Discussion AMD Threadripper 3000 series announcement and reviews

[Kocicak]

24 cores for 1400 and 32 cores for 2000 USD?

That is expensive. It does not put much pressure on Intel. It also does not motivate people too much to change platforms from AM4 to sTRX4. I must say I am surprised.

I bet that some people at Intel are pretty relieved now.

 

[beginner99]

I suspect with the whole epyc design AMD kind of created the problem of not being able to offer a low end HEDT version. I'm pretty sure they need at least 4 dies probably for technical reasons but also due to heatsink pressure distribution. Hence anything below the 24-core doesn't make much economical sense.

But IMHO it's not that bad, since x570 is pcie4 it does offer enough lanes even for some exotic use cases.

 

[Topweasel]

Well, I'd say quadroupling the chipset I/O bandwidth and doubling the total I/O bandwidth is more than worth it.

Yeah and its a bit more than than doubling. The flexibility of using the IO die versus the lanes of each cpu die is paying off. I also wouldn't be surprised if the Zen 4 version isn't using DDR5 and therefore they made the choice now while forced by configuration limitations to move TR upmarket to a workstation only type of lineup to upgrade the platform configuration in order to feel like it can hold out till after TR5 to replace the platform. When you think of where TR came from, what they have at their disposal now and going forward, the socket change seems almost worth it.

It's always disappointing when they force a platform change but we really should be looking at why a change was made, what you get from it. When we look at the platform changes between this and the X99-299 or Z70-Z390 vs the X399-TRX40 or AM4. You start to see a trend. There is pretty much nothing new from the X99 up and Z77 an up. It has been one of the reasons their hasn't been any platform problems (beyond the launch confusion of the X299 due to PCIe differences between CPU's), but those changes haven't been to support new features, or that the CPU's changed in anyway that made supporting the new products any harder. AM4 on the other hand has had a few handicaps that restricting new CPU's specially the new Ryzen 3k series (specifically the 3900x and 3950x) could have made better use of if the platform had more breathing room and wasn't limited by design choices made during Zen 1's development. But that is what a consumer product is supposed to be, a compromised design to satisfy 90% of the market. HDET and specifically TR is supposed to be a no compromise product, this is the first time it really isn't and in that sense, I think it was smart for AMD to re-approach the platform and change some of the weaknesses they had from their original design restrictions (64 lanes only due to die pairing) and the original design decision made to keep development cost down (370X chipset re-purposed for X399). This is really the first time in either party where you get a well defined change in the platform when making a platform change like this.

 

[Timorous]

From AMD's perspective, that's not a great end result when they clearly could price a 16C TR3 @ 900 USD.

They could but the 16C Ryzen would likely have a higher margin so why compete with yourself?

I think this might be AMD saying to Intel we can give you a niche you can actually somewhat compete in but it is going to mean sacrificing in another segment because you are supply constrained. Making too many of these parts may not be worth it to Intel in the long run so even if they have a part that can compete in theory it may not actually be that easy to get hold of one and that will mean retailers increase the price making the price difference for stepping up to TR3 lower or the savings from going with the 16c Ryzen higher.

If the 18c does have a niche (and we need to wait for benchmarks to really see) it could be a pyrrhic victory for Intel to pursue it.

 

[IndyColtsFan]

I don't like it that AMD is pushing for the older Threadrippers to those who don't want to pay for the 3960X. Those chips have their own issues with NUMA and lack the FP capabilities of Zen 2 which makes them a lot less attractive. I bet that even a 3950X would easily beat a 2970X in many workloads.

Yeah, I do wish AMD would’ve released a 16 core TR3 this round. I think the $750 to $1300 pricing gap is huge and I hoped they would slot a new TR in there instead of the 29xx series. On the bright side, the 29xx series will likely see some huge discounts so I would imagine the 2970 may even fall into low $800 range soon. It’s already $916 on Amazon.

 

[CHADBOGA]

As far as the comment on AMD "fans" and saying "but its OK all of a sudden because its' AMD" is just trolling.

But is what he is saying, true?

Here AMD fan favourite Coreteks, is very unimpressed with AMD's pricing choices and makes very obvious Intel comparisons.

 

[coercitiv]

Here AMD fan favourite Coreteks, is very unimpressed with AMD's pricing choices and makes very obvious Intel comparisons.

His intro reproduced my exact initial reaction to TR announcement info: is complacency setting in already at AMD?

I guess Christmas came early at Intel.

 

[dnavas]

Well, I'd say quadroupling the chipset I/O bandwidth and doubling the total I/O bandwidth is more than worth it.

