The Ryzen "ThreadRipper"... 16 cores of awesome

[crashtech]

I think leaving ECC enabled but not officially supported is a great balance between fusing it off (like Intel) or making consumer CPUs too close to the equal of their server counterparts, which would be bad for business.

 

[Topweasel]

I think leaving ECC enabled but not officially supported is a great balance between fusing it off (like Intel) or making consumer CPUs too close to the equal of their server counterparts, which would be bad for business.

Again I am going to point out lack of overlap. There isn't an 8c server variation. There isn't going to be a 16c AM4 option. Even the difference in the 12/16c option and the 24/32c options are going to be platform independant from each other. The only overlap is going to be between the R3 and the Ryzen APU offerings. So really it should only come down to core count and IO throughput.

 

[crashtech]

Again I am going to point out lack of overlap. There isn't an 8c server variation. There isn't going to be a 16c AM4 option. Even the difference in the 12/16c option and the 24/32c options are going to be platform independant from each other. The only overlap is going to be between the R3 and the Ryzen APU offerings. So really it should only come down to core count and IO throughput.

I don't know why you think overlap matters that much. If consumer Ryzen or Threadripper had official ECC support, it would erode sales of upcoming Epyc regardless of core count, imo. I don't think that should be controversial.

 

[Topweasel]

I don't know why you think overlap matters that much. If consumer Ryzen or Threadripper had official ECC support, it would erode sales of upcoming Epyc regardless of core count, imo. I don't think that should be controversial.

I don't agree. ECC is important on servers and basically a requirement. But honestly I really doubt that buyers of AMD in corporate would be allergic to their high core counts that the would try to run AM4 or TR4 desktops in servers and again if TR4 pricing is what is Rumored then AMD is literally selling at a linear core and clock speed pricing. Which screams buy my Zeppelins not an attempt to segment the market by artificial feature restrictions.

That said AMD didn't work with customers on validating ECC in Ryzen because it isn't meant for that range. But we aren't talking about AMD fusing off capability just because it might be "missused" by some enthusiasts. Got to get out of the mindset of Intel's product placement. It's always been anti consumer.

 

[crashtech]

@Topweasel , I don't think I exactly understand what you are trying to say, and it's possible you might be misinterpreting me, so I'm just going to leave it at I am glad AMD has ECC support in their CPUs, and I think it is okay that it's not official in a consumer-level product.

 

[ericlp]

Intel unveils monster 18-core Core i9: 'First teraflop-speed' consumer PC chip ....

yes... I know... This is the AMD thread. But, my question is, how fast is the Thread Ripper? Wouldn't you think with more Lanes "faster communication" it could beat Intel?

Also, now that most people think Intel never had a plan for it's i9, just rushed out the numbers to rain on AMD's parade, with no clue how they going to develop it's chip by next year. If AMD pulls off a thread ripper by the end of this summer, They will have tons of business and risks of being sold out, all the cash will just help them along to a 25-30 core CPU by the time or maybe even sooner than Intel can release it's crap CPU. I think Intel is going down in a handbasket, Even apple is considering dropping them. Yes, it has gotten that bad!

 

[Topweasel]

@Topweasel , I don't think I exactly understand what you are trying to say, and it's possible you might be misinterpreting me, so I'm just going to leave it at I am glad AMD has ECC support in their CPUs, and I think it is okay that it's not official in a consumer-level product.

I understand you. I just don't think A.) Validated ECC on Ryzen 7 would erode server sales. B.) That AMD frankly cares that much. EPYC is a very specific configuration and is priced if TR is any example right inline with the Ryzen 7 x 4. If that holds up Why would AMD care if someone bought a Ryzen 7 for $300-$500 or an Epyc at $1200-$2000? same per die profit.

 

[crashtech]

I understand you. I just don't think A.) Validated ECC on Ryzen 7 would erode server sales. B.) That AMD frankly cares that much. EPYC is a very specific configuration and is priced if TR is any example right inline with the Ryzen 7 x 4. If that holds up Why would AMD care if someone bought a Ryzen 7 for $300-$500 or an Epyc at $1200-$2000? same per die profit.

Okay, well, do we know Epyc's price structure yet? I would assume Epyc would have all that extra validation cost built into the price, making it somewhat more expensive.

 

[Atari2600]

@Topweasel

Snowy Owl is supposedly an exact server counterpart to Threadripper - 2 Zeppelin dies.

They likely *will* release an 8 core version too.


The below is well out of date - but outlines their thinking. Ryzen with ECC would cannibalise EYPC sales.

