Xeon W-3175X Review

CAD4466HK

CAD4466HK

2[H]4U
Joined
Jul 24, 2008
Messages
2,700
I wondered if this CPU "needed" a 1700W chiller to hit 5GHz...it just might.

At 4.3 GHz, we were hitting over 500W peak load (confirmed by wall meter), which is the limit of the cooling setup provided. Compared to the 4.0 GHz result, we calculated that the CPU actually used 17% more power overall to get a 7% increase in performance.
Click to expand...

We took some of our benchmark values for power and frequency, extrapolated them with a power curve, and we estimate that at 5.0 GHz, this chip is likely to be drawing in excess of 900W, perhaps as high as 1200W. Yes, Intel really did need that 1700W water chiller.
Click to expand...

https://www.anandtech.com/show/13748/the-intel-xeon-w-3175x-review-28-unlocked-cores-2999-usd
 
Damn, that is an impressive chip! $3000 is actually not that bad considering what this CPU can do. It beats the 2990wx in most tests and by a significant amount in some. This thing does not seem to compromise anywhere. Six channel memory is pretty cool too.
 
Steve at GN was able to hit 5.1Ghz using a ice chiller setup with a Mora 3 radiator and was pulling 800w just from the cpu probably would hit 900 easy in p95.. i haven't finished watching der8aur's LN2 overclock video on it yet but i think he did 5.8 or 5.9Ghz at something like 1100w.

even the 4.5Ghz overclock steve did using a 360 cooler it was pulling 600+.
 
Last edited:
On the Anandtech system, the motherboard actually seemed like the most overpriced item. $1500?? GTFO. Maybe ASUS thought the CPU was going to be far pricier so as to get their prices overlooked. The cooling was not that impressive either. With that giant case and CPU, something like a 420mm open loop would have been more capable, and at the very least, far more quiet.

It would be neat if someone launched a smaller 6 dimm slot version of this motherboard.

I find it kind if funny that G. skill is launching 6-channel memory kits. Is buying 3 dual channel not an option?
 
Nightfire said:
On the Anandtech system, the motherboard actually seemed like the most overpriced item. $1500?? GTFO. Maybe ASUS thought the CPU was going to be far pricier so as to get their prices overlooked. The cooling was not that impressive either. With that giant case and CPU, something like a 420mm open loop would have been more capable, and at the very least, far more quiet.

It would be neat if someone launched a smaller 6 dimm slot version of this motherboard.

I find it kind if funny that G. skill is launching 6-channel memory kits. Is buying 3 dual channel not an option?
Click to expand...

and that 1500 dollars is the price SI's are paying for the board, retail it would probably be closer to 2k dollars for the board if asus decides to sell it separately which they haven't said whether or not they will.
 
CAD4466HK said:
I wondered if this CPU "needed" a 1700W chiller to hit 5GHz...it just might.

At 4.3 GHz, we were hitting over 500W peak load (confirmed by wall meter), which is the limit of the cooling setup provided. Compared to the 4.0 GHz result, we calculated that the CPU actually used 17% more power overall to get a 7% increase in performance.
Click to expand...
We took some of our benchmark values for power and frequency, extrapolated them with a power curve, and we estimate that at 5.0 GHz, this chip is likely to be drawing in excess of 900W, perhaps as high as 1200W. Yes, Intel really did need that 1700W water chiller.
Click to expand...

https://www.anandtech.com/show/13748/the-intel-xeon-w-3175x-review-28-unlocked-cores-2999-usd
Click to expand...

500W (4.0/4.3)^2 = 433W

500W is a 16% more than 433W. So the measured 17% is rigth.

500W (5.0/4.3)^2 = 676W

Even assuming cubic scaling it is 786W. I don't get from where they got that 1200W estimation.
 
sirmonkey1985 said:
Steve at GN was able to hit 5.1Ghz using a ice chiller setup with a Mora 3 radiator and was pulling 800w just from the cpu probably would hit 900 easy in p95.. i haven't finished watching der8aur's LN2 overclock video on it yet but i think he did 5.8 or 5.9Ghz at something like 1100w.

even the 4.5Ghz overclock steve did using a 360 cooler it was pulling 600+.
Click to expand...

