• Now's your chance win big! Join our community and get entered to win a RTX 2060 GPU, plus more! Join here.

    Pi Cast Episode 3 streams live on Tuesday, August 4th at 2:30 pm ET (7:30 PM BST). Watch live right here!

    Catch Scharon on the Tom's Hardware Show live on Thursday, August 6th at 2:00 pm ET (7:00 PM BST). Click here!

News AMD Delays Ryzen 9 3950X Launch, Announces Third-Gen Threadripper

Sep 20, 2019
4
2
15
0
I think what you are saying about AMD holding back some of the chips for their epyc lineup is likely true. The simple fact is that they get higher margins on epyc than they get with consumer. I mean if you can sell an 8c16t cpu for $575 as an EPYC 7262 it just doesn't make much sense to sell it for $330 as a 3700x. The margins are just much better in enterprise.
 

jimmysmitty

Champion
Moderator
AMD is definitely trying to take a piece out of the Xeon pie, but in doing so are reducing consumer availability, which sucks
The only problem I see is Intel is attacking HPC from many fronts and is offering many features that AMD might not have, like their FPGSa designed for Data Centers or Optane DIMMs. Its going to take a lot of work on AMDs end to really take a chunk out of Xeons and not just offering more cores per dollar.
 

bigdragon

Distinguished
Oct 19, 2011
551
28
19,010
0
I regret not buying a 3900X a month ago when I had the chance. The local Micro Center had no problem keeping those in stock for July and most of August. 3900X has been out of stock all September and Micro Center has delisted the product page a few times. Hopefully AMD can fix this availability problem sooner rather than later.

I can't blame AMD for prioritizing Epyc if that's what they're doing. I know even my work, a long time Intel-only shop, has now started ordering Epyc servers due to Intel's nonstop parade of security issues. There will be growing pains as AMD expands their market share.
 

Gillerer

Honorable
Sep 23, 2013
294
26
10,890
30
Still going to say TR will start at higher than 16c/32t. Doubt they want to cannibalize their higher margin market with their mainstream products.
In addition to the memory channels (4 or 8), registered & ECC memory support (on the workstation parts) and PCIe lanes, there is another way they can differentiate between mainstream Ryzen and Threadripper: Number of CPU dies. A 3rd gen. Threadripper could potentially have anything between 1 and 8 of them.

By using partially disabled dies, but more of them, AMD could increase both the L3 cache amounts as well as memory bandwidth(*. Workstation users could use more cache and more bandwidth to feed the 8 channel memory.

24 core could be made from 3 dies of 8 cores, 4 x 6 or 6 x 4, which would yield max. 96/128/192MB of L3 cache, and 1.5x/2x/3x the memory bandwidth(* (compared to 64MB and 1x of a 2 CCD design such as 3900X or 3950X).

If instead of a 16-core (2x8 or 4x4), they decided to go for a 18-core, they could build it from 3 dies of 6 cores each. (As an added bonus, you'd be getting 2 "extra" cores compared to the mainstream CPU, and 18 is halfway between 12 and 24, if a 12-core Threadripper were to be the "starting at" model for people who just need I/O.)


*) between the CCXes and the I/O die
 
Last edited:
Reactions: bit_user

InvalidError

Titan
Moderator
Still going to say TR will start at higher than 16c/32t. Doubt they want to cannibalize their higher margin market with their mainstream products.
The only sales AM4 16-cores would cannibalize is sales to people who need absolutely nothing else than the cores. If you need more RAM capacity, more RAM bandwidth, more IO than 20 host-based PCIe lanes, etc., then AM4 is no-go no matter how many cores AMD puts on it.
 

bit_user

Splendid
Ambassador
By using partially disabled dies, but more of them, AMD could increase both the L3 cache amounts as well as memory bandwidth(*. Workstation users could use more cache and more bandwidth to feed the 8 channel memory.
Good points, but note that TR motherboards only support 4 memory channels. I'm sure they literally don't wire up the other 4 in the socket.

What I think would be interesting is if Epyc CPUs supported disabling half of their PCIe lanes and memory channels to slot into TR4 boards where people wanted > 32 cores. That would avoid AMD needing to create more TR SKUs, just to serve that small market segment.
 