 

[Topweasel]

Okay, well, do we know Epyc's price structure yet? I would assume Epyc would have all that extra validation cost built into the price, making it somewhat more expensive.

I am basing it on TR's pricing because A.)that platform outside of 1S is very similar and I am pretty sure (though nowhere near positive) that AMD will be working on validating the platform OEM workstation options. B.) AMD still has one job to do this generation and maybe next and that is to disrupt Intel's price/perf and density so much that OEM's have to offer their stuff. AMD easily could have priced TR higher and still have been a deal, but capping TR16c at $1k sends a message. I think EPYC will as well.

 

[itsmydamnation]

You people realise AMD 's positions with Bulldozer was exactly the same position as Zen right.........

 

[crashtech]

I am basing it on TR's pricing because A.)that platform outside of 1S is very similar and I am pretty sure (though nowhere near positive) that AMD will be working on validating the platform OEM workstation options. B.) AMD still has one job to do this generation and maybe next and that is to disrupt Intel's price/perf and density so much that OEM's have to offer their stuff. AMD easily could have priced TR higher and still have been a deal, but capping TR16c at $1k sends a message. I think EPYC will as well.

Extrapolating upon a rumor is bound to leave you disappointed. I personally would expect, and think it very fair, that Epyc would be somewhat more expensive due to all the extra work that has to go into parts that serve mission-critical needs.

 

[Topweasel]

@Topweasel

Snowy Owl is supposedly an exact server counterpart to Threadripper - 2 Zeppelin dies.

They likely *will* release an 8 core version too.


The below is well out of date - but outlines their thinking. Ryzen with ECC would cannibalise EYPC sales.

Well out of date is right. Snowy Owl has been about 4 different things just since Zen launched. It was a 16c Quad APU, a 8c quad APUs (these two for compute), TR, and possible ThreadRipper EPYC. We will see. As it stands right now the only clear server chip is Naples. We have a month or two to guess pricing. But my guess is maybe a little bit extra margin it won't be too different from ThreadRipper and Ryzen pricing in terms of speed and cores. It's possible that in the future if they do have 16 and 8c EPYC's (kind of useless) they would only be a small amount more expensive than their Ryzen counterparts and it will be one of those things that sure they are losing a little money, anyone that matters will get the Validated server stuff and Enthusiasts will be happy AMD didn't make them jump on a more expensive platform (which is a side point unlike Intel, AMD doesn't really profit from board sales).

 

[Topweasel]

Extrapolating upon a rumor is bound to leave you disappointed. I personally would expect, and think it very fair, that Epyc would be somewhat more expensive due to all the extra work that has to go into parts that serve mission-critical needs.

Which AMD isn't responsible for. Literally the only thing they have a financial stake in now is the actual CPU and like TR it's just 4 dies on single package with IF interconnects. Whatever the extra cost is on TR it's the same as EPYC. AMD will have to work with more partners, send out more development kits (Thousands of CPU's for testing) but heck EPYC is offering so many lanes, and the USB and SATA (and even NVME) are so not important in this space that it is going to be chipsetless. A complete SoC. So everything is going to be on the hardware partners.

 

[jpiniero]

Extrapolating upon a rumor is bound to leave you disappointed. I personally would expect, and think it very fair, that Epyc would be somewhat more expensive due to all the extra work that has to go into parts that serve mission-critical needs.

At least for Broadwell-E, Most of the 2S Xeons were in the 1500-2500 range, and there were models well above and well below that. So 2500ish is probably the max they could sell the 32 core model for. 2k would be aggressive yes.

 

[crashtech]

@Topweasel , are you saying that AMD has no role in, nor incurs any costs validating server platforms? That sounds very wrong to me.

 

[Topweasel]

@Topweasel , are you saying that AMD has no role in, nor incurs any costs validating server platforms? That sounds very wrong to me.

I didn't say that. They will still have to work with their customers to validate the hardware. But that is a flat cost cost really, they don't have any real development costs for "parts that serve a mission critical need". OEM's are basically free to integrate the hardware that they are comfortable onto the platform rather than fit their requirements around AMD's implementation (like they do with Intel). This means quicker validation. So basically they just have AMD's design consultants, on site guys to help develop a validation suite, and whatever the supply of CPU's they need for the testing (which might be the largest portion of costs needed to apply to actual EPYC shipment costs). If they continue to use the platform for several generations, that will only quicken the ability for the OEM's to test what they need.