And even that was actually unstable. It did go through 3DMark Time Spy Extreme but Cinebench was an instant crash before it even started, IIRC they had to drop it to 4.8 to get through a single Cinebench test. Who knows how much power you need to pump through it to make it rock solid and actually stable 5Ghz overclocked system, if it is even possible.
 
So some reviews got power consumption wrong:

A comment about power from my W-3175X review. I've seen a couple of the reviews still using a 4x multiplier rather than the 2.25x multiplier, which erroneously shows the power >750W at 4.4 GHz. It should be nearer 500W.
Click to expand...

 
Nightfire said:
Damn, that is an impressive chip! $3000 is actually not that bad considering what this CPU can do. It beats the 2990wx in most tests and by a significant amount in some. This thing does not seem to compromise anywhere. Six channel memory is pretty cool too.
Click to expand...

Until Zen 2 cores come out this year and utterly smash intel in every meteric (potentially).

It is not a smart time to buy this $3000 cpu plus $1000 board unless you have a business model that can pay it off before Zen 2 drops. When the 3000 series Threadrippers go 7nm, especially with the projected IPC jump Intel will be sweating cannon balls. Their answer will be adding another + at the end of 14++++++++++++++
 
Nightfire said:
On the Anandtech system, the motherboard actually seemed like the most overpriced item. $1500?? GTFO. Maybe ASUS thought the CPU was going to be far pricier so as to get their prices overlooked. The cooling was not that impressive either. With that giant case and CPU, something like a 420mm open loop would have been more capable, and at the very least, far more quiet.

It would be neat if someone launched a smaller 6 dimm slot version of this motherboard.

I find it kind if funny that G. skill is launching 6-channel memory kits. Is buying 3 dual channel not an option?
Click to expand...

Probably high due to them only making a small batch of boards. It also mentions only 1500 processors being made.

"We were also told that even though there are only two motherboard manufacturers making motherboards, one of them only has plans to make a single run of 500 retail boards for OEMs, with the other expected to make up the deficit. The reason for this was simple: ‘Intel only ordered 500 from us’. These motherboards are expected to be ca $1500 apiece, but I still wonder if ASUS/GIGABYTE will break even designing these products."
 
tangoseal said:
Until Zen 2 cores come out this year and utterly smash intel in every meteric (potentially).

It is not a smart time to buy this $3000 cpu plus $1000 board unless you have a business model that can pay it off before Zen 2 drops. When the 3000 series Threadrippers go 7nm, especially with the projected IPC jump Intel will be sweating cannon balls. Their answer will be adding another + at the end of 14++++++++++++++
Click to expand...

(i)
RyZen and ThreadRipper are limited to dual-channel and quad-channel. So this Intel platform has the top on memory-bound unless AMD releases a new consumer platform around SP3 socket.

(ii)
This Intel platform support AVX512. So each core is capable of 64FLOP per cycle. Zen2 is a 32 FLOP per cycle muarch. So unless AMD releases a high-clocked 64C Threadripper this Xeon will continue being king on workloads that can take advantage of 512bit vectors

(iii)
AM4 is scheduled for Q3 or Q4. There is no schedule for TR4. There is not even any engineering sample known for TR4. How do you know 3000 series Threadrippers launch this year?

(iv)
What projected IPC? Even if Zen2 brings a ~15% IPC gain over 'Zen+', it will be only on pair with SKL-X cores in this Xeon.
 
Good points on all fronts. I did not see if the memory was overclockable for the W-3175x. Theoretically, it would take TR 4000 mhz in quad to match the 6 channel of the xeon. Latency would still likely be lower. In any case, I doubt most apps see a significant difference there, but it is still a plus for the Xeon.

Overall, the cpu is priced right. At $2000, it would have made even the 32 core TR look like a poor value. Priced closer to $4000 and most would just 'settle' for the TR for now and wait for the next big thing.
 
So 3K for the cpu and 1.5K for the MB? Maybe a dual cpu MB at that price.
 
Does anyone actually have one of these in the wild? I heard Intel isn't making very many. Seems more like a benchmark chip than something they want to sell.
 
Nightfire said:
Good points on all fronts. I did not see if the memory was overclockable for the W-3175x. Theoretically, it would take TR 4000 mhz in quad to match the 6 channel of the xeon. Latency would still likely be lower. In any case, I doubt most apps see a significant difference there, but it is still a plus for the Xeon.
Click to expand...

