Threadripper 3000, 3000g

M

Mega6

2[H]4U
Joined
Aug 13, 2017
Messages
4,037
Last edited:
Definitely want to see sustained clockspeeds on the 3950X; if AMD has binned it up, would be a great 'no compromise' CPU.

Also excited about the new TR; 4.5GHz on the top end is very, very cool for 24 and 32 total cores. Though the CPU likely won't be 'cool' itself, AMD has done a pretty good job of keeping TDPs under control.

At US$50, the 3000G is interesting for its price -- but it's a hard sell at just two cores. It's absolutely enough for what most consumers do and even for most modern low-intensity games (think League of Legends etc.), but the TechPowerUp article mentions HTPC and the like, and well, you generally want more cores for that if you're going to be doing more than your average Synology is capable of.
 
November 25.

Yikes. Want? Yes. Want badly enough to buy into new socket/chipset and a healthy price on new CPU?

Probably not. Might pick up a second-gen Threadripper off someone who does, however.
 
3950 is too expensive. Water cooling optimized? Not hard to crack that code. It all looks good, dont get me wrong, but that 3950 is an expensive ask when you factor in required water cooling.
 
Smoked Brisket said:
3950 is too expensive. Water cooling optimized? Not hard to crack that code. It all looks good, dont get me wrong, but that 3950 is an expensive ask when you factor in required water cooling.
Click to expand...
Like an extra $100 for a 240 or 360 AiO, at that price point the cooler isn’t significant.
 
  • Like
Reactions: N4CR
like this
Lakados said:
Like an extra $100 for a 240 or 360 AiO, at that price point the cooler isn’t significant.
Click to expand...

I think 750 is too much of a jump from the 500 3900x, so at 850 plus, the 3950 is too expensive. Just my opinion. There is a strong possibility that if AMD is recommending water cooling, and they are here, then that AIO is not going to be sufficient, at least to see temps that most would be happy with.
 
drescherjm said:
AMD knows Intel has no competitive product on the mainstream platform so they can get away with a much higher price than the 3900x
Click to expand...
They also have a shit tone of R&D to pay off with these chips. But I can’t imagine having something capable of handling quad channel memory and all those PCIE lanes doesn’t add to that price tag just a bit. Also it is a business product you always pay more in the professional space, it’s for all those extra compatibility checks and assurances and what nots.. So while expensive if any of those statements made by those guys at Blurr were true than that cost had an ROI of like 3 weeks.
 
What's it going to take to cool the new TR chips? I hope it's not foolish to presume a 360mm AIO would do the job.
 
Smoked Brisket said:
I think 750 is too much of a jump from the 500 3900x, so at 850 plus, the 3950 is too expensive. Just my opinion. There is a strong possibility that if AMD is recommending water cooling, and they are here, then that AIO is not going to be sufficient, at least to see temps that most would be happy with.
Click to expand...
I would imagine that these have a similar thermal profile to the EPYC chips, which you can happily cool on air so long as you don’t mind working next to a jet engine. A triple fan AiO should handle that with no issues.
 
Smoked Brisket said:
3950 is too expensive. Water cooling optimized? Not hard to crack that code. It all looks good, dont get me wrong, but that 3950 is an expensive ask when you factor in required water cooling.
Click to expand...

If you don't have a use for it, there are plenty of lower priced options in the Ryzen 3000 family. Even older 16 core CPU's that are much slower (Xeon e5-2698 v3 is cheapest I can find) has a much higher TDP and still costs in the $425-500 range used. So I actually think the price is fair.
 
Last edited:
"According to performance numbers put out by AMD, the Ryzen 9 3950X offers up to 22 percent higher single-threaded performance than the Ryzen 7 2700X"

Educate me here ppl, is this really such a massive jump over a not so spectacular single-thread performer?
 
Auer said:
"According to performance numbers put out by AMD, the Ryzen 9 3950X offers up to 22 percent higher single-threaded performance than the Ryzen 7 2700X"

Educate me here ppl, is this really such a massive jump over a not so spectacular single-thread performer?
Click to expand...
For them that is a massive jump, and it puts them “on par” with the Intel 8&9 gen.. So it is more than enough to keep them competitive.
 
I was waiting for Threadripper for my next upgrade, not so much for the core count, but more for the extra PCIe lanes, but in all honesty, the max boost of 4.5Ghz is a little bit of a turnoff.

That's the same as the 3800x.

Since I don't care about the extra core count above 8C/16T, I'd be paying a lot of money over the 3800x for the extra PCIe lanes I want. Disappointing. I was hoping this would be a 4.8Ghz boost part. :/

All I want is a no compromises chip.