The end point is that there is a cost. It won't be as high as it used to be. Even if it was it's not as high enough to cause AMD to price it out of its "Disruptive" price range. Intel's pricing is not a sign of actual CPU costs and a EPYC at $1500 is still $1300 in profit at 10k in sales per OEM you have 13 million in profit to cover RD, Validation, consulting per OEM client. That's on a single bare minimum order. If it's popular and they order 100,000 that is 130 million. If AMD can't make themselves profitable off a 500% manufacturing markup then they shouldn't be in business.

Edit: I will note that EPYC has the most room to stretch it's legs profit wise. So it would be a viable option for AMD to price it higher just because Intel's offering in the performance range is so far out there if there is a company in need of more profits it is AMD. But again I don't think they will delve to far from their Zepplin die pricing they seem to have established because it puts a tremendous amount of pressure on both Intel and also onto the OEM's to offer their solutions.

 

[crashtech]

@Topweasel , staying within what is considered a "disruptive" price range is not the same as selling server parts at prices consistent with consumer parts (which you seem to be wishing/advocating for).

 

[stockolicious]

https://www.youtube.com/watch?v=L3l9vZD7h_8&feature=push-u&attr_tag=mcPAmNS4noWp8b7N-6
I like this guys accent - but i like his vids too he comes off fair - threadripper info here

 

[Topweasel]

@Topweasel , staying within what is considered a "disruptive" price range is not the same as selling server parts at prices consistent with consumer parts (which you seem to be wishing/advocating for).

I am not advocating. I am just saying that AMD had all the potential in the world to scale up the ladder against Intel on ThreadRipper considering the HEDT setting. Not only that but unlike Intel they don't seem that intent on separating the Workstation offerings from the "Gaming" offering. But they didn't scale up the market at all and it seems more and more likely the insanely high clocked 16c32t will top off at $1k. I am not saying AMD will stick strickly to that pricing for EPYC, and I would understand if they pumped up the pricing. I am not sure AMD intends to.

 

[ZGR]

You people realise AMD 's positions with Bulldozer was exactly the same position as Zen right.........

How so?

 

[Topweasel]

@Topweasel , staying within what is considered a "disruptive" price range is not the same as selling server parts at prices consistent with consumer parts (which you seem to be wishing/advocating for).

Another point to add to this. Do you remember Opteron pricing back in the day? I could order Opteron 2S 12 core servers for $1500 less than a similarly built 8 Core Intel solutions. AMD doesn't do the market segmentation like Intel did. It's not like they never sold their CPU's as expensive as they could and still sell. But Opteron's that were simple rebadges of Athlon's or vice versa, where similarly priced within 5%. In fact there was a market for Opteron's that were undervolted for servers that people bought up by people because they were lower priced because they were clocked lower, but were great overclockers. AMD didn't do much to stop it because as long as it increased their ASP they were happy. AMD wouldn't be selling these chips at Ryzen costs, its still a $1500-$2000 CPU maybe they push 2500, but then you start to encroach on the pricing of Intel's HCC CPU's. Even if AMD is still beating Intel on core count and IPC by that point, it swings the favor back in Intel's favor. AMD has to overcome the "no one ever got fired for buying Intel" mindset and the way to do it is to offer the CPU at "an offer they can't refuse".

That said I can be wrong and AMD will price their 32C chips at $5k. I would be surprised personally if they even touched 3k.

 

[formulav8]

They likely *will* release an 8 core version too.

I personally expect 8/12/16 core TR's. I saw some chart with in-between core counts but I don't see it myself based on the CCX configuration. I have no specific paper or source though.

 

[Ajay]

https://www.youtube.com/watch?v=L3l9vZD7h_8&feature=push-u&attr_tag=mcPAmNS4noWp8b7N-6
I like this guys accent - but i like his vids too he comes off fair - threadripper info here

Is that a heavy Scottish accent? It's pretty brutal to listen too. I gave up.

 

[kalmquist]

How did you come to that conclusion? Is there any article that proves it? That linked article doesn't. No process termination was seen. Unless you have an article that demonstrates that.

I wrote that, “ECC RAM appears to work just fine under Linux.” That doesn't mean that there are no bugs in the implementation; it just means that the article doesn't identify any. I will readily admit that there could be problems that the writer of the article failed to uncover.

The writer misadjusted the memory timings and then ran a memory stress program, which runs multiple child processes. When the uncorrectable error occurred, the kernel should have sent a a signal to the process that was reading memory when the error occurred. The article does not tell us how the memory stress program reacts when it receives a SIGBUS signal. (The documentation for the program doesn't say, so it would be necessary to test the code to find out.) And even if we knew that answer to that question it wouldn't help us because the article doesn't contain any output from the memory stress program.