Memory is overcklocable. Tomshardware pushed up to 3200MHz
 
Nightfire said:
Damn, that is an impressive chip! $3000 is actually not that bad considering what this CPU can do. It beats the 2990wx in most tests and by a significant amount in some. This thing does not seem to compromise anywhere. Six channel memory is pretty cool too.
Click to expand...

It's not about competing with the 2990wx in all honesty.

The 2990wx mops the proverbial floor with the Xeon because of the price.

6 channel memory is a absolutely worthless. Is this thing going to be a dedicated transaction server for Amazon retail?

The AMD is half the price and it's only a few percent slower.

But yes you can pay double and get AVX 512 and for some I guess it's worth the premo.

Again Intel doesnt specify who this is for and they again over engineer and under deliver when price is concerned.

And at a minimum toss in the minimum of 1g for a motherboard and another 2 grand for RAM and your sitting on a machine that's 10% faster for hundreds of percents. Kre in cost.

And it's still friggin 14+++++++ process lol
 
The 2990WX is much slower in single-threaded applications - the 3175WX turbos aggressively, has a better internal layout, and has higher IPC. It's also not that more expensive - for an extra 50%, you get 10% higher multithreaded performance, 20% higher single-threaded performance, 50%+ better memory bandwidth, and likely 50%+ lower worst-case latency (since the 2990WX is missing memory channels). Unfortunately, the extreme price of the Dominus Extreme and the poor availability of the processor itself somewhat dampen the value proposition. The lack of choice also doesn't help - if the Dominus has some deal-breaking issue, you are stuck with a $2999 special-order processor and no board to put it in.
 
bwang said:
The 2990WX is much slower in single-threaded applications - the 3175WX turbos aggressively, has a better internal layout, and has higher IPC. It's also not that more expensive - for an extra 50%, you get 10% higher multithreaded performance, 20% higher single-threaded performance, 50%+ better memory bandwidth, and likely 50%+ lower worst-case latency (since the 2990WX is missing memory channels). Unfortunately, the extreme price of the Dominus Extreme and the poor availability of the processor itself somewhat dampen the value proposition. The lack of choice also doesn't help - if the Dominus has some deal-breaking issue, you are stuck with a $2999 special-order processor and no board to put it in.
Click to expand...
Pretty sure you do not buy either of these for single threaded apps?
Have you done a price comparison?
 
Hagrid said:
Pretty sure you do not buy either of these for single threaded apps?
Have you done a price comparison?
Click to expand...

You do still care about single-threaded performance, because most workstation workflows have many single-threaded steps. For example, engineering applications typically want one or two very fast cores to manipulate and rebuild the model, but potentially scale to an unlimited number of cores for simulation and validation (e.g. Solidworks for design, LS-DYNA for validation). Ditto for the creative fields - rendering an animation can happen on hundreds of thousands of cores, but compositing, VFX, and grading cannot.

The value of a powerful workstation is that it offers a well-balanced mix of single-threaded and multithreaded performance while maintaining the familiarity and ease of use of a single, local system. Most larger firms have caught on to the fact that the large scale simulation and rendering should be done remotely, and workstations should be relatively low core count, highly clocked systems (hence the proliferation of single-socket workstations and the relative demise of the dual socket ones), but a smaller shop or independent artist might not have the infrastructure available to deploy and maintain dozens of servers.

Now, 20% extra single-threaded throughput is probably not too noticeable, except when you are trying to hit a performance target. For example, in the visual effects industry, that extra 20% might allow the artist to achieve real-time playback, which is a huge deal and absolutely worth the extra $999 on what is probably a $5K+ workstation running $5K+ of pro software. It's less valuable for the engineering fields - 20% faster rebuilds aren't going to make or break anything, but on the other hand, CAD workstations typically sport $2K graphics cards and potentially tens of thousands of dollars of software, so the extra thousand dollars is not really an issue...
 
bwang said:
You do still care about single-threaded performance, because most workstation workflows have many single-threaded steps. For example, engineering applications typically want one or two very fast cores to manipulate and rebuild the model, but potentially scale to an unlimited number of cores for simulation and validation (e.g. Solidworks for design, LS-DYNA for validation). Ditto for the creative fields - rendering an animation can happen on hundreds of thousands of cores, but compositing, VFX, and grading cannot.

