For me it was just convenience - I bought a complete workstation (tower) with that CPU in it (and with all its capabilities, like 128 PCIe lanes and 8 memory channels). Does the EPYC server chip use a different socket (?), so I'd have to build a tower workstation from the components myself, a server mobo, server CPU squeezed into a tower?
I once spent a good while trying to figure out why a 2U server wasn't recognizing new drives. No messages, no errors, no lights, nothing.
Finally it turned out that it's not a 2U server with 12 bays, but two separate servers in one box, side by side. Half the bays go to one, half to the other. Each side has two CPUs in it, and one was completely powered off.
Be careful, sometimes these server boards will be missing stuff that you'll want in a desktop. ASRock Rack boards often have insomnia. That is, they don't support low power states.
In the last generation (and presumably true for this one), Threadripper Pros retained higher single core boost frequencies. Useful for mixed workstation loads, where you might still end up with some applications that are single/low threaded, and the extra boost is noticeable.
Also, since these will mostly sell through OEMs, I imagine many of these might get put together with the whole validated hardware+software package deal thing.
An appropriate cooling system for a 2x Epyc mobo will cause so much noise I'm not sure we could call it a workstation. Well, you could try to, but nobody else in the room will hear you.
If that had been true, then I guess there would be no offerings of "workstations" that pack server-grade CPUs into the tower cases ready to be used as more powerful desktop machines. I happen to have one at home (dual socket) and it's silent just as a regular machine.
Noise that we find in server machines is because of the fans pushing loads of air basically out through the 1U/2U tunnel. With tower cases that is not the case and hence there's no noise.
One would be fine - absolute flat out pair of epyc chips is still under 800W and you can't overclock them so running a water delta of 20 degrees doesn't matter very much.
Workstation motherboards are different things to server boards. Threadripper pro blurs the lines a bit but there's a decent chance it still comes with conveniences like audio and a load of USB ports that the server boards won't bother with. The chassis will look more like (and maybe sound more like) a desktop too.
I'm using a single socket epyc milan as my daily driver and it's great. 64 cores and 512gb of memory is is good times. But that only works for me because it's headless and the thin clients talking to it can do things like make noises and connect to bluetooth keyboards.
This isn't the announcement yet. Don't bother reading the article, it's just speculation. Probably well informed speculation, but since the actual announcement is coming out in just a few hours, why not just wait?
You're right that it probably won't run that hot, but AMD will probably keep the TDP up at 350W with everything else - they did that with the last generation of Threadripper Pro - everything has a 280W TDP including their 12C/24T model.
Since this will mostly be an OEM part, I think that makes sense. By slapping the same claimed TDP on all cores, vendors can offer all SKUs using the same thermal solution (though obviously the low end will probably be overspecced).
Also, I don't think it'll be down at 120-130W - the 7900X (also 12c/24T) has a TDP of 170W, and a max socket draw (which it will hit) of 230W.
That's with a base clock of 4.7GHz, though. It's very high, but not impossible.
The 12-core Ryzen 9 7900X has a TDP of 170W, and that's with a significantly smaller package. The 16-core EPYC 9174F already has a TDP of 320 W with a base frequency of 4.1GHz and a boost frequency of 4.4GHz, in a package quite similar to Threadripper.
This might be anathema to some people, but how much does the TDP change if you activate ECO mode on these bigger chips? From my testing on an older Ryzen, ECO mode left single core performance unchanged, multicore numbers dropped like 4%, and temperatures went down over 5C. All of which is totally worth it to me.
Theoretically they could pack in 192 cores with the 16-core die, but haven't done so (on EPYC) for reasons unknown to me. Maybe a physical packaging limitation?
Probably thermals, at ~3W/core TDP, 192 cores is gettingg close to 600W to dissipate from a single socket, which is beyond all off the normal heatsink/watercooling solutions I'm aware of
There are some pretty interesting cooling solutions in development. Will be interesting to see which ones become economical enough for mass market use.
That's Zen 4c, which I don't think is ready yet? I'm pretty sure 192c/384t* at ~4Ghz EPYC is on the way, although that might be a Zen 5c product (so... 2025, 26?)
*wow, core counts so big I found it slightly difficult to multiply by 2 in my head.
