Question AMD's chicken and egg Threadripper problem?

[igor_kavinski]

From another thread:

AMD could easily make a 24/32 core Ryzen but they choose not too.

Instead AMD also wants to make profit from Threadripper sales. shocker

I have to wonder. Do they really make that much profit from Threadripper? It seems to be their lowest priority product line in terms of launch cadence. For Zen 5, they released consumer CPUs first, then came Turin and now it's time for Shimada Peak but not even a launch event yet or any juicy leaks of whether it will have V-cache or not.

Threadripper could be a LOT more popular if AMD simply released it in quad channel form. But they don't want to because it is supposedly expensive. BUT how do they bring their cost down if they can't sell it in volume which requires them lowering the price of entry?

Chicken and egg. Something's gotta give.

Come on, AMD. Be bold. Invest in the R&D for a quad channel Threadripper with minimum $400 motherboards and then see your Threadripper sales soar in a year or two!

 

[SteinFG]

Threadripper could be a LOT more popular if AMD simply released it in quad channel form.

They did, 7960x/70x/80x are quad channel and only work in quad channel TRX50 motherboards.
And some of them are under 700 dollars. WS from asrock or Aero from gigabyte. Considering X870E prices, it's not gonna get better.

 

[poke01]

Still too expensive, 7960X is $1500 for 24 cores and work in expensive motherboards.


Hopefully by the time we get AM6, 32 cores are top end Ryzen parts and X series motherboards are quad channel. By then bandwidth will become super important

 

[DrMrLordX]

Don't expect quad channel on consumer motherboards. That being said, anything beyond 16c is going to be starved for bandwidth. 16c is already pushing it.

 

[Shmee]

What we need is a return to glorious HEDT systems. Like we had with X99. TR is really more workstation currently.

 

[Gideon]

What we need is a return to glorious HEDT systems. Like we had with X99. TR is really more workstation currently.

This! there would definitely be a niche for it. But so far it seems it has not been deemed big enough for any of the OEMs to care.

 

[igor_kavinski]

People obviously have money to burn, when they buy 7950X3Ds and turn off the HF CCD to get better gaming performance. Give them a single CCD 16 cores at the very least. $800 for the CPU and something like $400-$650 for the mobo with quad channel DDR5 and AMD/OEMs WILL have sales.

 

[poke01]

People obviously have money to burn, when they buy 7950X3Ds and turn off the HF CCD to get better gaming performance. Give them a single CCD 16 cores at the very least. $800 for the CPU and something like $400-$650 for the mobo with quad channel DDR5 and AMD/OEMs WILL have sales.

People buy $2000+ GPUs, AMD can definitely do more here like you say.

 

[MS_AT]

People obviously have money to burn, when they buy 7950X3Ds and turn off the HF CCD to get better gaming performance. Give them a single CCD 16 cores at the very least. $800 for the CPU and something like $400-$650 for the mobo with quad channel DDR5 and AMD/OEMs WILL have sales.

Then they will burn it on the Threadrippers You forget Threadrippers bring with them lots of IO capabilities, all these things have to be routed on the PCB. PCIe 5 has tighter requirements what increases cost. The socket itself takes more space on the PCB what makes it harder and more expensive to design something that is smaller than ATX.

The alternative would be to produce another niche SKU, without the big IOD, just with 4 mem channels, up to 32 cores maybe, with severly limited IO compared to threadrippers but somewhat bigger than what AM5 provides. Who would buy that? I mean how big the target market would be for these parts. Do they make economical sense? Would you buy it yourself?

 

[igor_kavinski]

Would you buy it yourself?

I would buy the older gen used or on sale once it gets refreshed

Lots and lots of people out there who are much more able financially to pocket the early adopter tax.

I do have a 128 thread Zen 2 ES on ASROCK mobo.

 

[SteinFG]

I would buy the older gen used or on sale once it gets refreshed

Well here's the answer why they are not making it.

Just buy older epyc or older threadriper then.

 

[igor_kavinski]

Well here's the answer why they are not making it.

Everyone isn't like me

There are probably even workstation users who are being forced to buy older gen TRs right now and that money isn't going into AMD's pockets. If a cheaper TR platform were available, AMD would see more fresh sales that would have a more tangible effect on their bottom line.

 

[MS_AT]

Everyone isn't like me

There are probably even workstation users who are being forced to buy older gen TRs right now and that money isn't going into AMD's pockets. If a cheaper TR platform were available, AMD would see more fresh sales that would have a more tangible effect on their bottom line.

But are there Zen 5 CCDs kept in storage they are unable to sell? Or is everything already selling as fast as they can make it?

 

[Thibsie]

In most EU countries, older gens Epyc or so expensive it isn't worth it at all.

