Epyc vs threadripper

Which is going to be better for a 3d rendering workstation? 1P 32 core epyc or 16 core threadripper? The leaked epyc specs show 2ghz base 3ghz turbo, will it turbo all cores to 3ghz if cooled properly? What are the chances of anyone offering a decent ATX epyc motherboard?

Other urls found in this thread:

videocardz.com/70266/amd-epyc-7000-series-specs-and-performance-leaked
amd.com/Documents/Radeon-Pro-SSG-Technical-Brief.pdf
twitter.com/SFWRedditVideos

Depends.
Are you running 3 full speed GPUs and NVME, and PCIe SSD?
If so, probably best to go EPYC.
But if your IO is mostly unpopulated, then go threadripper and save the wallet a little hit.

>epyc

I don't think we'll ever see atx boards for that, it's just too many ram slots to fit in

More cores will be better.

Since both chips use the TR4 socket and both support 128 PCI-E lanes, there's a good chance there will be compatibility between Epyc and X399, so you may not have to choose.

Performance will have to wait until we know more, the 16-core TR should get upwards of 3000CB at stock clocks/turbo, but the question is at what temperatures and voltage?

Technically there's nothing stopping any processor based on the Zen core from scaling to ~4Ghz, so depending on just how efficient Epyc will be and how good the cooling solution is, chances are we'll see some extraordinary clocks from it. At the very least, simply based on the core increase, Epyc should get like 5000CB.

Don't forget that AMD adds cores with discrete dies, not a monolithic die like Intel, so the heat output is spread over a much larger surface area. TR4 coolers will on average have to be at least 50% bigger than anything else before just to cover the die, so expect good performance. In the case of water cooling, the potential is even more immense.

>Since both chips use the TR4 socket and both support 128 PCI-E lanes, there's a good chance there will be compatibility between Epyc and X399, so you may not have to choose.
EPYC being a server CPU is designed to run without a chipset.
And TR tops out at 64 lanes.
From a physical standpoint, sure, they're compatible in eachother's sockets as they're identical, but certain design features have changed and that leads to incompatibility.

Will it run 32 cores @ 3ghz turbo? If it's 2ghz only then a 3.6ghz threadripper is going to be a better option.

videocardz.com/70266/amd-epyc-7000-series-specs-and-performance-leaked

With 180w TDP rated, then it'll run damn near to 3Ghz, turbo is rated for 3.2, but that is likely with the 200Mhz XFR on 1-2 cores.

the temps will probably be similar to an 1800x and the voltages a bit lower, accounting for the b2 stepping

Ebyn

>And TR tops out at 64 lanes.
Ah fuck me, I had a brain flop.
Well, since Epyc can run without a chipset, doesn't that just mean that it matters even less whether or not X399 will support it, if it can run autonomously?
It would be a pretty big missed opportunity for AMD if they don't allow for this.

>Don't forget that AMD adds cores with discrete dies, not a monolithic die like Intel, so the heat output is spread over a much larger surface area. TR4 coolers will on average have to be at least 50% bigger than anything else before just to cover the die, so expect good performance. In the case of water cooling, the potential is even more immense.
The question is will AMD let us make epyc house fires with unlocked chips or atleast the ability to force all cores to max turbo. A 300w+ water cooled epyc is still less power hungry than 2p xeon setups.

Probably Threadripper. If you could get all 16 cores/32 threads up to 3.6-3.8 Mhz, you could demolish rendering and encoding/decoding (and it won't cost $4,000).

I've no idea.
We know that intel locks out server CPUs from identical workstation sockets. In some instances, this is technical. The rest of the time, it's them being jewish.

So whether EBYN will bring that much more to the table over Treadripper to make it incompatible is yet to be seen.
EYPC NDA lifts on the 20th.

but you can't ever make use of all 128 pci lanes on a standard atx board, nor use of the 8 memory channels

it's a waste to put epyc on it

I think it would be possible to put it at 4GHz, seeing that the voltage requirements will be smaller and that you have a high tier cooling solution

3.8 is a safe estimate.
We can get Ryzen to 3.9/4.0 with relative ease and the temps are kept well in control. But anything past that and the voltage stepping is insane.

well, the process in which the zen dies are built wasn't made for high clockspeeds... that comes with zen2

In a certain sense, yeah, but Asrock have also made ITX motherboards for X99 and now X299, which also throw away most of the capabilities of the system for the sake of packing a lot of cores into a very small space. And yet it seems to be a desirable enough effort for the company to actually waste R&D time to figure out how to make such a product.

