It's about freaking time. And we have ThreadRipper to thank for that.
It's about freaking time. And we have ThreadRipper to thank for that.Freaking 14C/16/18C options for desktop!
And more importantly above 12.000 usd. We are way past what is relevant. But who knows. Perhaps a new supply of high price hedt cpu will move demand for 2k cpu.So sorry but Xeon Platinum tops out at 28 cores.
28 Skylake cores are worth far more than 32 Zen cores, especially given how the latter is actually a 4-way NUMA system, with independent memory controllers on each node.So sorry but Xeon Platinum tops out at 28 cores.
AMD TDPs are imaginary anyway. Once you OC to 3.5 GHz, the power draw in the same workloads will probably be the same. In AVX workloads, obviously Intel will draw more power, but it will also do twice (or maybe 4x) the work.So at what base (MT) clock will 18C/36T part run? Can we assume 2.5Ghz? If AMD targets 1800X clocks for their top of the line Threadripper equivalent, the 18C/36T part running at 2.5Ghz base will have a tough time competing in non-AVX256 bit optimized MTed workloads, which is basically a 90+% of what we have in use today. The saving grace could be the advanced Turbo intel plans to use on it though, we will have to wait and see how that one works in practice.
That is not only a rumor, but also incomplete information as the Platinum 8180 replaces the E7 product line. The E5-2699Av4 has a MSRP of only $4500, so there will likely be a Xeon SKU in the $4000-5000 range with 24+ cores, perhaps up to 26 cores.And more importantly above 12.000 usd. We are way past what is relevant. But who knows. Perhaps a new supply of high price hedt cpu will move demand for 2k cpu.
These prices are disappointing but not unexpected. I said as much in other threads, but those expecting Zen to cause an Intel HEDT price crash are going to be disappointed. Since Intel does not actually sell any HEDT CPUs, they have no incentive to price competitively. It appears their response to AMD "ThreadRipper" is just to put out another SKU at the top end (at a price point nobody wants), so they can claim to have more cores for marketing/PR reasons.Lifted from overclockers.co.uk forum.
Thai Prices: https://gameolo.com/product-category/คอมประกอบ/cpu/cpu-intel?orderby=price-desc
INTEL i9-7920X - 12c/24t @ 2.90GHz (No turbo listed) - 160W.
L2 CACHE -12 x 256KB
L3 CACHE - 16.5MB
£1300 + VAT / $1671
INTEL i9-7900X - 10c/20t @ 3.30/4.50GHz - 160W
L2 CACHE - 10 x 256KB
L3 CACHE - 13.75MB
£855 + VAT / $1099
INTEL i9-7820X - 8c/16t @ 3.60/4.20GHz - 140w
L2 CACHE - 8 x 256KB
L3 CACHE - 11MB
£515 + VAT / $659
INTEL i9-7800X - 6c/12t @ 3.50/4.00GHz - 140W
L2 CACHE - 6 x 256KB
L3 CACHE - 8.25MB
£345 + VAT / $439
If that is the price then nobody sane will buy it. They will buy either 10/12C parts from intel or Threadripper.On the other hand, the much-rumored "competition" has apparently failed to affect pricing. I expect the 18C SKU to cost over $3000, given the $1700 price point of the 12C.
Maybe Intel will make bank with their larger-core-count HEDT CPUs. They should thank AMD for the introduction of Ryzen. I think that Ryzen has single-handedly caused enthusiasts (and even mainstreamers) to move to that platform, which is itself rather HEDT-like, being a direct competitor to BDW-E, Intel's HEDT platform. Meaning, I think AMD is growing the HEDT market with Ryzen, or at least getting open-minded Intel fans to think in terms of "more cores", more being 6-8, rather than the core-anemic i7-7700K.On the other hand, the much-rumored "competition" has apparently failed to affect pricing. I expect the 18C SKU to cost over $3000, given the $1700 price point of the 12C.
'Worth' is determined in part by cost.28 Skylake cores is worth far more than 32 Zen cores, especially given how the latter is actually a 4-way NUMA system, with independent memory controllers on each node.
So at what base (MT) clock will 18C/36T part run? Can we assume 2.5Ghz?