I still wanna grab an old Epyc server for my classroom but just can't afford it and public schools can't either.

 

[igor_kavinski]

I still wanna grab an old Epyc server for my classroom but just can't afford it and public schools can't either.

I got mine from Markfw for a super reasonable price. Contact him

 

[igor_kavinski]

But are there Zen 5 CCDs kept in storage they are unable to sell? Or is everything already selling as fast as they can make it?

Good question but only AMD knows that.

 

[igor_kavinski]

Raise your hands if you felt frustrated as heck when the P4 with DDR instead of RDRAM came out. It wasn't as good due to not getting the bandwidth it needed and for which it was engineered specifically for.

This is the same situation with Zen. It's a server first design, with the inherent expectation right from the design phase being that gobs of multi-channel memory bandwidth will be available. And then they go ahead and chop up it into pieces for the plebs. Oh, a pleb doesn't care about a gimped part. Yeah maybe but a well informed pleb certainly does!

This is why minimum quad channel RAM support is needed for 16 cores and higher. Intel being in the situation that they are, I wouldn't be surprised if they fire the first shot and market a cheap TR killer, completely obliterating the need to get a 16 core or 24 core TR.

 

[DrMrLordX]

Raise your hands if you felt frustrated as heck when the P4 with DDR instead of RDRAM came out. It wasn't as good due to not getting the bandwidth it needed and for which it was engineered specifically for.

DDR was fine once the dual-channel DDR boards were available. It was funny watching SiS come out with the first viable dual-channel DDR chipset. That is so far off-topic though . . .

anyway it's gonna be hella expensive for Intel to roll out a consumer quad channel platform, especially since they can't use their mesh Xeons as a basis. Well okay they could, but if they do then they'll Xeon-Ws based on Granite Rapids which will not necessarily make for a great experience.

 

[igor_kavinski]

anyway it's gonna be hella expensive for Intel to roll out a consumer quad channel platform, especially since they can't use their mesh Xeons as a basis.

Shouldn't they just need to substitute the I/O die or refine the one in Arrow Lake?

I know that there might be a lack of pins in the socket for the purpose but maybe they can do something clever with what they have available?

 

[Thibsie]

Raise your hands if you felt frustrated as heck when the P4 with DDR instead of RDRAM came out. It wasn't as good due to not getting the bandwidth it needed and for which it was engineered specifically for.

You mean, bottlenecked SDR they were trying to sell just to upsell to patent troll Rambus ?
LOL, no. DDR dual channel were just fine.
SDR P4 was LOL and RDR P4 was just trying to get Intel of the proprietary RAM (Rambus).

No thanks.

 

[igor_kavinski]

LOL, no. DDR dual channel were just fine.

https://www.aceshardware.com/pdfs/AcesHardware-45000300.pdf

Maybe you missed these kinds of benchmarks where the boring Willamette beat the formidable Northwood thanks to RDRAM?

 

[Thunder 57]

https://www.aceshardware.com/pdfs/AcesHardware-45000300.pdf

View attachment 112326

Maybe you missed these kinds of benchmarks where the boring Willamette beat the formidable Northwood thanks to RDRAM?

I never thought I'd see the day where one was promoting RDRAM. It was super expensive, hot, and had terrible latency. There was only a small period of time where it had better performance before dual channel DDR made it irrelevant.

Also that article is comparing Via and Sis DDR chipsets to an Intel RDRAM one. Third party chipsets for the P4 were garbage.

 

[igor_kavinski]

I never thought I'd see the day where one was promoting RDRAM.

Not promoting that particular RAM type but trying to drive home the point that right now, we are in a similar situation with Zen 5. It needs wayyy more bandwidth to perform than it has available. Either AMD needs to fix that by Zen 6 with more RAM channels in consumer mobos or we'll have the same lukewarm reception of Zen 6 (only a boring 10% perf uplift across the board because the cores are wasting their time waiting for data to be available so they can start munching and crunching).

OK fine, maybe quad channel is asking for too much. AMD could give us just one more. Three channels of roughly DDR5-10000 should be plenty!

 

[Makaveli]

This is why minimum quad channel RAM support is needed for 16 cores and higher. Intel being in the situation that they are, I wouldn't be surprised if they fire the first shot and market a cheap TR killer, completely obliterating the need to get a 16 core or 24 core TR.

I doubt we will ever see quad channel on consumer. When AMD revamps IOD for Zen 6 they will probably focus on getting the 1:1 speed higher. So if you can run DDR5 8000-10000 at those speeds will help alot with bandwidth.

 

[MS_AT]

Not promoting that particular RAM type but trying to drive home the point that right now, we are in a similar situation with Zen 5. It needs wayyy more bandwidth to perform than it has available.

Could you name workloads that you find bandwidth starved? How many of them you would consider typical client workloads?