- "Previous Ryzen Threadripper launches have seen a DIY launch happening several months later after the launch of the OEM/Pre-built platforms."
This anti-competitive behavior should be illegal and it should be prosecuted. Clearly, there's some kind of backroom agreement where these "preferred" OEM's are giving AMD money, or some consideration of value, in exchange for locking out their competition. What other, benign explanation is there?
(Not unlike Intel paying OEM's illegal kickbacks to lock out AMD [a], come to think of it).
If they actually paid in advance for the whole production run for those months how is that any different compared to the product being released but shelves being empty?
Because they product product of which that CPU is a fundamental part of the supply chain.
So they have at least initial contract for supply before they even start designing this (mostly work station) products (which often involve e.g. custom motherboards).
I.e. not only are they buying before release, they _need_ supply guarantees long long before release.
The customer can't do that for multiple reasons.
- They haven't signed a NDA contract nor are they in direct contact with first party sells personal so when it's time to early order they don't even know it exist.
- They buy in quantities of 1,2 maybe 10 instead of multiples of thousands (initial charge) with follow up guarantees. E.g. if you look at server CPUs you will see 1kU prices, because this CPUs are sold only (by AMD or Intel directly) in multiples of 1000 units.
If you (or lets say a PC retailer) do buy multiple thousand thread ripper CPUs at once and are willing to enter a contract for this sells long before they are sold (which likely involves an initial % of the price as down payment) you probably could get them early one. Through you would still be competing with DELL and similar for limited supply so you might still have a hard time to get the very first charge. And you also might have some problem getting compatible motherboards as many ORMs produce their own.
Well, sure. The difference is in intent. It's something where I would want court-appointed lawyers reviewing these company's internal communications, and taking depositions, to figure out.
"Oops we sold out" is one thing. "Oops; we conspired with other companies to limit supply and to form a cartel" is entirely another. The industry's been caught doing the latter too often, and governments' oversight is IMO too slow, too weak, and lacking meaningful deterrent power. I think they need to be much more aggressive. This stuff victimizes consumers.
and as far as I can tell it's quite unlikely that there is bad intent involved
as far as I understand DIY sells of Threadriper solely exist for PR, mind share reasons a bit of edge case handling and maybe some book keeping trick to even out certain marked OEM demand fluctuations.
but I don't think they make relevant amounts of profit with DIY sells
I mean think about it most (by far) end users of Threadrippers are companies needing it for work stations.
Most larger companies only buy from OEMs like DELL for various reasons like warranty handling, uniform platforms, certain "enterprise" BIOS features etc.
Even for small companies the DIY part is often an issue, they might not have IT experts and even if this system goes e.g. to a software developer they might not be willing to hardware maintain their PC (and take responsibility for it).
This leafs a relative small number of private enthusiasts and small companies, to small to really be profitable.
Add to that that the demand of DIY is less predictable then that of OEMs, and production capacities for CPUs tend to be limited and even if less so currently tend to still often doesn't allow fast dynamic changes of the amount.
Idk. but if I would be AMD I think I would pin down my big customers on contracts with a more fixed demand (both fixed initially and fixed+%variety follow up) and then plan CPU production of that (with maxed out variety) and if there are any spare factory resources I would use them for either server CPUs or desktop CPUs whatever currently sells hotter. Then if the OMS don't max out their variety I would stock the remainder and once enough stock accumulated pass it the the consumer facing department to sell as DIY parts.
How is selective selling the same as paying a company to not buy from your competitor? The knee-jerk reaction here to call things they don't like "enshitification" or "anticompetitive" is really annoying.
you don't need backroom deals, it's just normal marked because
1. it's not a consumer product but mainly an OEM product
2. it's supply limited (much more so then server or desktop CPUs as the marked is in general more niche)
3. it's an OEM product, i.e. it only sells to companies always in large amounts and they put it into their product. BUT for such supply chains a minimal quantity is needed to be sold per customer. Or the customer won't buy because it's to little to do a product launch.
5. Only if there are left overs is selling to end users even considered, not because of backroom deals, but because it financially makes no sense at all. I would even go as far as to say AMD makes no (relevant degree of money) profit from the DIY sells of Threadripper.