From what I recall, all 2011-3 Xeons could be used on any X99 motherboard, sometimes even with ECC RAM.

Agreed. Should tear through any workstation's load with ease.

it will both tear through highly multithreaded workloads while not fucking up performance on things that can't make use of 16 cores, like those shitty xeons with shitty st performcance

The X99 ITX board gave away half the memory channels which sucked, but the X299 one has three m.2 slots and quad-channel using sodimms

It can make pretty good use of the 28 lane cpus at least (though the sweet spot for it is probably cramming 10 or 12 cores into a small case for a mini workstation, for 8 cores or less just use ryzen)

You could fit 8 slots on the board (we get that with quad channel ATX boards already) but PCB routing for it might be fucked with twice as many traces present.

With PCIe you'd be able to fully feed all 7 possible slots on ATX but a more likely use would be a bunch of M.2 SSDs on a riser card like ANUS are doing and/or offering a bunch of U.2 ports for the 2.5" nvme units.

>I don't think we'll ever see atx boards for that, it's just too many ram slots to fit in
EATX or some customized supermicro case+board would be ok. It's more if anyone will sell a workstation board, the one in the picture is EATX but it's lacking in terms of USB, though with 128 pcie lanes you could add anything. Then it depends if they artificially limit epyc with locked multipliers.

They were really stupid with the opterons back during the phenom II x6 days. The ES versions of their 12 core server versions reached 3.5ghz without any trouble then they locked the production versions to 2ghz and 100w. They could have held the highend desktop market by just releasing unlocked 12core opterons instead they decided no one has any need for more than 6 threads and let their entire lineup be crushed by core2duo products.

If the full 4-module EPYC chips are capable of sustaining 3.6Ghz all-core at 300w I will eat my favorite belt, record it, and post it here. Simple math using 1800Xs should hint at ~380w for a server bin.

you should use dual threadrippers, since they both use b2 steppings, this puts it at 310~360w (I don't remember if they're 155W or 180W)

2P threadripper is currently not possible. AMD would likely make a 1P version of Epyc for a 24 or 32c threadripper setup.

Threadripper is higher clocked again.
These EPYCs base at 2.0-2.4 with turbos ranging from 2.7 to 3.2
videocardz.com/70266/amd-epyc-7000-series-specs-and-performance-leaked

If anyone with an 1800x wants to crack the core clock down to 2.9Ghz and run a torture test to see what wattage it's yanking out the wall we should get some interesting numbers.

>These EPYCs base at 2.0-2.4 with turbos ranging from 2.7 to 3.2
How does the turbo work on the newest AMD chips? Does it mean they will do 3ghz on all cores provided temperatures are fine? Or 3ghz only when many cores are idle.

There's a 50% clockspeed difference between base and turbo on 32 core epyc.
Base clock might be with all PCIe populated and mem channels running full tilt.

Normal all core turbo in x86 without overkill I/O should be close to 3.0ghz

I mean for the wattage and clocks calculation he's making

I believe the turbo is dependant upon cooling yes. Given that these are server CPUs they're never going to run hot, so they should turbo up nicely when required.
Though how much of that turbo is XFR if at all. The XFR boost will only effect 1-2 cores.

AMD missed the memo again, Intel has a 205W part, why doesn't Epyc? More clockspeed certainly wouldn't hurt.

>Efficiency is bad
I love how Intel shills turned this around in the space of 4 months of Ryzen.

Efficiency is a product of performance/power
205W epyc would barely be less efficient than a 180W one, it's just another missed market.

The only reason Intel is hitting 205w is because of how inefficient their monolithic dies are compared to the MCM design employed by AMD.
More TDP is not a "Missed market" in servers.

>205W epyc would barely be less efficient than a 180W one, it's just another missed market.
The problem is both intel and AMD refuse to acknowledge workstations exist. For 3d all I care about is getting the most performance out of a single PC a 300watt CPU is not much of an issue, the alternative is shoving multiple GPUs which are 300 watts each.

If there was easy access to highly clocked server chips there wouldn't be such a push towards GPU raytracing in the first place. A highly clocked 32 core epyc would likely be pretty competitive with GPU rendering but without the limitations that come with GPU use. For some reason intel and AMD have decided a single PC must never have more than 200-250 watts of power while nvidia will happily let me cram 4x titans in a single PC, needless to say 1200 watts of GPUs beat a 200 watt cpu even if efficiency is lower. It's not feasible to run the real time viewport or texture baking on server farms or the cloud, the bottleneck is purely how much processing power you can cram into a single computer.

