes to move up to PCIe Gen 5, but that wasn't particularly likely to begin with, and really not necessary yet, IMHO.

You bought a Threadripper system from Lenovo?
Click to expand...
I have 3 of them, they have been problematic. Bios updates finally got them stable but for a good while they would just power off under load it was very not fun to troubleshoot. And Lenovo support was slightly better than useless on the matter. They kept blaming power, so when I put them on UPS’s and it continued they thought it was a PSU issue, when that didn’t fix it they thought it was software related. Finally one of the Bios updates improved memory related things and it resolved it self from that but Before it got to that point I was going back and forth with my rep on a refund and having them pay to haul them away.
 
Lakados said:
I have 3 of them, they have been problematic. Bios updates finally got them stable but for a good while they would just power off under load it was very not fun to troubleshoot. And Lenovo support was slightly better than useless on the matter. They kept blaming power, so when I put them on UPS’s and it continued they thought it was a PSU issue, when that didn’t fix it they thought it was software related. Finally one of the Bios updates improved memory related things and it resolved it self from that but Before it got to that point I was going back and forth with my rep on a refund and having them pay to haul them away.
Click to expand...
Jeez.

And people go to the likes of Lenovo to NOT have to deal with this kind of DIY troubleshooting.

Someone screwed the pooch on that design.
 
Zarathustra[H] said:
Jeez.

And people go to the likes of Lenovo to NOT have to deal with this kind of DIY troubleshooting.

Someone screwed the pooch on that design.
Click to expand...
My Dell EPYC’s had a similar issue but they soft lock, not exactly what you want your VM hosts to be doing…. But again fixed with bios updates that dealt with power states.
 
Lakados said:
My Dell EPYC’s had a similar issue but they soft lock, not exactly what you want your VM hosts to be doing…. But again fixed with bios updates that dealt with power states.
Click to expand...

I guess there is some learning involved for the big guys to get the new AMD systems right...

Is this in an enterprise setting? Because otehrwise I'd just buuild my own and never deal with a large OEM ever again.
 
Zarathustra[H] said:
I guess there is some learning involved for the big guys to get the new AMD systems right...

Is this in an enterprise setting? Because otehrwise I'd just buuild my own and never deal with a large OEM ever again.
Click to expand...
Yeah, enterprise. But I don't know how much of it was AMD and how much was the OEMs' because AMD controls 90% of their Bios and doesn't give the OEM's much to play with there. In either event, they are all up and running and they likely stay AMD for the future as all I need is core count and ram.
 
Lakados said:
Yeah, enterprise. But I don't know how much of it was AMD and how much was the OEMs' because AMD controls 90% of their Bios and doesn't give the OEM's much to play with there. In either event, they are all up and running and they likely stay AMD for the future as all I need is core count and ram.
Click to expand...

Hmm. I'm thinking it must have been in the parts that are not controlled by AMD or we would have seen issues like these in boards from the likes of Asus, Gigabyte, MSI etc. as well, which at least I haven't. (I've seen other problems with my Gigabyte TRX40 board to the point where I am trying to get them to give me my money back, but that's a different story)

I'm wondering if there is just something different enough about designing a board for AMD that threw them off, which DIY brands who have more AMD experience have already learned in past designs.
 
Zarathustra[H] said:
Hmm. I'm thinking it must have been in the parts that are not controlled by AMD or we would have seen issues like these in boards from the likes of Asus, Gigabyte, MSI etc. as well, which at least I haven't. (I've seen other problems with my Gigabyte TRX40 board to the point where I am trying to get them to give me my money back, but that's a different story)

I'm wondering if there is just something different enough about designing a board for AMD that threw them off, which DIY brands who have more AMD experience have already learned in past designs.
Click to expand...
I do recall the desktop chips having issues with the C6 power states causing lockups and crashes when running stock speeds, the "fix" for most AMD boards is to not run stock speeds and run their overclocking tools to OC and ultimately stabilise the systems.
 
You must log in or register to reply here.
Back
Top