I'm not surprised by this, in fact i expected it to be like this, considering its the same setup as their servers. I'm curious though, does that mean a 16 core chip is 4+4+4+4? Hey look it's Mr. Nonsensical back to tell us all about his mass knowledge of absolutely nothing taking information and turning it around without absolutely any proof all to hound on AMD for no reason. A whole lot of nonsense. It's nice to see you, i really hope you get banned soon.

if they had perfect yields than thread ripper would not exist(at least in it's current configuration) as that 32core could sell on EYPC for much higher profit margins. With perfect yields there would only be 2 dies, 4 perfect ccx = 16 cores. But then I'm not sure if it could still be quad channel memory in that configuration.

so that means threadripper is 2+2+2+2+2+2+2+2 ???? so is not 2x 1800x is 4x1400 ? and if yes what is the 12 core ? 2x1600 or is like 2+2+2+2+1+1+1+1 ? if the second case is true ....is kind of wow that way they can use dual ccx's that can not even make it to be quad cores since they are not going to sell tri-cores and dual cores but it seems they can make em on 12 core that way ? i am a bit confused ! it does not seem like a bad choice to me on thermals standpoint the heat will come from 4 different sources on a huge heat spreader so i expect em to do really well comparing to intels offerings on that aspect ...and if that is the case let me make a prediction of our beloved dual tower like(noctua nh-d15) air coolers having 2 sets of heat pipes leading on each tower !

Yeah, just saw the video by Der8auer. Quite interesting, although not too surprised. Perhaps Threadripper is just EPYC CPUs that didn't quite make it?

Exactly, they probably use same production line for Epyc and Threadripper. That why Threadripper comes after Epyc, Threadripper is salvaged Epyc dies. Just like Ryzen 6 / 4 cores CPUs. That explains low cost.

Holy mother of cores... !! If they are able to build these at a decently low production cost, it means that they can compete with Intel's offerings indefinitely. Sometime 2018: Intel (finally) launches their 18-core uber-expensive 1999$ top of the line HEDT. "We have moar corez !" AMD: "yawn" ... launches 24 cores TR Intel ups the frequency 100Mhz in despair... TDP goes through the startosphere. AMD: "more yawn" ... launches 32 cores TR. BANG. BANG. Ur Dead.

Last year Intel already released a 22 core Xeon Broadwell, if you have x99 mobo you can use it. Haswell-E goes up to 18 cores/32t and Turbo 3.6Ghz, you can find second hand one on ebay for 700-800usd, x99 compatible of course.

so that means threadripper is 2+2+2+2+2+2+2+2 ???? so is not 2x 1800x is 4x1400 ? and if yes what is the 12 core ? 2x1600 or is like 2+2+2+2+1+1+1+1 ? if the second case is true ....is kind of wow that way they can use dual ccx's that can not even make it to be quad cores since they are not going to sell tri-cores and dual cores but it seems they can make em on 12 core that way ? i am a bit confused ! it does not seem like a bad choice to me on thermals standpoint the heat will come from 4 different sources on a huge heat spreader so i expect em to do really well comparing to intels offerings on that aspect ...and if that is the case let me make a prediction of our beloved dual tower like(noctua nh-d15) air coolers having 2 sets of heat pipes leading on each tower !

Watch the video, by the end of the video he says that he contacted AMD and they answered him, its 2CCX out of 4

Watch the video, by the end of the video he says that he contacted AMD and they answered him, its 2CCX out of 4

oh i did not watched the video ...so it seems those are epyc chips that something went wrong with the infinity fabric then ?

if they had perfect yields than thread ripper would not exist(at least in it's current configuration) as that 32core could sell on EYPC for much higher profit margins. With perfect yields there would only be 2 dies, 4 perfect ccx = 16 cores. But then I'm not sure if it could still be quad channel memory in that configuration.