The value of a powerful workstation is that it offers a well-balanced mix of single-threaded and multithreaded performance while maintaining the familiarity and ease of use of a single, local system. Most larger firms have caught on to the fact that the large scale simulation and rendering should be done remotely, and workstations should be relatively low core count, highly clocked systems (hence the proliferation of single-socket workstations and the relative demise of the dual socket ones), but a smaller shop or independent artist might not have the infrastructure available to deploy and maintain dozens of servers.

Now, 20% extra single-threaded throughput is probably not too noticeable, except when you are trying to hit a performance target. For example, in the visual effects industry, that extra 20% might allow the artist to achieve real-time playback, which is a huge deal and absolutely worth the extra $999 on what is probably a $5K+ workstation running $5K+ of pro software. It's less valuable for the engineering fields - 20% faster rebuilds aren't going to make or break anything, but on the other hand, CAD workstations typically sport $2K graphics cards and potentially tens of thousands of dollars of software, so the extra thousand dollars is not really an issue...
Click to expand...
Yes, if you are using old software, then yeah, single threaded.
 
Hagrid said:
Yes, if you are using old software, then yeah, single threaded.
Click to expand...
Current versions of Solidworks and Fusion360 (CAD/CAM) use a single core for model loading and rebuilding. Most Adobe products (2D design, video editing) don't scale past 5-6 cores. Capture One (RAW development) doesn't scale well either. Davinci Resolve (grading, compositing) is usually GPU driven, but when it does touch the CPU, its only a few cores. I don't have numbers for Maya, but I think the interactive viewport is pretty poorly threaded as well.

The list goes on and on, and even for current versions of software, having poorly-threaded interactive functionality is the norm and not the exception.
 
There is just nothing this thing is needed for that justifies the target market and price sorry. Just doesnt. Intel just chunked together some server stuff, unlocked it, priced it retarded, and said oh lookie we beat AMD by a little bit. It's not innovation. AMD going chiplet was innovation, even if bot the first out there, and I bargain it will shine come Zen 2 architecture. Intel still on monolithic and by the time they get 12 or 10nm going they will be left in the dust. Thier super huge monolithic redunculous chip design has hit the roof. I dont look at this as a win but a last attempt to make giant monolithic power hungry overpriced complexity relevant. And it's relevant but this is the final generation.

My opinion and nothing more. To each thier own. Clearly this is not my kind of platform. So I speak with my own limited pov.
 
tangoseal said:
There is just nothing this thing is needed for that justifies the target market and price sorry. Just doesnt. Intel just chunked together some server stuff, unlocked it, priced it retarded, and said oh lookie we beat AMD by a little bit. It's not innovation. AMD going chiplet was innovation, even if bot the first out there, and I bargain it will shine come Zen 2 architecture. Intel still on monolithic and by the time they get 12 or 10nm going they will be left in the dust. Thier super huge monolithic redunculous chip design has hit the roof. I dont look at this as a win but a last attempt to make giant monolithic power hungry overpriced complexity relevant. And it's relevant but this is the final generation.

My opinion and nothing more. To each thier own. Clearly this is not my kind of platform. So I speak with my own limited pov.
Click to expand...

The merits of chiplets can be debated forever, but I don't see how the pricing is retarded - it usually matches or exceeds the performance of the 2990WX, while pulling far ahead in a few cases (single threaded, worst-case latency, memory bandwidth, AVX-512).
It's less of a good deal than Threadripper, for sure, but I wouldn't call the price unrealistically high. The value argument is a dangerous one on the high end; you wind up with arguments like "the 1950X has the performance of two 1700's and the price of four", which leads to fast consumer processors not existing.
 
For someone that can actually take advantage of this $3K for the CPU doesn't seem that bad, however the mobo cost is insane. If your workload can scale across many machines it would make way more sense just building multiple TR or Eypc boxes.
 
I wonder if it works in regular Xeon boards - you would lose the unlocked multiplier, but it would still be a Xeon with good Turbo for much less than the price of a 8180.
 
So we are discussing this ridiculous slab of a pointless over priced rebuttal to Threadripper. Does anyone here on the forums actually have or going to buy one?

Here is an awesome review from Newegg! I would totally give this guy 5 eggs for this review.

upload_2019-3-7_17-51-4.png
 
You must log in or register to reply here.
Back
Top