Or in other words it's the standard procedure for companies selling to companies in case of limited demand to first sell to your second best customer as little as you can to satisfy their minimal needs, then continue to your 2nd best customer, 3rd best etc. If there is a huge step down between nth and n+1th best also maybe revisit the first n best customers first and increase their quantity.
"the new AMD Ryzen Threadripper 7000 Series and select Ryzen Threadripper PRO 7000 WX-Series processors will be available from DIY retailers starting on November 21st"
So this time it appears that it will be a real product launch for everyone, not just for Lenovo or the like.
Multiple 7950X systems will always be faster than any of these Threadrippers for a given amount of money, but 7950X is limited to only 64 GB of fast memory (populating two sockets per channel slows a lot the memory), so anyone who needs a lot of memory (up to 1 TB) in a single system might want a 7960X (24 cores @ $1500).
The models with more cores have worse performance per dollar, so it is hard to justify their use instead of using multiple servers, for higher performance at a lower cost.
You can use them for servers, but they're meant for workstations with resource-hungry applications.
There even are companies that make a threadripper-box-with-a-handle-and-wheels when you need that oomph to be deployable to locations.
Can't speak to whether there are backroom deals, but from a development perspective, we do this because AMD relies on it's partners to help qualify their products for production. We get early samples and work with AMD to sort out any bugs (and even ship with some bugs that are mitigated) in order to achieve faster time to market. Once AMD feels the product is stable enough for general availability, they then open up sales to the general public.
I wonder how this beast with AVX-512 compares to GPUs for AI model training and execution?
I'm sure GPUs would still be more efficient for the specialize task of heaps of matrix math, but given the insane price premium on high-end nVidia cards I wonder if this could be an alternative?
96 AVX512 cores is roughly 3072 parallel fp16 calculations per cycle.
This lacks context. Avx-512 is disabled (fused off) in the big cores of Intel big.little. In the server and workstation chips that only have big cores, it is functional.
Any software developer can easily use any number of available cores for a proportional speed-up of compiling projects.
Nevertheless, Threadrippers are not really useful for software development, because using two computers with 7950X and 64 GB memory each as compilation servers would provide a much higher speed than one computer with a Threadripper 7960X and 128 GB memory, which would cost about the same.
Threadrippers are not useful for who wants many cores, because the 7950X cores are much cheaper, they are useful for those who want a large amount of memory in a single box or a large amount of PCI lanes (e.g. for multiple GPUs) in a single box.
For me personally this is too little too late. About 5 years ago I could have really used a super powerful multicore processor in my local desktop. But my company has since moved to the cloud and I just use ssh on a 128 core VM based on AMD Epyc.
The ryzen 9's with 16 cores a fantastic little machines for general development work, and burn through compiling most things. But I do a fair bit of linux kernel work too, and I would really like more cores. The old threadrippers were the obvious target, but the MSRP on this 32 core is fully 4x+ the 16 core Ryzen. I get they are protecting their low end epyc business but I am desperately praying for some intel competition in this space. I still use my intel HEDT machine on a daily basis and while I picked it up right after a refresh and received a significant discount that machine with a 1080TI still cost less than the ryzen 9.
So, hey AMD figure out a reasonable way to product segment your workstation users so you can flatten the price curve a bit so I don't have to keep buying used servers.
Well, in my case I have access to a couple servers with >250 threads, and it will happily peg all those out until it has to put vmlinux together, then it goes down to a thread or two for a bit, and then they all go back to 100% while the modules are processed.
So, yes, but presumably given enough threads the limit becomes how fast one can collect and link the vmlinux which is mostly still done serially ala Amdahl's law. Oh and there is some IO/etc bounding at the beginning checking the config/etc. Some sites have kernel compile benchmarks and for the most part they tend to indicate nearly linear scalability. Although there tend to be two major config differences, some people benchmark with something like an allmodconfig vs people who benchmark with a more minimal kernel module setup. The former are the ones that can be as slow as an hour or so on machines without a lot of cores, and the latter are the ones that finish in sub 60 seconds on fairly modest hardware. But given enough cores its possible to turn the former into he latter.