It's worse when you figure out nobody ever mentions all core turbo.

CPUs just aren't that great scaling up for pure thoroughput though. SISD/Superscalar is just not dense enough to achieve the general performance scaling for general purpose computing people wanted in the day which were mostly statistical, linear algebra, convolution and other signal processing in nature. RISC SIMD with huge vector pipelines (like GCN) works the best and is much more power efficient in those use cases. I think the tables will start turning for GPU driven content creation with Radeon Pro SSG and other massive texture buffer devices. I'm not sure if TDP has ever been a hard restriction for AMD/Intel, I think it's just because it's harder to scale their architectures that way, and since multisocket exists, there's no reason to make larger dies. IF helps to some extent I suppose.

>8c/16t EPYC
what would be the use case of something like that?

Not every small business needs a monsterous 32 core server CPU.
Low core count, but ALL of the same IO and memory capacity.

why would a small business need 64MB of L3, octa channel ram and 128 PCI lanes while not needing a lot of cores? it just doesn't make sense

The same people who bought 2 socket xeons to get 88 lanes.
Nowadays you're more I/O and memory bound(Dimm count, bandwidth is mostly for cloud) than pure compute bound.

I/O server.

Thin clients with dedicated GPU passthrough?
It's just another option for individuals.

Hell. Or even think of it like a light workstation chip on a server socket. 128 lanes and 2TB of RAM on a workstation would pack a punch.
Ultimately it's just a 4 CCX design with 2 cores on each complex due to binning and failures. Why throw them out when you can make money on them.

There's bound to be use-cases where you just want to stick 8 GPUs into the system and don't care much about CPU power.

motherfuckin miners would eat that shit up like a wet dream that came to life

Why? They need a lot of x1 slots and risers.

128 pci-e lanes, so they can have 8 gpus in one system. not sure what a riser is, and does mining only need 1x?

Are you dumb?

yeah, seems like it. i think i don't know what im talking about, ill go google it

>CPUs just aren't that great scaling up for pure thoroughput though. SISD/Superscalar is just not dense enough to achieve the general performance scaling for general purpose computing people wanted in the day which were mostly statistical, linear algebra, convolution and other signal processing in nature. RISC SIMD with huge vector pipelines (like GCN) works the best and is much more power efficient in those use cases. I think the tables will start turning for GPU driven content creation with Radeon Pro SSG and other massive texture buffer devices. I'm not sure if TDP has ever been a hard restriction for AMD/Intel, I think it's just because it's harder to scale their architectures that way, and since multisocket exists, there's no reason to make larger dies. IF helps to some extent I suppose.
The issue is the modeling program runs on the CPU, modern CPU renderers barely care about polygon count as long as it fits inside ram. It's common to have hundreds of millions of polygons, due to instancing, displacement and hair, I can use about 120 million poly on 16 gb of ram for professionals you could expect them to be using 32gb of geometry and another 32gb of uncompressed 16bit images. Even if you have 64gb of VRAM it takes ages to export all the polygons and convert image maps to something the GPU can handle. For quick test renders or animations where it only takes a few minutes to render each frame, the time wasted on initializing the gpu renderer is too much.

I'm not sure if vega can solve these issues with it's new memory controller if it can stream data from system ram it would help but I'm not sure if it's possible to stream the geometry.

The ideal solution would probably be an 8 core ryzen + vega and 16gb of HBM on a MCM as a coherent cache but AMD has so far refused to make high end APUs.

>128 pci-e lanes, so they can have 8 gpus in one system. not sure what a riser is, and does mining only need 1x?
They don't care about bandwidth, If you could put 10x low clocked rx480 GPUs with 1mb of ram on a single card and have them all communicate via an MCU over serial it would be better than anything else.

>not sure what a riser is
Think of them like extenders but they can also be used as 1x to 16x adapters.

>does mining only need 1x
It doesn't need more lanes. Mining isn't IO bound.

Housefires are acceptable now, goy.

this is exactly why I'm still considering stock 1700, because in terms of 24/7 use I still see it as best server tdp, even beats out e3-12xx 80w

I get the impression threadripper and epyc are the same chip, just one made for dual socket. I don't think most home/power user scenarios will get value from epyc unless you've got serious server load.