You do know that doesn't fully make sense, right? How long did they take to start releasing 4-core Ryzen processors? If their yields were "so low" why wouldn't they have started releasing them sooner, or for that matter, at the same time? They would have had them... Do you think if their yields were 100%, that would mean there would be no 4-core ryzen processors? That wouldn't make sense, from AMDs perspective. Could they get more money from every chip if every single chip they made they sold at a premium? Sure, if they can sell it all, and if people will buy it, being the fact its so expensive (most people don't want to spend $330 or more on a processor, let alone the people buying a whole PC for a PC sake (not gaming, etc.) that would like it to be $300-400 total system.) No matter how good the yields are, there would still be processors with cut off CCXs and disabled cores, so that way there would be a tier system. As long as they are making a profit, gaining a lot more sales from having cut down products will make them more money, faster, then just having high end products that won't sell as well. Same thing goes for threadripper. They can obviously make a profit on these chips still at the prices they are selling them for. Now, regardless of yields, they could have just decided they were for the server and server only, but why? Why would they cut out the possibility for of more sales from the enthusiast/workstation market, when they don't have to? What would be the purpose of that, other then to shoot themselves in the foot? If i remember the costs correctly, a $1000 processor still has about $700 profit? even if i'm wrong, i'm certain it'd be at least $500, given what we know about the costs for the CCXs and etc. That's $500, as much as the most expensive R7. And just to be clear, i'm not saying that the yields are 100%, i'm sure that'd be impossible, the post you replied to, however, was in regards to a guy deciding this information as being the yields are not good, which is a nonsense thing to conclude, but with his post history, makes sense he'd state it. I mean, if he were actually paying attention, or at least, didn't purposefully leave information out, he'd know that it was announced that zen yields are above 80%, which is quite good.... I mean cmon, athlon 64/barton yields were around 60%, but it seems a little difficult to find out the actual percentages of yields, especially from intel, their graphs are very hush hush, lets show you an angle, but no actual numbers lol

Some new possible insights on the TR design. According to a PCWorld article (http://www.pcworld.com/article/3211409/computers/why-ryzen-threadripper-has-two-mysterious-chips.html), the two extra dies are nothing more than space savers, to help maintain the structural integrity of the chip. In other words, they aren't real dies, and only exist to prevent the IHS from warping/bending under pressure.

Some new possible insights on the TR design. According to a PCWorld article (http://www.pcworld.com/article/3211409/computers/why-ryzen-threadripper-has-two-mysterious-chips.html), the two extra dies are nothing more than space savers, to help maintain the structural integrity of the chip. In other words, they aren't real dies, and only exist to prevent the IHS from warping/bending under pressure.

I wonder if they are literal dummies (meaning they contain no transistors) or if they're just total failed dies. After all, what else are they going to do with the defective dies? Either way, at least this confirms that there are only 2 working dies, which explains why all the resources are cut in half. This also suggests there definitely will not be any TRs with more than 16 cores, at least for this first generation. As far as I'm aware, Epyc is compatible with the same socket so not like it matters anyway. What I don't really understand though is there seems to be a VERY blurred line between enthusiast class workstation hardware and server class hardware. X399 motherboards aren't exactly what you'd want in a server - too desktop-oriented.

That is until ASUS releases a x399 WS.

I can't say this surprises me either. But hey, at least now we know delidding one of these things isn't a viable option. "We are a bit puzzled as to how things are configured for the to be released 16-core threadripper. Common sense dictates that two dies would be disabled, then again there could be four dies active with 2 disabled cores per CCX. But ergo, Threadripper really is 32-core EPYC. And that opens up so many question like is that 1 DRAM channel per die ?" I'm sure it's not as complicated as it might seem. Kind of the point of the CCXs and infinity fabric is to exchange data across the entire die. So it wouldn't surprise me if in some cases an entire CCX has no active cores, but the L3 cache, memory controller, and PCIe lanes are still available for the others to utilize. Maybe I'm way off, but how else would AMD account for defective cores in such a complex CPU? Perhaps this is why AMD never made any single-CCX Ryzens, because otherwise that might have forced it into single-channel memory.

I'm expecting the Vega/Ryzen APU's to use just one CCX (4 cores/8 Threads), where the single memory controller could be a problem. The APU part would have to have its own DDR4 Memory controller, or onboard/package memory. Or the APU's will be on 7nm and a full dual-CCX with Dual channel memory, but can't see that happening by early next year... I thought AMD had stated that their yields were about 70% good ?

I wonder if they are literal dummies (meaning they contain no transistors) or if they're just total failed dies. After all, what else are they going to do with the defective dies? Either way, at least this confirms that there are only 2 working dies, which explains why all the resources are cut in half. This also suggests there definitely will not be any TRs with more than 16 cores, at least for this first generation. As far as I'm aware, Epyc is compatible with the same socket so not like it matters anyway. What I don't really understand though is there seems to be a VERY blurred line between enthusiast class workstation hardware and server class hardware. X399 motherboards aren't exactly what you'd want in a server - too desktop-oriented.