I don't care about core count, but I am unwilling to sacrifice on per thread performance or PCIe lanes. Why is this so goddamned hard?

Why on earth do they always have to tie having an even halfway reasonable PCIe lane count to having excessive cores, and lower max clocks? These things are unrelated!

My ideal chip would be a super-binned 8C/16T with the best possible boost clocks and all the Threadripper PCIe lanes. Hvaing to compromise on PCIe lanes to get the chip I want is really goddamned frustrating.
 
Me want that 24c TR but I'm not going to drop $2k into it. I am cool chugging along with my 7820x.
 
Auer said:
"According to performance numbers put out by AMD, the Ryzen 9 3950X offers up to 22 percent higher single-threaded performance than the Ryzen 7 2700X"

Educate me here ppl, is this really such a massive jump over a not so spectacular single-thread performer?
Click to expand...

It seems that AMD is hinting that the 3950X will sustain higher clockspeeds, so in addition to the IPC issues they sorted on Zen 2, there should be more performance on tap.

Of course, this isn't likely to be more than a few percent faster than other Zen 2 CPUs per core. How much faster is what I find interesting.
 
Zarathustra[H] said:
My ideal chip would be a super-binned 8C/16T with the best possible boost clocks and all the Threadripper PCIe lanes. Hvaing to compromise on PCIe lanes to get the chip I want is really goddamned frustrating.
Click to expand...

Yeah, life is full of compromises. Pay up.. isn't that the real answer?
 
All that said, according to AMD the difference in single threaded performance between a 4.5Ghz boost 3800x and a 4.8Ghz boost is only about 3%, despite the 6-7% clock speed difference. (I guess they have diminishing returns at higher clocks)

That suggests that I would not be sacrificing much in the way of single threaded performance by going with a 24C Threadripper. I mean, I'd be paying a lot of money for 14 cores, 16 of which I have no use for, but I'd still get my precious PCIe lanes....

It's a pricy chip. Has anyone seen any predicted pricing on the TR40x Motherboards yet? I've been looking through the launch materials, but I'm not finding anything.

It would be good to know just how much of a price premium I'd be paying just for the luxury of not having a gimped number of PCIe lanes.

Even the lowest of low end chips should not have only 24 lanes. I don't see how anyone gets way with under 32, with 40 lanes being the cutoff for me where I am no longer unhappy with PCIe lane count.
 
Last edited:
Zarathustra[H] said:
All that said, according to AMD the difference in single threaded performance between a 4.5Ghz boost 3800x and a 4.8Ghz boost is only about 3%, despite the 6-7% clock speed difference. (I guess they have diminishing returns at higher clocks)

That suggests that I would not be sacrificing much in the way of single threaded performance by going with a 24C Threadripper. I mean, I'd be paying a lot of money for 14 cores, 16 of which I have no use for, but I'd still get my precious PCIe lanes....

It's a pricy chip. Has anyone seen any predicted pricing on the TR40x Motherboards yet? I've been looking through the launch materials, but I'm not finding anything.

It would be good to know just how much of a price premium I'd be paying just for the luxury of not having a gimped number of PCIe lanes.

Even the lowest of low end chips should not have only 24 lanes. I don't see how anyone gets way with under 32, with 40 lanes being the cutoff for me where I am no longer unhappy with PCIe lane count.
Click to expand...


So get gen2k or gen 1k, or buy used. You want your cake and to eat it too.
 
that must suck when there are chips that have 48 pci lanes

and have been out for 3 years!
 
sunruh said:
that must suck when there are chips that have 48 pci lanes

and have been out for 3 years!
Click to expand...

And which would they be?

There are no chips with top level max clocks from either Intel or AMD that have OK numbers of PCIe lanes.

Intel has low PCIe lane counts on all of their consumer (lga1151) chips. They also limit their numbers of PCIe lanes on their lower core count LGA2066 chips and the higher core count versions which do have the full number of PCIe l have lower max boost clocks.

I don't understand why neither AMD not Intel have any top clocked 8C/16T chips with 40+ PCIe lanes.

You get to choose. Get the PCIe lanes (and pay a pretty penny for cores most people don't need) or go for max per thread performance and get anemic numbers of PCIe lanes.
 
Zarathustra[H] said:
And which would they be?

There are no chips with top level max clocks from either Intel or AMD that have OK numbers of PCIe lanes.

Intel has low PCIe lane counts on all of their consumer (lga1151) chips. They also limit their numbers of PCIe lanes on their lower core count LGA2066 chips and the higher core count versions which do have the full number of PCIe l have lower max boost clocks.