If AMD targets 1800X clocks for their top of the line Threadripper equivalent, the 18C/36T part running at 2.5Ghz base will have a tough time competing in non-AVX256 bit optimized MTed workloads, which is basically a 90+% of what we have in use today. The saving grace could be the advanced Turbo intel plans to use on it though, we will have to wait and see how that one works in practice.
I don't think Ryzen is 'stealing' a lot of market share. There has been a lot of pent up demand for more cores for a long time now. By people unwilling to pay the Intel 'premium' for anything more than 4 cores. Those are sales Intel wouldn't have gotten anyway. And at this point, they don't have to bring down prices. Just fiddle with the SKU's will be enough to keep most Intel fans happy.Maybe Intel will make bank with their larger-core-count HEDT CPUs. They should thank AMD for the introduction of Ryzen. I think that Ryzen has single-handedly caused enthusiasts (and even mainstreamers) to move to that platform, which is itself rather HEDT-like, being a direct competitor to BDW-E, Intel's HEDT platform. Meaning, I think AMD is growing the HEDT market with Ryzen, or at least getting open-minded Intel fans to think in terms of "more cores", more being 6-8, rather than the core-anemic i7-7700K.
Naples is certainly not going to improve memory latency (which is already poor on 1-node Zen) by introducing NUMA and a 75% probability of cross-node memory traffic in "1S" installations. Perhaps it will also have a 4-way NUMA programming model like Intel Cluster-On-Die, with all the performance scaling problems you would imagine from a 4S cluster trying to compete with a 1S CPU.'Worth' is determined in part by cost.
Ryzen engineering motherboards had 2:1 DF:Memclock in debug mode, and it is not known how Snowy Owl and Naples would have it configured. Broadwell-E already showed latency issues with 22+ cores, and my hunch is that the revised cache configuration in Skylake-X is aimed primarily at alleviating that problem. What it'll do in case of lower core counts is anybody's guess.
This statement is one I have seen countless times on these forums now. It is also one that makes zero sense, as NUMA scaling problems are physical problems and have nothing to do with emotions or "comfort." We already have seen how well Zen scales on even a single node, with all the threads about CCX performance catastrophes, and Naples will have four times as many of those node boundaries.I don't want to go nitpicking about the details regarding to what extent Naples is NUMA-like(the technical thread has tons of info about that), but historically AMD has been more comfortable with NUMA than Intel since the Opteron heydays.
Base clock speeds which are listed in the specifications have got nothing to do with AVX512. Full/half speed AVX512 doesn't affect the listed base and boost clocks because there are separate multipliers for AVX workloads since the v3 Xeons.I see you missed Xeon Gold specs. They have a 2.7 GHZ base 18C part with full speed AVX-512 enabled. If my source is correct it might be even higher for desktops with more limited AVX-512 support.
We will have to wait and see if AVX-512 is a feature of the XCC die or merely an e-fuse configuration. The lineup of Skylake-SP shows full AVX-512 only being enabled on XCC SKUs (which starts at 16C), but that could easily be a segmentation strategy instead of a technical detail. Chances are still very high that AVX-512 will be half speed on HEDT, and even if it were full speed, we would just get endless threads of "enthusiasts" bitching about it messing with their overclocking.I see you missed Xeon Gold specs. They have a 2.7 GHz base 18C part at 165W with full speed AVX-512 enabled. If my source is correct this monolithic 7980XE beast might be clocked even higher for desktops with more limited AVX-512 support. And then we have the >4 GHz Turbo and possible gains from the new cache structure improving performance per clock
FTFYAMD just lost any hope of having the performance lead on desktops. But look on the bright side, Threadrippercouldwill do well againstthe cheaper 10C-12Call Skylake-X variants.
Really amazing how Intel was able to keep this secret until (almost) Computex. Everyone I talked to just a few days ago reiterated the previous leak, even Intel in the presentation below.
8th Gen Intel Core 'Coffee Lake-S' + Z370 Chipset to Launch in August/September (4C and 6C 'K' Options)
These prices are disappointing but not unexpected.
Naples is hardly any more NUMA than Ryzen. There is a possibility that DF clocks may be enabled to run at twice the memory speed.Naples certainly isn't going to improve memory latency (which is already poor on 1-node Zen) by introducing NUMA and a 75% probability of cross-node memory traffic in "1S" installations. Perhaps it will also have a 4-way NUMA programming model like Intel Cluster-On-Die, with all the performance scaling problems you would imagine from a 4S cluster trying to compete with a 1S CPU.