By Robert's own words, products that will utilize this I/O won't come quickly. Additionally by his own admission, there are basically two sets of customers -- people who 'pay for the best' and people who 'want the best value.' And yes, this bifurcation exists even in the expensive TR market -- he was specifically talking about that market. Anyway, I agree with you, for one of those two customer sets, doubling I/O is worth spending the extra money. For the rest of us, it is not. Robert's no fool, though. Either this new platform is providing *value* in some way not currently visible, or the value market is being given cheaper 2970/2990 parts. As far as we've been told, it's both.
We've been told that the wider chipset I/O was required for reasons that are not yet apparent (but will be real soon, trust him) AND that the 2970 and 2990 will continue to be manufactured. I continue to want to know what the devil they need 8 lanes of pcie4 for in the chipset (or, really, 4 additional lanes). They're on record as saying NO to the dual socket TR -- is this really only for nvme? That's a lot of disk I/O! It's true that 120fps 8k raw is something like 20GBps, and maybe that's the point. But why not put that use on an expansion slot? There's got to be something else here.... They're not planning on putting a rasterizer on the I/O, are they?! A TR with a heterogenous set of compute cores (cpu + gpu) would be sort of slick. Again, though, why not use the expansion slot for that? It's a head scratcher to me, anyway . :shrug:

 

[LightningZ71]

I personally believe that the biggest Gap in D's lineup isn't the lack of a 16 core TR3000. Instead, I believe that the biggest Gap is the lack of a lower end workstation platform to take better advantage of the 3900x/3950x processors. Hear me out:

The 3900/3950 should easily be able to out process any of the 1900/1920/1950/2920/2950x threadrippers in 99% of the tasks that you throw at them that aren't entirely memory bandwidth bound AND that have a working set that can't fit into the local L3 cache. That's a very very tiny selection of tasks out there that aren't entirely focused at the very top of the market. The one main limitation of the AM4 socket that will hold it back as a workstation platform is PCIe I/O.

That X16 PCIe 4.0 link could be repurposed to link to a special chipset that has all the needed I/O AND drives four full length PCIe slots that can do bifurcation like the following: pair 1, 4.0 x16 or 3.0 x16/3.0 x16 or 4.0 x8/4.0 x8. Second pair same as the first. Throw in a pair of 4.0 x4 channels that can flip to nvme slots and the usual I/O, and you have a solid chipset and platform. This also leaves the PCIe 4.0 x4 channel that was meant for the chipset free, and, as it is essentially just another x4 pci channel, it can be used as a second CPU connected nvme slot.

The resulting machine would have 2 x PCIe 4.0 m.2 slots from the CPU, two from the chipset, four high speed full length PCIe slots, and either a bunch of sata channels, or fewer, but 10Gbps Ethernet and or tunderbird plus a few smaller PCI slots. It would still have the two DDR4 channel limit, but, with 64 MB of L3, as compared to the past TRs, that is rarely going to be a big handicap. As for RAM, it's still got 128 GB capacity with 32GB Dimms, AND supports ECC.

That would bridge the products nicely.

 

[beginner99]

I personally believe that the biggest Gap in D's lineup isn't the lack of a 16 core TR3000. Instead, I believe that the biggest Gap is the lack of a lower end workstation platform to take better advantage of the 3900x/3950x processors. Hear me out:

On the other hand if you only need the processing power but not the pcie lanes then 3900x/3950x are already fine for cheap workstation, if you need both well you need the new threadrippers and if you only need the lanes, get and TR2.

EDIT: As I wrote before AMD dug themselves a bit of a hole with the large IO die and probably needing at least 4 chiplets + not worth it to make a 16-core part due to higher demand / too low output at TSMC.
In the end it does give cascadelake-x a small niche.

 

[dnavas]

As I wrote before AMD dug themselves a bit of a hole with the large IO die ....
In the end it does give cascadelake-x a small niche.

Phfft -- are the hardware differences between 2x3900x and 3960x really all that large? -- 16 more pcie 4.0 lanes (2x24 vs 64) for two fewer wraith coolers. Doesn't seem like that's worth a $400 premium. You're probably right on the 16 core, but as the 24 core should have slotted closer to that price point.... :shrug: The platform cost of TR is really on the motherboard side, not the CPU -- the I/O hub on TR seems much more capable.

 

[LightningZ71]

(Board acting wierd on mobile, sorry if this is a duplicate. Mods, please delete if it is)

thinking about things... There's no real hard reason that AMD can't use the AM4 socket with the TR40X chipset. The processor already has no issues with bifurcating the x16 channel, so that could easily drive the TR40 AND an x8 PCIe 4.0 slot while also driving a pair of processor linked M.2 PCI drives.

That's already a solid lite workstation board without reinventing the wheel...

 

[StefanR5R]

I will add to this post as more details and review come out.

servethehome:
"STH asked if this was an EPYC 7002 series derived I/O die but we were told that it is a distinct I/O die but are awaiting a follow-up."