C and C++ compilation model allows compilation of separate translation units in parallel so yes. Linking compiled translation units together is also taking advantage of multiple cores but it's not easily "parallelizable" as compilation step is.
71 comments
[ 4.6 ms ] story [ 136 ms ] threadYou can get perfectly normal ATX motherboards for Epyc. There's even a mini-ITX one for some reason:
https://www.asrockrack.com/general/productdetail.asp?Model=R...
So that you cat fit 2 motherboards in that 2 CPU setup! a.k.a. redundant motherboards. /s
Finally it turned out that it's not a 2U server with 12 bays, but two separate servers in one box, side by side. Half the bays go to one, half to the other. Each side has two CPUs in it, and one was completely powered off.
Also, since these will mostly sell through OEMs, I imagine many of these might get put together with the whole validated hardware+software package deal thing.
Noise that we find in server machines is because of the fans pushing loads of air basically out through the 1U/2U tunnel. With tower cases that is not the case and hence there's no noise.
The real issue is that there are no non-Pro threadrippers anymore. Those were even cheaper.
I'm using a single socket epyc milan as my daily driver and it's great. 64 cores and 512gb of memory is is good times. But that only works for me because it's headless and the thin clients talking to it can do things like make noises and connect to bluetooth keyboards.
WCCF is known for jumping the gun. They are not SEO trash or a tech tabloid, but grains of salt are required.
Since this will mostly be an OEM part, I think that makes sense. By slapping the same claimed TDP on all cores, vendors can offer all SKUs using the same thermal solution (though obviously the low end will probably be overspecced).
Also, I don't think it'll be down at 120-130W - the 7900X (also 12c/24T) has a TDP of 170W, and a max socket draw (which it will hit) of 230W.
The 12-core Ryzen 9 7900X has a TDP of 170W, and that's with a significantly smaller package. The 16-core EPYC 9174F already has a TDP of 320 W with a base frequency of 4.1GHz and a boost frequency of 4.4GHz, in a package quite similar to Threadripper.
Theoretically they could pack in 192 cores with the 16-core die, but haven't done so (on EPYC) for reasons unknown to me. Maybe a physical packaging limitation?
*wow, core counts so big I found it slightly difficult to multiply by 2 in my head.
This anti-competitive behavior should be illegal and it should be prosecuted. Clearly, there's some kind of backroom agreement where these "preferred" OEM's are giving AMD money, or some consideration of value, in exchange for locking out their competition. What other, benign explanation is there?
(Not unlike Intel paying OEM's illegal kickbacks to lock out AMD [a], come to think of it).
[a] https://news.ycombinator.com/item?id=30087546
So they have at least initial contract for supply before they even start designing this (mostly work station) products (which often involve e.g. custom motherboards).
I.e. not only are they buying before release, they _need_ supply guarantees long long before release.
The customer can't do that for multiple reasons.
- They haven't signed a NDA contract nor are they in direct contact with first party sells personal so when it's time to early order they don't even know it exist.
- They buy in quantities of 1,2 maybe 10 instead of multiples of thousands (initial charge) with follow up guarantees. E.g. if you look at server CPUs you will see 1kU prices, because this CPUs are sold only (by AMD or Intel directly) in multiples of 1000 units.
If you (or lets say a PC retailer) do buy multiple thousand thread ripper CPUs at once and are willing to enter a contract for this sells long before they are sold (which likely involves an initial % of the price as down payment) you probably could get them early one. Through you would still be competing with DELL and similar for limited supply so you might still have a hard time to get the very first charge. And you also might have some problem getting compatible motherboards as many ORMs produce their own.
"Oops we sold out" is one thing. "Oops; we conspired with other companies to limit supply and to form a cartel" is entirely another. The industry's been caught doing the latter too often, and governments' oversight is IMO too slow, too weak, and lacking meaningful deterrent power. I think they need to be much more aggressive. This stuff victimizes consumers.
and as far as I can tell it's quite unlikely that there is bad intent involved
as far as I understand DIY sells of Threadriper solely exist for PR, mind share reasons a bit of edge case handling and maybe some book keeping trick to even out certain marked OEM demand fluctuations.
but I don't think they make relevant amounts of profit with DIY sells
I mean think about it most (by far) end users of Threadrippers are companies needing it for work stations.