Well for starters, Infinity Fabric means that in theory Naples can approach node-interleaved configuration more than NUMA, and node-interleaving had significant performance benefits on AMD with no effect on Intel in purely I/O bound scenarios.This statement is one I have seen countless times on these forums now. It is also one that makes zero sense, as NUMA scaling problems are physical problems and have nothing to do with emotions or "comfort."
Are you asking how four nodes are "hardly any more" NUMA than one node? Even if Naples has this amazing INFINITY interconnect with infinite bandwidth, it does nothing for latency, which will be made even worse by the fact that each node's L3 and memory controllers will be working independently. In fact, even within one node, the L3 isn't uniform! If 4-way NUMA scales so well, why are there hardly any 4S installations in the field?Naples is hardly any more NUMA than Ryzen. There is a possibility that DF clocks may be enabled to run at twice the memory speed.
Irrelevant stuff from last decade.
Such authority, much wow.Are you asking how four nodes are "hardly any more" NUMA than one node? Even if Naples has this amazing INFINITY tearsofjoy interconnect with infinite bandwidth, it does nothing for latency, which will be made even worse by the fact that each node's L3 and memory controllers will be working independently.
I don't wish to do the work for you, but when software correctly identifies 16MB L3 as 2x8MB, performance improves a lot. So it is a matter of optimization.In fact, even within one node, the L3 isn't even uniform! If 4-way NUMA scales so well, why are there hardly any 4S installations in the field?
So CFD which is purely I/O bound, the real test for NUMA latency issues, is irrelevant? Whatever floats your boat. Are you telling me that people no longer do fluid dynamics on CPUs?Irrelevant stuff from last decade.
So CFD which is purely I/O bound, the real test for NUMA latency issues, is irrelevant? Whatever floats your boat. Are you telling me that people no longer do fluid dynamics on CPUs?
This is precisely what an I/O bound workload looks like. CFD, or for that matter solving any differential equation on a grid is that - I/O bound.Actually it is you who has no clue. Even in that article it's written that it is FP load that has completely RANDOM memory access in certain sized chunks.
Even with suboptimal latency it is far better than having to do remote access. The Intel Xeon CPU had already got a more advanced protocol than the Opteron, hence it does not benefit from interleaving substantially.With Interleave, all 4 controllers get to work, even if suboptimal latency for all 4 CPUs it increases scaling from disastrous to okayish.
So please take your BS claims about "AMD benefiting from Interleaving" to some other forum, it is workload that is specific, not AMD.
65W and 95W as expected. Some people should wake up finally.
14C/16C/18C for SKL-X is a big surprise.
You should better wait for the official price list and for the sales start. Such pre-release listings are often much higher, especially when we talk about an unknown asian shop.
This is precisely what an I/O bound workload looks like. CFD, or for that matter solving any differential equation on a grid is that - I/O bound.
Even with suboptimal latency it is far better than having to do remote access. The Intel Xeon CPU had already got a more advanced protocol than the Opteron, hence it does not benefit from interleaving substantially.
Exciting times ahead.
Lifted from overclockers.co.uk forum.
Thai Prices: https://gameolo.com/product-category/คอมประกอบ/cpu/cpu-intel?orderby=price-desc
INTEL i9-7920X - 12c/24t @ 2.90GHz (No turbo listed) - 160W.
L2 CACHE -12 x 256KB
L3 CACHE - 16.5MB
£1300 + VAT / $1671
INTEL i9-7900X - 10c/20t @ 3.30/4.50GHz - 160W
L2 CACHE - 10 x 256KB
L3 CACHE - 13.75MB
£855 + VAT / $1099
INTEL i9-7820X - 8c/16t @ 3.60/4.20GHz - 140w
L2 CACHE - 8 x 256KB
L3 CACHE - 11MB
£515 + VAT / $659
INTEL i9-7800X - 6c/12t @ 3.50/4.00GHz - 140W
L2 CACHE - 6 x 256KB
L3 CACHE - 8.25MB
£345 + VAT / $439