I would have also thought that each dual-CCX complex would be tested before being put on a substrate, so they could be deactivating fully enabled dual-CCX's, just to counter Intels HEDT platform. Maybe it was just to push Intel to do something stoopid, like make 18 core HEDT CPU's on the cheap (compared to the Xeon chips) 🙂

I'm expecting the Vega/Ryzen APU's to use just one CCX (4 cores/8 Threads), where the single memory controller could be a problem. The APU part would have to have its own DDR4 Memory controller, or onboard/package memory.

I can see why you'd think that, but I'm pretty confident AMD wouldn't make such a bad move. Keep in mind the northbridge/chipset has access to the memory controller (and therefore dictates where the bandwidth is going), not the individual processors or CCXs. Though outdated, this diagram is likely still relevant in explaining the order of how things work: https://upload.wikimedia.org/wikipedia/commons/thumb/b/bd/Motherboard_diagram.svg/300px-Motherboard_diagram.svg.png Except in the case of AM4, the memory controller (which is in the northbridge) seems to directly feed into the L3 cache. From what I recall, AMD intends the IGP to feed from the same L3 as the CPUs. In other words, rest assured, APUs will be fully dual-channel where the bandwidth is for both the CPU and GPU.

Or the APU's will be on 7nm and a full dual-CCX with Dual channel memory, but can't see that happening by early next year...

Yeah I wouldn't hold my breath for that. I don't think we're going to see 7nm until "Zen 2", and that probably won't happen until AMD made a CPU for each segment; they still have servers (Epyc), APUs, mobile, and maybe embedded to make.

Also with some of the Phenom II quad cores you could unlock them to hex cores. Also I brought the possibility of unlocking more cores on the Ryzen chips and I was told AMD possibly disabled the cores on the chip physically rather than though some kind of software that was on the chip or how ever they disabled cores back during the Phenom II days where users could unlock the cores themselves via the BIOS.

You do know that doesn't fully make sense, right? How long did they take to start releasing 4-core Ryzen processors? If their yields were "so low" why wouldn't they have started releasing them sooner, or for that matter, at the same time? They would have had them... Do you think if their yields were 100%, that would mean there would be no 4-core ryzen processors? That wouldn't make sense, from AMDs perspective. Could they get more money from every chip if every single chip they made they sold at a premium? Sure, if they can sell it all, and if people will buy it, being the fact its so expensive (most people don't want to spend $330 or more on a processor, let alone the people buying a whole PC for a PC sake (not gaming, etc.) that would like it to be $300-400 total system.) No matter how good the yields are, there would still be processors with cut off CCXs and disabled cores, so that way there would be a tier system. As long as they are making a profit, gaining a lot more sales from having cut down products will make them more money, faster, then just having high end products that won't sell as well. Same thing goes for threadripper. They can obviously make a profit on these chips still at the prices they are selling them for. Now, regardless of yields, they could have just decided they were for the server and server only, but why? Why would they cut out the possibility for of more sales from the enthusiast/workstation market, when they don't have to? What would be the purpose of that, other then to shoot themselves in the foot? If i remember the costs correctly, a $1000 processor still has about $700 profit? even if i'm wrong, i'm certain it'd be at least $500, given what we know about the costs for the CCXs and etc. That's $500, as much as the most expensive R7. And just to be clear, i'm not saying that the yields are 100%, i'm sure that'd be impossible, the post you replied to, however, was in regards to a guy deciding this information as being the yields are not good, which is a nonsense thing to conclude, but with his post history, makes sense he'd state it. I mean, if he were actually paying attention, or at least, didn't purposefully leave information out, he'd know that it was announced that zen yields are above 80%, which is quite good.... I mean cmon, athlon 64/barton yields were around 60%, but it seems a little difficult to find out the actual percentages of yields, especially from intel, their graphs are very hush hush, lets show you an angle, but no actual numbers lol

Sure it makes sense. A CCX is a single 4 core core die. The tier system would still work. Why would the entry Ryzen 3 come with 2CCX(aka 2 dies glued together) if yields are perfect? Perfect yields would mean it would have a single CCX instead they are using the perfect ones in other higher end products. The low end Ryzen products are getting the imperfect chips. Of course the 6 core ryzens would have to have cores disabled so they can fit in the price point consumers want. That's why the high end parts come first. You said 80% are perfect, well that's exactly why Ryzen 3 was last. Takes time to amass the 20% imperfect chips for a reasonable stock to sell.

Also with some of the Phenom II quad cores you could unlock them to hex cores. Also I brought the possibility of unlocking more cores on the Ryzen chips and I was told AMD possibly disabled the cores on the chip physically rather than though some kind of software that was on the chip or how ever they disabled cores back during the Phenom II days where users could unlock the cores themselves via the BIOS.