Most larger companies only buy from OEMs like DELL for various reasons like warranty handling, uniform platforms, certain "enterprise" BIOS features etc.
Even for small companies the DIY part is often an issue, they might not have IT experts and even if this system goes e.g. to a software developer they might not be willing to hardware maintain their PC (and take responsibility for it).
This leafs a relative small number of private enthusiasts and small companies, to small to really be profitable.
Add to that that the demand of DIY is less predictable then that of OEMs, and production capacities for CPUs tend to be limited and even if less so currently tend to still often doesn't allow fast dynamic changes of the amount.
Idk. but if I would be AMD I think I would pin down my big customers on contracts with a more fixed demand (both fixed initially and fixed+%variety follow up) and then plan CPU production of that (with maxed out variety) and if there are any spare factory resources I would use them for either server CPUs or desktop CPUs whatever currently sells hotter. Then if the OMS don't max out their variety I would stock the remainder and once enough stock accumulated pass it the the consumer facing department to sell as DIY parts.
1. it's not a consumer product but mainly an OEM product
2. it's supply limited (much more so then server or desktop CPUs as the marked is in general more niche)
3. it's an OEM product, i.e. it only sells to companies always in large amounts and they put it into their product. BUT for such supply chains a minimal quantity is needed to be sold per customer. Or the customer won't buy because it's to little to do a product launch.
5. Only if there are left overs is selling to end users even considered, not because of backroom deals, but because it financially makes no sense at all. I would even go as far as to say AMD makes no (relevant degree of money) profit from the DIY sells of Threadripper.
Or in other words it's the standard procedure for companies selling to companies in case of limited demand to first sell to your second best customer as little as you can to satisfy their minimal needs, then continue to your 2nd best customer, 3rd best etc. If there is a huge step down between nth and n+1th best also maybe revisit the first n best customers first and increase their quantity.
"the new AMD Ryzen Threadripper 7000 Series and select Ryzen Threadripper PRO 7000 WX-Series processors will be available from DIY retailers starting on November 21st"
So this time it appears that it will be a real product launch for everyone, not just for Lenovo or the like.
Multiple 7950X systems will always be faster than any of these Threadrippers for a given amount of money, but 7950X is limited to only 64 GB of fast memory (populating two sockets per channel slows a lot the memory), so anyone who needs a lot of memory (up to 1 TB) in a single system might want a 7960X (24 cores @ $1500).
The models with more cores have worse performance per dollar, so it is hard to justify their use instead of using multiple servers, for higher performance at a lower cost.
But yes, there's definitely a memory speed limitation there.
I'm sure GPUs would still be more efficient for the specialize task of heaps of matrix math, but given the insane price premium on high-end nVidia cards I wonder if this could be an alternative?
96 AVX512 cores is roughly 3072 parallel fp16 calculations per cycle.
It's right here in black and white. https://www.intel.com/content/www/us/en/products/sku/233483/...
The other 3 are running at 15% utilization
Nevertheless, Threadrippers are not really useful for software development, because using two computers with 7950X and 64 GB memory each as compilation servers would provide a much higher speed than one computer with a Threadripper 7960X and 128 GB memory, which would cost about the same.
Threadrippers are not useful for who wants many cores, because the 7950X cores are much cheaper, they are useful for those who want a large amount of memory in a single box or a large amount of PCI lanes (e.g. for multiple GPUs) in a single box.
So, hey AMD figure out a reasonable way to product segment your workstation users so you can flatten the price curve a bit so I don't have to keep buying used servers.
So, yes, but presumably given enough threads the limit becomes how fast one can collect and link the vmlinux which is mostly still done serially ala Amdahl's law. Oh and there is some IO/etc bounding at the beginning checking the config/etc. Some sites have kernel compile benchmarks and for the most part they tend to indicate nearly linear scalability. Although there tend to be two major config differences, some people benchmark with something like an allmodconfig vs people who benchmark with a more minimal kernel module setup. The former are the ones that can be as slow as an hour or so on machines without a lot of cores, and the latter are the ones that finish in sub 60 seconds on fairly modest hardware. But given enough cores its possible to turn the former into he latter.