I doubt it. Even the 18 Core Intel® Xeon® W-2195 has a MSRP of $2553.00. Why would a 28 Core cost 3 times as much? A MSRP of $4000 for the 28 Core sounds realistic to me.
8-core Core i7-9800X officially costs $599. Yet the 18-core Core i9-9980xe that operates on the same platform costs $1999 and not $1348 as it should if it were proportionately priced to the i7-9800X. Higher end means higher premium.
You've got a higher base clock speed (which I think all cores are guaranteed to maintain). Based on that you're talking 41.4Ghz-cores vs 86.8 - over double the raw power. You get it in one socket, albeit at almost twice the TDP. And in theory you can go even higher...
I imagine the workloads might be quite specialized, though! And because of that, potentially less volume, which again increases the price a bit.
It would eat into their 8176 and 8180 sales, especially since those require 2s+. I servers need 1s compute without the lanes, pricing this low would cannibalize that niche. They're not aiming this at consumers and they're sacrificing a premium XCC die, so they're not going to sell it for 4k when they could sell it as an 8176 or 8180 for 8k. Just sayin' TM.
So you're saying you would buy one of these @ $8k?
I'm curious what your workload is, that is specifically 1S only.
At $4-5k, I would have pushed to have this CPU added to our workstation spec at work. At $8k, even if we wanted to spend that kind of money, I don't think it would win out vs. comparably-priced systems with 2 regular, non-W Xeons.
Intel's prior press event was to align these chips up against the coming 32 core Threadrippers. With a price of $8000, Intel is flirting with what AMD could price their 32 core, high clocked Zen 2 based Epyc chips with a full 512 bit wide memory interface. *If* Zen 2 lives up to the promised IPC increases and clocks a notch higher than Zen+, AMD could be dethroning this chip in pretty much every metric that matters: core count, stock clock speeds, memory channel count, PCIe lane count, maximum socket support and cost. And that ignores the potentially of AMD serving other niche markets with higher clocked/lower core parts or higher cores per socket offerings (up to 64 cores).
The one advantage that Intel can leverage are the in-package options they offer: FPGA and higher performance fabric (Omnipath). From the picture in the article, the CPU socket supports the chips that have additional fabric, which is a nice touch.
Cascade Lake cannot come soon enough for Intel. With the delay of this part into 2019, it would not surprise me if the W3175X changed from Sky Lake-SP to Cascade Lake to leverage some process improvements for higher clocks.
One thing of the matters its that there are already single socket LGA 3647 motherboards in the server space: they just don't support a 255W unlocked CPU. I'm also torn seeing 2 DIMM per channel layouts with this chip. If anything, the platform would be focused on higher memory clocks with tighter timings which takes a hit when the second DIMM is placed into a system. Capacity is offset by supporting 8-rank, registered ECC DIMMs across six slots for those who want it alongside the unlocked CPU. Those that need really insane memory capacity are better off jumping to a dual/quad socket with 2 DIMM per channel configurations. Even the existing single socket boards are split between one and two DIMM per channel offerings based upon target performance and density.
AMD isn't going to charge that much. Neither does Intel really, for that matter. The W3175x - as a single-processor-only high-clocked 28 core CPU - would have made sense as a workstation chip around $4,000.
People who buy their $10,000 28-core server variant CPUs do so because they want maximum performance in servers that will fit in an ordinary rack. They buy a 4-socket server from Dell or HP, and get 108 cores, 216 threads, in a machine that will easily fit into their existing server rack.
If that's your workload/mindset, you can make a case for those $10k CPUs. What you can't make a case for is a single-socket-only 28-core CPU for $8k. If you have that kind of money to burn on workstation CPUs, you'll probably buy a machine with a pair of 24 core Xeons for a similar price instead.
Or, if it's about single-core performance rather than number of cores, then you wouldn't' have been looking at the W3175x anyway - a 9900K or whatever its Xeon equivalent is would have been a better fit for your needs.