I read some time ago that AMD use laser to disable cores so activating them is impassible

@Agent-A01 I think Aura89's confusion is where you said "if they had perfect yields than thread ripper would not exist(at least in it's current configuration) as that 32core could sell on EYPC for much higher profit margins." Which isn't necessarily true. For socket TR4, the CCXs appear to have 8 cores each instead of 4. Since apparently Threadrippers have only 2 of the 4 dies active, that means the 16-core variants are also a result of perfect yields; they're merely just lesser models vs Epyc. The thing to consider is getting 4 perfect CCXs for a single CPU is uncommon, and as a result will likely disproportionately cost more. Most people don't have a need for 32 cores, let alone 64 threads. Despite a 16 core costing less, it will be much easier to sell. Therefore, AMD will make more money in the end by focusing on 16 core Threadrippers. Due to the die arrangement, that should mean we'll see 16 core Epycs (where the only difference between them and a TR is the extra PCIe lanes). In a setup like that, you might find 4 cores per CCX.

Why would the entry Ryzen 3 come with 2CCX(aka 2 dies glued together) if yields are perfect?

Because that's how they come? Honestly, they probably have no desire to use just one CCX, as that would get rid of the potential latency issue, and then everyone would (potentially) be buying those, undercutting their high end stuff. If it would actually cause an difference is a different story. But simple fact that we have is, CCXs come in packages of 2, we have no information that states it can come in 1. Again though, i'm not saying the yields are perfect, my original post was about how assuming the yields are "bad" based off the information of this thread doesn't make sense.

For socket TR4, the CCXs appear to have 8 cores each instead of 4.

As far as i can tell, they still have 4 per CCX, there's 2 CCXs per "die" area, or whatever you want to call it. If i'm not mistaken, when i was watching the youtube video, the guy said the reason the top of the heat spreader looks like this: https://www.extremetech.com/wp-content/uploads/2017/07/Der8auer-Indium-640x375.jpg With 8 "sections", is because there are actually 8 CCXs there. Just like the Ryzen processors http://hexus.net/media/uploaded/2017/3/1f34ebdb-fd8d-40de-9085-54cf055c02bf.jpg

Because that's how they come? Honestly, they probably have no desire to use just one CCX, as that would get rid of the potential latency issue, and then everyone would (potentially) be buying those, undercutting their high end stuff. If it would actually cause an difference is a different story. But simple fact that we have is, CCXs come in packages of 2, we have no information that states it can come in 1. Again though, i'm not saying the yields are perfect, my original post was about how assuming the yields are "bad" based off the information of this thread doesn't make sense.

Well we do know it comes in 1ccx configuration. APU based ryzen will have 1 CCX with added GPU chip. Probably like you said, 1CCX would see a performance boost in memory + latency as well.

@Agent-A01 I think Aura89's confusion is where you said "if they had perfect yields than thread ripper would not exist(at least in it's current configuration) as that 32core could sell on EYPC for much higher profit margins." Which isn't necessarily true. For socket TR4, the CCXs appear to have 8 cores each instead of 4. Since apparently Threadrippers have only 2 of the 4 dies active, that means the 16-core variants are also a result of perfect yields; they're merely just lesser models vs Epyc. The thing to consider is getting 4 perfect CCXs for a single CPU is uncommon, and as a result will likely disproportionately cost more. Most people don't have a need for 32 cores, let alone 64 threads. Despite a 16 core costing less, it will be much easier to sell. Therefore, AMD will make more money in the end by focusing on 16 core Threadrippers. Due to the die arrangement, that should mean we'll see 16 core Epycs (where the only difference between them and a TR is the extra PCIe lanes). In a setup like that, you might find 4 cores per CCX.

No way the CCXs are 8 cores. https://*******-*******ltd.netdna-ssl.com/wp-content/uploads/2017/07/threadripper-delidded-1-800x467.jpg There are 8 ccxs here. That means a total of 24core possible any a perfect threadripper CPU. If there were 8 cores per CCX that means this chip fully unlocked is 64cores.. No way that's possible yet. Idk what the product line looks like but the 16 core thread ripper has 8 2 core CCXs to get to the 16 core amount. These CCXs are what could have been on EPYC if they were good.

Well we do know it comes in 1ccx configuration. APU based ryzen will have 1 CCX with added GPU chip. Probably like you said, 1CCX would see a performance boost in memory + latency as well.

That's confirmed?