>threadripper and epyc are the same chip
no, threadripper has 2 dies and epyc has 4

the dies are built equal tho

thanks, i get it now

How many GPUs can it support?

As many as you can physically fit, I guess around 12 single slot cards?

AMD Threadripper 16c/32t @4ghz will draw 330W from the socket.
2x 1800X@4ghz(165W) = 330W.

That's why you don't overclock 16 cores.

Intel core i9 10c/20t housefire edition draw 340W @ 4.7ghz

I'm not defending Intel's housefires, I just generally advise against overclocking high core count chips due to stability, heat and power.
Hugh core chips hit the latter two far sooner than low core count chips, thus getting 10% more performance for twice the power used is dumb.

Yet two 1700 at 3.2Ghz would use around 130 watts..

That's how it works. As soon as you have to touch the voltage, the power draw scales to the square of the frequency.

no way that is correct

1700 using a cTDP of 35W scores 1000 in Cinnebench.
The 7900k at stock does around 2100, it pulls over 200W during turbo.

The power efficiency is on a whole different magnitude.

I wish that Epyc started at a high-clocked 2d/12c model so I could get a longer single-system upgrade curve.

now:
> expensive 1S / 16DIMM board
> 4x 32 GB ECC RDIMMs
> 2d/12c Threadripped-like CPU with ~3.5/4.0 GHz clocks, plus 1:1 clocked IF mesh

possible midpoint upgrade:
> 8x 32 GB ECC RDIMMs
> 2d/16c high clocked Epyc2

as far as I would ever reasonably take it:
> 12x 64GB ECC RDIMMs (and only if 64 GB DIMMs are reasonable in the future)
> 3d/36c Epyc2/3

max (this will never happen):
> 16x 128 GB ECC RDIMMs
> 4d/48c 7nm Epyc2/3

Pointless to have so few cores on such a large chip. If you wanted single threaded performance you would just buy a consumer Ryzen like the 1600X to get high turbo when few cores are used. If you wanted multi thread performance then why would you care if each core is clocked low when you could have way more cores instead?

Just because you have a high core chip doesn't mean you trade ST performance, turbo is a thing, and turbo on server chips with huge TDPs can push a single core way over the chips rated turbo clocks.

So 3.9Ghz single core on a 2.3ghz base 20+ core chip is not uncommon

main problem is that Ryzen/Epyc are UDIMM only, limiting you to 64/128 GB total, and even ECC support is fuzzy.

the recent model rumors show 2/3/4 die Epyc variants with 4/6/8 DDR4 channels, but the lowest 1s/2d/16c variant has pretty poor base clocks, and no info exists on how reliably it can hit its all-core turbo.

You mean TR not Epyc? Who would even use epyc if it didne support RDIMM,even bulldozer did

yeah, sorry, clearly mean threadripper.

ECC RDIMMs aren't cheap nor will even 1S Epyc boards, but if the Threadripper platform is limited to only twice the memory (8*16GB UDIMMs) I currently have in my workstation, it's hard for me to seriously consider it a real 3-4 year investment.

I only need TR-level compute power now (and would prefer decent single core performance), but I can definitely imagine wanting 256+ GB in a year or two.

Do we even know if TR is UDIMM only? ASUS (or asrock, can't remember) was advertising ECC support for X399 during computex.
RDIMMs shouldn't be outside the realm of possibility.

Word on the street is UDIMM ECC yes, RDIMM(/LRDIMM/etc) anything no, so low-end Epyc sales don't get cannibalized.

Not everybody in the market for a 16c server needs all the memory encrypting crap or up to 128 PCIe lanes that Epyc provides, but they surely want at least 256 GB memory support.

Do UDIMMs top out at 16GB? Seems low.

SSG was pretty much invented for just what you're describing (GPU requiring access to way more shit than it could ever fit in memory)

amd.com/Documents/Radeon-Pro-SSG-Technical-Brief.pdf

Sadly yes (along with 2DPC), which is why Ryzen max capacity is specified as 64GB, and why TR is all but guaranteed stuck at 128GB.

Guess this is thr big difference between epyc and TR, also it's known that dual rank memory gives a nice perf boost over single rank, quad rank should be even better.

Disappointing, but hey, if TR could do everything EPYC could, there'd be no point in EPYC

Efficiency is only good when Intel does it.

Software with per-core licensing.