I think this was a phantom product, never really intended to be something anyone would actually buy, intended to slow AMD down. Ironically, it may have had the opposite effect. People who would have gone ahead and bought systems with a pair of $2k Intel 18-core CPUs may have waited to see if the W3175x would be better. Now, with W3175x having turned out to be a marketing stunt, maybe if you can afford to wait a couple months' longer before replacing your last-gen machines, you can get something with Epyc 2 in it.
"This hardware is perfect for high frequency traders who make 6 or 7 figures a day on the market – they will happily drop 5 figures on a faster machine if it means they can make more money."
If they can benefit from so many cores at these lower frequencies then why not have some kind of fast networking use a lot of 5 GHz processors instead? I assume the high-frequency trading bottleneck is primarily latency. My understanding is that, more than any other market (other than possibly emulation), high-frequency traders needed high processor frequencies (combined, of course, with maximal IPC). Is the latency tradeoff from networking multiple 5 GHz systems so bad that a jumbo chip like this one is more effective?
It's a speculative claim to say that high-frequency traders will like it, based on the tiny venn-diagram intersection of "people with enough money to buy this" and "Intel's remaining advantages over AMD".
I'd like to make a more solid speculative claim: there is no market for this chip. It might find a niche being sold to people who want a "pet Xeon" and can deal without ECC support, and perhaps the odd enthusiast with too much money. Much like the supposed release for revenue of 10nm silicon, Intel are launching this so they can say they "kept" a promise, not because doing so makes any sense.
Also, even maxed out, the W3175x is not the ultimate CPU for max clock frequency. It would be the ultimate CPU for high clockspeed with 28 cores.
But what is the high-frequency trader using 28 cores for, where differences in clockspeed still matter, and getting a system with even more cores through multi-socket doesn't?
It's purely wishful thinking. HFT traders who need this level of extreme performance and ultra low latency are already moving on to implementing their algorithms onto FPGAs.
HFT is also very space limited. It's a lot of companies competing for the same connections that are near the exchanges. Many CPUs means many boxes or larger boxes.
Many CPUs maybe, but at least up to 4 CPUs it's not that hard to find reasonably-sized systems/motherboards.
If the W3175x is not single-CPU limited like the rest of the W series, then it might make sense. But if it is limited to 1S configurations, then $8k is way too much. For that money, I'd rather have a pair of slightly lower-clocked 24-core regular Xeons.
This makes me think it was never intended to be a real product - just a diversion. Even if it is based on a $10k server chip, they must know this is too much - by a factor of about 2x.
If you could dual-socket it, they might sell a few. But otherwise they'll sell almost none of them - probably by design.
Haha! $8K for a workstation chip. Good luck with that. I retired from software development. The last company I worked for was not perturbed about spending $5K for a dual Xeon workstation for developers, expected to last about 4 years. But the additional productivity (if any) from this $8K chip just won't justify the additional cost. This will be a very niche product. As others have suggested, maybe this is just a halo product, and Intel doesn't really expect to sell any.
No matter how high it actually gets priced for H1-2019 (if it is even released in H1), as soon as Threadripper comes along this summer, I have a feeling things will change.
"have learned that Intel's original pricing for the chip was around $8000. One of our sources is saying that the initial documents put the price up at that high, however discussions with distributors have confirmed that the on-shelf price is more likely to be around $4500"
Gee, they really seem to know their market... Seems totally amateurish. This is Intel?
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
27 Comments
Back to Article
rquared - Sunday, January 6, 2019 - link
I doubt it. Even the 18 Core Intel® Xeon® W-2195 has a MSRP of$2553.00. Why would a 28 Core cost 3 times as much? A MSRP of $4000 for the 28 Core sounds realistic to me.
ZoZo - Sunday, January 6, 2019 - link
8-core Core i7-9800X officially costs $599.Yet the 18-core Core i9-9980xe that operates on the same platform costs $1999 and not $1348 as it should if it were proportionately priced to the i7-9800X.
Higher end means higher premium.
GreenReaper - Sunday, January 6, 2019 - link
You've got a higher base clock speed (which I think all cores are guaranteed to maintain). Based on that you're talking 41.4Ghz-cores vs 86.8 - over double the raw power. You get it in one socket, albeit at almost twice the TDP. And in theory you can go even higher...I imagine the workloads might be quite specialized, though! And because of that, potentially less volume, which again increases the price a bit.