I don't understand why neither AMD not Intel have any top clocked 8C/16T chips with 40+ PCIe lanes.

You get to choose. Get the PCIe lanes (and pay a pretty penny for cores most people don't need) or go for max per thread performance and get anemic numbers of PCIe lanes.
Click to expand...
I wonder what you have plugged in that chews up all those PCIE lanes that isn't helped by having larger core counts. I need but loads of cores and large numbers of lanes but I don't need clock speeds nearly as much as the pure throughput volumes, not judging or ciricising just curious.
 
Lakados said:
I need but loads of cores and large numbers of lanes but I don't need clock speeds nearly as much as the pure throughput volumes
Click to expand...

To be clear -- you don't need more performance, you just need more cores?

Is this a licensing thing?
 
Zarathustra[H] said:
They also limit their numbers of PCIe lanes on their lower core count LGA2066 chips and the higher core count versions which do have the full number of PCIe l have lower max boost clocks.
Click to expand...

Only on the 7000 series. 9000 and 10000 series all have max lanes on all sku's (44 for 9k series, 48 for 10k)

Plus you can always overclock or adjust the max turbo multipliers on them as they are all unlocked..
 
IdiotInCharge said:
To be clear -- you don't need more performance, you just need more cores?

Is this a licensing thing?
Click to expand...
More a VM/RDP thing, 64 users running Excel/Outlook/Firefox don't need a lot of CPU for general office production work, but what really helps is the larger core counts and the larger PCIE lane support 64 + cores at 2.3 GHZ performes much smoother than 32 cores at 5.0, individual tasks may not be as fast but overall average performance for all the connected users is much smoother with less latency. Add in some Tesla GPU's for things that require the vGPU access, give it a solid amount of ram 512 GB in this machines case, and users are happy and maintenance costs are down.
 
Lakados said:
but what really helps is the larger core counts and the larger PCIE lane support 64 + cores at 2.3 GHZ performes much smoother than 32 cores at 5.0
Click to expand...

So, theoretically, this shouldn't be the case. Seeing as that's what you're... seeing, I wonder if the delta is related to other architectural differences, like having more cache and / or memory bandwidth per core, etc.
 
IdiotInCharge said:
So, theoretically, this shouldn't be the case. Seeing as that's what you're... seeing, I wonder if the delta is related to other architectural differences, like having more cache and / or memory bandwidth per core, etc.
Click to expand...
I am not totally sure, but the overall experiance with the higher core counts is much much smoother, like I said if one user puts together one of their disgusting Dynamics reports and compiles a 120MB Excel file on the EPYC it does take longer not quite twice as long but close too, so lets say 5 min instead of 3. But other users don't notice the slow down when they do run that report they are none the wiser, so I suspect it has something to do with underlieing task scheduling as much as the other factors.

EDIT:
I would like to point out that we have noticed that for our use cases that this holds true for the Intel based servers as well core count > clock speed (within reason)
 
Last edited:
Smoked Brisket said:
I think 750 is too much of a jump from the 500 3900x, so at 850 plus, the 3950 is too expensive. Just my opinion. There is a strong possibility that if AMD is recommending water cooling, and they are here, then that AIO is not going to be sufficient, at least to see temps that most would be happy with.
Click to expand...
Agreed I think it'll need decent water cooling to see that magic 4.7 but it's like the 2080Ti. It's the best at what it does, it's very expensive in terms of bang for buck, targeted for the high end users; buy It or get off the shitter.
 
Probably because more CPUs means less Context Switches.
Smoked Brisket said:
I think 750 is too much of a jump from the 500 3900x, so at 850 plus, the 3950 is too expensive. Just my opinion. There is a strong possibility that if AMD is recommending water cooling, and they are here, then that AIO is not going to be sufficient, at least to see temps that most would be happy with.
Click to expand...

From the article - AMD RECOMMENDS AIO at least so should be no problem.

"Available for purchase from November 25, 2019, the Ryzen 9 3950X is priced at USD $749 (MSRP). The retail PIB box package lacks a cooling solution, and AMD recommends at least a 240 mm x 140 mm AIO liquid CPU cooler to go with this chip."
 
It's a serious chip for work that requires it. If you don't get paid for your work then you'll have to decide whether or not it's for you. If you are paid for your work it will (hopefully) pay for itself.
 
Mega6 said:
Probably because more CPUs means less Context Switches.
Click to expand...

While this is true for some small percentage of workloads, everything else being equal, it's really inconsequential. There just aren't enough context switches to make a difference.
 
You must log in or register to reply here.
Back
Top