InvalidError

Titan
Moderator
Good points, but note that TR motherboards only support 4 memory channels. I'm sure they literally don't wire up the other 4 in the socket.
Rumors say ThreadRipper 3 will be available in EPYC-like (octo-channel RAM, 96-128x PCIe4 depending on how many get re used for HSIO) variants, so we'll see TR4 motherboards for each variant, probably under the "WRX" brand to reduce the likelihood of confusion.
 

bit_user

Splendid
Ambassador
Rumors say ThreadRipper 3 will be available in EPYC-like (octo-channel RAM, 96-128x PCIe4 depending on how many get re used for HSIO) variants, so we'll see TR4 motherboards for each variant, probably under the "WRX" brand to reduce the likelihood of confusion.
That would be very interesting. Could be seen as (though I'm sure it's not) a F-you to Apple, for going with Intel in their newest Power Mac.

I just think it'd be nice to see them take advantage of the abstraction provided by the I/O die, in order to unify their workstation & server chip markets. Create more options for everybody, without the burden of so many SKUs.
 

InvalidError

Titan
Moderator
I just think it'd be nice to see them take advantage of the abstraction provided by the I/O die, in order to unify their workstation & server chip markets. Create more options for everybody, without the burden of so many SKUs.
Then you'd just end up with marketplace confusion from having no clear boundaries on what chips support which features on what motherboard similar to how many people got confused when Intel put mainstream desktop chips on its HEDT platform, leading to most motherboard features being unusable due to acute PCIe lane deficit and mainstream CPUs only having dual-channel memory.

I'm perfectly fine with 8-channel PCIe x128 products having separate branding from 4-channels PCIe x64 stuff despite sharing the same physical socket and possibly being interchangeable, albeit at the expense of functionality loss.
 

bit_user

Splendid
Ambassador
Then you'd just end up with marketplace confusion from having no clear boundaries on what chips support which features on what motherboard similar to how many people got confused when Intel put mainstream desktop chips on its HEDT platform, leading to most motherboard features being unusable due to acute PCIe lane deficit and mainstream CPUs only having dual-channel memory.
It's possible, but I think that's a rather extreme example.

I prefer to think of it more like how Intel long shared the same socket between their workstations & servers. You use the same CPUs in either, except that some were limited to single-socket. Starting with Haswell-E, I think some lower-end E5 Xeons were also crippled by having only 28 PCIe lanes.

Then, Intel went and introduced a 3rd socket, splitting their servers off from the workstation line. Recently, they seem to be backtracking (by introducing a bunch of Xeon-W branded CPUs for LGA 3647), and I'll be interested in seeing whether they introduce any more Xeon-W CPUs for the LGA 2066 platform.

Truth be told, I was waiting to buy into the LGA 2066 platform, but disappointed at some of the performance regressions of the first-gen Xeon W's. I figured I'd sit back for a couple generations, to see if they could tune the mesh interconnect for better efficiency and lower latency. Only now, I'm starting to think pretty hard about a ThreadRipper.
 

Gillerer

Honorable
Sep 23, 2013
294
26
10,890
30
Good points, but note that TR motherboards only support 4 memory channels. I'm sure they literally don't wire up the other 4 in the socket.
Sorry, I was jumping a gun a bit and not taking the end of August rumors about the upcoming Threadripper chipsets with a grain of salt. The again, my post was pure speculation anyways.

Supposedly there'll be three chipsets: TRX40 with 4 channels as usual, and TRX80 and WRX80 with 8 channels, with the WRX being for workstations with ECC & RDIMM, but no overclocking support.

There's no doubt that 8 channels would require a new motherboard with one of the new "80" chipsets. What is less clear whether (and for which lines - HEDT or WS; 4 or 8 channel CPU SKUs) there would be backward compatibility if you were willing to settle for 4 channels, no official ECC support and RDIMMs having no chance of working.

I doubt the chipsets would be different silicon or have differences in functionality (of the chipset itself), seeing as how Threadripper is a relatively low volume product. Instead the "chipset" designations would be useful in motherboard names and marketing material to signal which boards support which functionality.

(To add to my speculation: With the "40" and "80" numbering AMD would leave room for new chipsets or versions by just incrementing the ones digit.)