KOneJ - Wednesday, January 9, 2019 - link
It would eat into their 8176 and 8180 sales, especially since those require 2s+. I servers need 1s compute without the lanes, pricing this low would cannibalize that niche. They're not aiming this at consumers and they're sacrificing a premium XCC die, so they're not going to sell it for 4k when they could sell it as an 8176 or 8180 for 8k. Just sayin' TM.twtech - Thursday, January 10, 2019 - link
So you're saying you would buy one of these @ $8k?I'm curious what your workload is, that is specifically 1S only.
At $4-5k, I would have pushed to have this CPU added to our workstation spec at work. At $8k, even if we wanted to spend that kind of money, I don't think it would win out vs. comparably-priced systems with 2 regular, non-W Xeons.
HurleyBird - Sunday, January 6, 2019 - link
Good luck with that, IntelMikewind Dale - Sunday, January 6, 2019 - link
"and I can very easily see Intel making the argument that they’re not competing against 32-core Threadripper"Let's get us some benchmarks!
Kevin G - Sunday, January 6, 2019 - link
Intel's prior press event was to align these chips up against the coming 32 core Threadrippers. With a price of $8000, Intel is flirting with what AMD could price their 32 core, high clocked Zen 2 based Epyc chips with a full 512 bit wide memory interface. *If* Zen 2 lives up to the promised IPC increases and clocks a notch higher than Zen+, AMD could be dethroning this chip in pretty much every metric that matters: core count, stock clock speeds, memory channel count, PCIe lane count, maximum socket support and cost. And that ignores the potentially of AMD serving other niche markets with higher clocked/lower core parts or higher cores per socket offerings (up to 64 cores).The one advantage that Intel can leverage are the in-package options they offer: FPGA and higher performance fabric (Omnipath). From the picture in the article, the CPU socket supports the chips that have additional fabric, which is a nice touch.
Cascade Lake cannot come soon enough for Intel. With the delay of this part into 2019, it would not surprise me if the W3175X changed from Sky Lake-SP to Cascade Lake to leverage some process improvements for higher clocks.
One thing of the matters its that there are already single socket LGA 3647 motherboards in the server space: they just don't support a 255W unlocked CPU. I'm also torn seeing 2 DIMM per channel layouts with this chip. If anything, the platform would be focused on higher memory clocks with tighter timings which takes a hit when the second DIMM is placed into a system. Capacity is offset by supporting 8-rank, registered ECC DIMMs across six slots for those who want it alongside the unlocked CPU. Those that need really insane memory capacity are better off jumping to a dual/quad socket with 2 DIMM per channel configurations. Even the existing single socket boards are split between one and two DIMM per channel offerings based upon target performance and density.
twtech - Wednesday, January 9, 2019 - link
AMD isn't going to charge that much. Neither does Intel really, for that matter. The W3175x - as a single-processor-only high-clocked 28 core CPU - would have made sense as a workstation chip around $4,000.People who buy their $10,000 28-core server variant CPUs do so because they want maximum performance in servers that will fit in an ordinary rack. They buy a 4-socket server from Dell or HP, and get 108 cores, 216 threads, in a machine that will easily fit into their existing server rack.
If that's your workload/mindset, you can make a case for those $10k CPUs. What you can't make a case for is a single-socket-only 28-core CPU for $8k. If you have that kind of money to burn on workstation CPUs, you'll probably buy a machine with a pair of 24 core Xeons for a similar price instead.
Or, if it's about single-core performance rather than number of cores, then you wouldn't' have been looking at the W3175x anyway - a 9900K or whatever its Xeon equivalent is would have been a better fit for your needs.
I think this was a phantom product, never really intended to be something anyone would actually buy, intended to slow AMD down. Ironically, it may have had the opposite effect. People who would have gone ahead and bought systems with a pair of $2k Intel 18-core CPUs may have waited to see if the W3175x would be better. Now, with W3175x having turned out to be a marketing stunt, maybe if you can afford to wait a couple months' longer before replacing your last-gen machines, you can get something with Epyc 2 in it.