 

[Panino Manino]

I, again, welcome higher prices.
This may allow Intel to maintain most of it's margins, but for Intel this extra cash makes no difference. For AMD any dollar makes a difference.

I know that everyone wants to pay as little as possible, but "can't be helped".

 

[TheGiant]

It feels surreal to see the same arguments being spat here to justify higher prices, only this time, by the AMD fans. Oh how the times have changed! At 280 watts, $1,400 entry fee ($2,000 for mobo and cpu alone), the abrupt death of x399, at best a one time upgrade path for TRX4, and topped by a paper launch, AMD has outdone Intel in raising the stakes in the HEDT segment. This is not what most were expecting, but it's okay all of a sudden because it's AMD.


Don't broad brush the board, straw man, or engage in hypocrisy fallacies.

AT Moderator ElFenix

stop arguing about ...., this is not a place for such things, just enjoy the world created, maybe that is all the cores are used for (or not)

anyway, with such number of cores - HTF can I achieve high OEE on desktop content creation machine
and what about programs with high per core license cost ? in calculation of price/perf CPU price is definitely the top spot of that ratio

 

[Kocicak]

I, again, welcome higher prices.

Good that all consumers are not like you, because otherwise the world would collapse in a inflation spiral.

Anyway, I watched the video with the two bald guys, and it seems that AMD changed the socket so that they can have two socket motherboards. Imagine 128 cores monster that fits in a "normal" PC case! WOW!

 

[StefanR5R]

I watched the video with the two bald guys, and it seems that AMD changed the socket so that they can have two socket motherboards.

Coreteks said: "I asked AMD if we should expect dual-socket motherboards for this new generation of Threadrippers, and the answer was a clear No." (Whatever this generation might mean.)

 

[Topweasel]

It's a stupid question anyways. A two socket 8-16 channel board wouldn't be anything close to ATX. AMD is trying to sell the idea that EPYC is so good you don't even need 2 socket support to be better then everyone else and you can even do 1u super servers by using their 64c CPU's. The last thing they would ever support is 2P on a enthusiast/prosumer/Workstation CPU like TR. If you really need a 2p 128c beast, petition Supermicro to sell a board and case to support it for an EPYC workstation.

 

[IEC]

From the released specs, it appears to me that the chiplets used are not the best 5% like previous Threadrippers.

From the released pricing, it appears that AMD is confident they can sell every chip they can produce for the foreseeable future.

 

[Kocicak]

The reason for higher TDP of the 3rd gen. compared to 2nd gen. may be the massive IO die (produced with an "old" power hungry technology), not necessarily bad quality of chiplets.

 

[Kocicak]

... the answer was a clear No.

What else should they say? "No" can sometimes mean "yes but I cannot tell".

 

[Markfw]

The reason for higher TDP of the 3rd gen. compared to 2nd gen. may be the massive IO die (produced with an "old" power hungry technology), not necessarily bad quality of chiplets.

No. thats only a few watts.

 

[StefanR5R]

Coreteks said: "I asked AMD if we should expect dual-socket motherboards for this new generation of Threadrippers, and the answer was a clear No." (Whatever this generation might mean.)

What else should they say? "No" can sometimes mean "yes but I cannot tell".

There are several and better ways to say "I can't tell" than "a clear (all-uppercase) NO".

The last thing they would ever support is 2P on a enthusiast/prosumer/Workstation CPU like TR.

Edit -- Incidentally, their 3960X/3970X/TRX40 announcement slides show an "8x Chipset Downlink (Reserved)" at the processor and an "8x Chipset Uplink (Reserved)" at the chipset. It will be interesting to see if and how these reserved links will be put to use in later products of this platform.

 

[Kenmitch]

The reason for higher TDP of the 3rd gen. compared to 2nd gen. may be the massive IO die (produced with an "old" power hungry technology), not necessarily bad quality of chiplets.

More like it's....

TR 3960X is 4x 3600(260 TDP) under the hood
TR 3970X is 4x 3700x(260 TDP)under the hood

....but with higher binned chiplets to achieve higher clockspeeds. I just used the 3600 and 3700x chiplets as TDP is close.

That would be my guess.

I'm not in the market for either, but the reviews will be interesting to say the least.

 

[Mopetar]

AMD might not make a 2P workstation, but Intel always could. That would let them be competitive on performance against AMDs higher core count chips without having to push top binned chips beyond the point of efficiency.

If that were successful for Intel then AMD might have to respond in kind. Obviously this is the kind of thing that happens over several years, but if there’s an actual market for 2P workstations we’ll get there eventually as long as the market remains competitive.

 

[jpiniero]

Dell (and I'm sure others) sells 2P workstations, but it uses the SP Xeons as-is. They want an extra 24 grand for the top model which includes dual 8280Ms.