*

With the way AMD seems to have used up most of the high 3000 series numbering with mainstream Ryzen, I was thinking maybe they'll go for a different numbering for 3rd gen. Threadripper (also to differentiate it more). Maybe:
  • (7002 series: 2nd Gen EPYC)
  • 6003 series: 3rd Gen Threadripper Workstation (support 8-ch. registered & ECC memory)
  • 5003 series: 3rd Gen Threadripper HEDT (4-ch. memory, supports overclocking)
It would make a lot of sense at least for the WS parts.

Then, to make the SKU numbering really easy to decode, just put the number of cores in the middle two digits: 5243 would be a 24-core 3rd gen HEDT Threadripper. Will work as long as they have less than 100 cores (which I doubt will happen on desktop anytime soon, since frequencies would suffer too much), and until they make it to the 10th generation.

Two variants with the same number of cores (maybe with different cache amounts) can have a +1 added to the "number of cores" of the better one - it's unlikely they'll ever release TR CPUs with an odd number of cores.

Alternatively, if they didn't feel like segmenting the workstation features in a line of their own, they might make the CPU split according to the number of memory channels supported: 6003 for 8-channel, 5003 for 4-channel. If you put either in a workstation board, you'd get ECC & registered memory. If you put either in an HEDT board, you could overclock.

(I suppose they could also go for 63xx and 53xx series, but I like the symmetry with the EPYC numbering.)
 
Last edited:
Reactions: bit_user

jimmysmitty

Champion
Moderator
Sorry, I was jumping a gun a bit and not taking the end of August rumors about the upcoming Threadripper chipsets with a grain of salt. The again, my post was pure speculation anyways.

Supposedly there'll be three chipsets: TRX40 with 4 channels as usual, and TRX80 and WRX80 with 8 channels, with the WRX being for workstations with ECC & RDIMM, but no overclocking support.

There's no doubt that 8 channels would require a new motherboard with one of the new "80" chipsets. What is less clear whether (and for which lines - HEDT or WS; 4 or 8 channel CPU SKUs) there would be backward compatibility if you were willing to settle for 4 channels, no official ECC support and RDIMMs having no chance of working.

I doubt the chipsets would be different silicon or have differences in functionality (of the chipset itself), seeing as how Threadripper is a relatively low volume product. Instead the "chipset" designations would be useful in motherboard names and marketing material to signal which boards support which functionality.

(To add to my speculation: With the "40" and "80" numbering AMD would leave room for new chipsets or versions by just incrementing the ones digit.)

*

With the way AMD seems to have used up most of the high 3000 series numbering with mainstream Ryzen, I was thinking maybe they'll go for a different numbering for 3rd gen. Threadripper (also to differentiate it more). Maybe:
  • (7002 series: 2nd Gen EPYC)
  • 6003 series: 3rd Gen Threadripper Workstation (support 8-ch. registered & ECC memory)
  • 5003 series: 3rd Gen Threadripper HEDT (4-ch. memory, supports overclocking)
It would make a lot of sense at least for the WS parts.

Then, to make the SKU numbering really easy to decode, just put the number of cores in the middle two digits: 5243 would be a 24-core 3rd gen HEDT Threadripper. Will work as long as they have less than 100 cores (which I doubt will happen on desktop anytime soon, since frequencies would suffer too much), and until they make it to the 10th generation.

Two variants with the same number of cores (maybe with different cache amounts) can have a +1 added to the "number of cores" of the better one - it's unlikely they'll ever release TR CPUs with an odd number of cores.

Alternatively, if they didn't feel like segmenting the workstation features in a line of their own, they might make the CPU split according to the number of memory channels supported: 6003 for 8-channel, 5003 for 4-channel. If you put either in a workstation board, you'd get ECC & registered memory. If you put either in a HEDT board, you could overclock.

(I suppose they could also go for 63xx and 53xx series, but I like the symmetry with the EPYC numbering.)
I don't see why AMD couln't make it so an 8 channel board could run a 4 channel chip. Intel did the very thing with X299 when they has the i7 7740x, it was essentially a 7700K on LGA2066 and would only run two channels.

The other way would be harder though unless they can figure a way out to make the CPU shut off memory channels. However with the chiplet design since I/O, if similar to Zen 2, like the MC will be independent from the CPUs themselves they might be able to.

Or they may just say screw it and TR3 will require a new board all together. Or the TRX40 is replacement for current TR, TRX80 is a new one designed, as you said, specifically for more workstation/low lever server tasks and WRX80 is pure workstation design.

Guess we have to wait and see.
 

ASK THE COMMUNITY