DanNeely - Sunday, January 6, 2019 - link
at this price the chilled water cooler Intel demoed it with @5ghz suddenly doesn't seem so unreasonable.Spunjji - Monday, January 7, 2019 - link
Hell, I'd expect it to come included :DOxford Guy - Sunday, January 6, 2019 - link
"This hardware is perfect for high frequency traders who make 6 or 7 figures a day on the market – they will happily drop 5 figures on a faster machine if it means they can make more money."If they can benefit from so many cores at these lower frequencies then why not have some kind of fast networking use a lot of 5 GHz processors instead? I assume the high-frequency trading bottleneck is primarily latency. My understanding is that, more than any other market (other than possibly emulation), high-frequency traders needed high processor frequencies (combined, of course, with maximal IPC). Is the latency tradeoff from networking multiple 5 GHz systems so bad that a jumbo chip like this one is more effective?
Spunjji - Monday, January 7, 2019 - link
It's a speculative claim to say that high-frequency traders will like it, based on the tiny venn-diagram intersection of "people with enough money to buy this" and "Intel's remaining advantages over AMD".I'd like to make a more solid speculative claim: there is no market for this chip. It might find a niche being sold to people who want a "pet Xeon" and can deal without ECC support, and perhaps the odd enthusiast with too much money. Much like the supposed release for revenue of 10nm silicon, Intel are launching this so they can say they "kept" a promise, not because doing so makes any sense.
twtech - Wednesday, January 9, 2019 - link
Also, even maxed out, the W3175x is not the ultimate CPU for max clock frequency. It would be the ultimate CPU for high clockspeed with 28 cores.But what is the high-frequency trader using 28 cores for, where differences in clockspeed still matter, and getting a system with even more cores through multi-socket doesn't?
r3loaded - Monday, January 7, 2019 - link
It's purely wishful thinking. HFT traders who need this level of extreme performance and ultra low latency are already moving on to implementing their algorithms onto FPGAs.Bp_968 - Monday, January 7, 2019 - link
Nope. Look up Infiniband. Its basically designed for super low latency, super high speed networking.jordanclock - Tuesday, January 8, 2019 - link
HFT is also very space limited. It's a lot of companies competing for the same connections that are near the exchanges. Many CPUs means many boxes or larger boxes.twtech - Thursday, January 10, 2019 - link
Many CPUs maybe, but at least up to 4 CPUs it's not that hard to find reasonably-sized systems/motherboards.If the W3175x is not single-CPU limited like the rest of the W series, then it might make sense. But if it is limited to 1S configurations, then $8k is way too much. For that money, I'd rather have a pair of slightly lower-clocked 24-core regular Xeons.
twtech - Wednesday, January 9, 2019 - link
How many of those are there, really? Maybe they'll sell 100 of these CPUs.twtech - Tuesday, January 8, 2019 - link
This makes me think it was never intended to be a real product - just a diversion. Even if it is based on a $10k server chip, they must know this is too much - by a factor of about 2x.If you could dual-socket it, they might sell a few. But otherwise they'll sell almost none of them - probably by design.
guyr - Friday, January 11, 2019 - link
Haha! $8K for a workstation chip. Good luck with that. I retired from software development. The last company I worked for was not perturbed about spending $5K for a dual Xeon workstation for developers, expected to last about 4 years. But the additional productivity (if any) from this $8K chip just won't justify the additional cost. This will be a very niche product. As others have suggested, maybe this is just a halo product, and Intel doesn't really expect to sell any.SharpEars - Friday, January 11, 2019 - link
No matter how high it actually gets priced for H1-2019 (if it is even released in H1), as soon as Threadripper comes along this summer, I have a feeling things will change.Gadgety - Thursday, January 24, 2019 - link
"have learned that Intel's original pricing for the chip was around $8000. One of our sources is saying that the initial documents put the price up at that high, however discussions with distributors have confirmed that the on-shelf price is more likely to be around $4500"Gee, they really seem to know their market... Seems totally amateurish. This is Intel?