Mostly storage.
That 7007$ Intel quad core is a steal
I know someone who's likely to buy it. It's basically a specialist bin that extracts the absolute most memory bandwidth and cache of the huge die while only enabling 4 cores, meant for those poor bastards who license software by the core of the CPU on which it runs, and who would pay much more than $7000 extra for the software license if they moved to a CPU with more cores.
Excellent result, that have take me a bit by surprise. ( ofc theres the cores numbers who are in question ) but as you said, with thoses aggressive prices... I dont see much what could do intel outside bring their price down.
They can't play a price war, yields for those monster dies (the 28core ones) must be terrible in comparison to yields for 4 zeppelin dies. Hence, "It's 4 desktop cores re-purposed for the server space! This is not what you want!" in their official release slides That whole presentation reeks of desperation, almost half of it is dedicated to Epyc. Not to mention, they did cite wccftech (lol) and even left a session id in one of their links ( ). Surely not what I expected from Intel. But hey! Maybe it works with investors ¯\_(ツ)_/¯
That whole presentation reeks of desperation, almost half of it is dedicated to Epyc. Not to mention, they did cite wccftech (lol) and even left a session id in one of their links ( ). Surely not what I expected from Intel. But hey! Maybe it works with investors ¯\_(ツ)_/¯
Last edited:
They can't play a price war, yields for those monster dies (the 28core ones) must be terrible in comparison to yields for 4 zeppelin dies. Hence, "It's 4 desktop cores re-purposed for the server space! This is not what you want!" in their official release slides
I aggree that yields are surely all but stellar ( and price / wafer ).. This said, i doubt that they cant cut 20-25% of the actual prices ( 10K+ for 2P )...
here you have a 32 cores 2P Epyc who cost less than half of the 28 cores Skylake one. That is pretty disturbing when you think about it..
I really need to look at this official slide lol.
Last edited:
Not just 4 dies but the same die so AMD and GloFo didn't need to spend I+D perfecting several Die designs, they just needed to spend all their I+D into perfecting a single design and making it the best chip their I+D budged let them. The yields most be close to the theoretical perfect yield possible by the process.
Also knowing that the die is the same in every single SKU is not a desktop cpu re-purposed to server use its a server CPU re-purposed to desktop use.
I think in the future we will look at zen as a CPU ahead of its time even with its flaws, the concept of develop a single Die, using all ur resources into making it the best die you can create and then use it in ur entire line-up its something that I think we will see not just in the futures CPU but GPU as well. Its not perfect but no design is, looking at what zen is able to do(from low TDP laptops to high end servers) this concept has shown a incredibly good balance in its concessions.
Also knowing that the die is the same in every single SKU is not a desktop cpu re-purposed to server use its a server CPU re-purposed to desktop use.
I think in the future we will look at zen as a CPU ahead of its time even with its flaws, the concept of develop a single Die, using all ur resources into making it the best die you can create and then use it in ur entire line-up its something that I think we will see not just in the futures CPU but GPU as well. Its not perfect but no design is, looking at what zen is able to do(from low TDP laptops to high end servers) this concept has shown a incredibly good balance in its concessions.
Maybe it's not at all an economic gain to cut prices in these market segments. Discounts can oth mean "sign of weakness" for the corporate resposibles for purchases.
AMD is on a different position (0% marketshare) and lower prices are one of their few cards to play.
AMD is on a different position (0% marketshare) and lower prices are one of their few cards to play.
Anarchist4000
Veteran
Well, theres a little rule anyway, if you sell less ( a bit less ) increase margin or price will have the same outcome in profit. If your plan is to increase the number of product solds, ofc margin can be reduced and price lowered.
No doubt they are, since AMD does binning for Ryzen Pros too
Anarchist4000
Veteran
Not just binning, but having some extreme bins. A 99% or higher quality bin could be feasible. Just consider Epyc volumes as a fraction of all the Ryzens they are dumping to consumers. Normally the server products are their own chips and will lack that volume.
https://www.techpowerup.com/235092/...cessors-glued-together-in-official-slide-deck
When an official web use this kind of lenguage you know this is a thing.
Threadripper 1950X will be practically a dual socket Ryzen 1700X Pro. But without needing an expensive dual socket server motherboard and without needing special 2P processors (that tend to cost up to 2x more). And the latency across dies should be lower than traditional dual socket systems as well. It is a much better choice than dual socket 8-core Xeon.
We can of course compare this "glued-together" Threadripper with similar dual socket Xeon Gold 6134 (Skylake-SP, 8 cores, 3.2 GHz base, 3.7 GHz turbo). These Xeons cost $2214 each, making the total CPU cost = $4428. Threadripper 1950X will likely cost around $1000-$1200. Intel doesn't win the "glued-together" competition either. Also Xeon Gold 6134 has only 2 UPI links, so if you clue four of them together to form EPYC, you won't get direct connection between each CPU pair, increasing latency. You need to buy a Xeon Platinum to get full CPU<->CPU connections in 4P system. And then we are talking about CPUs that cost each more than the full 32 core EPYC (which is practically a 4P system in package).
Intel actually seems to have one aggressively priced Platinum Xeon model. The 16-core Xeon 8153 at $3115. Four of them = $12460. Compare to dual socket EPYC 7601 = 8400$. Both systems have 64 cores (128 threads). AMD has 200 MHz clock advantage, but Intel has IPC advantage to tie the comparison. In this comparison Intel is only 48% more expensive. If you compare this Intel setup to the existing dual Xeon 8176 ($8719, 28 core) benchmarks, both the performance and perf/$ should be better for Intel as their quad socket 16-core gives you 8 cores (16 threads) more (but a 100 MHz lower clock rate). Price for the CPUs is 29% lower. 4P motherboard obviously costs a bit more vs 2P.
We can of course compare this "glued-together" Threadripper with similar dual socket Xeon Gold 6134 (Skylake-SP, 8 cores, 3.2 GHz base, 3.7 GHz turbo). These Xeons cost $2214 each, making the total CPU cost = $4428. Threadripper 1950X will likely cost around $1000-$1200. Intel doesn't win the "glued-together" competition either. Also Xeon Gold 6134 has only 2 UPI links, so if you clue four of them together to form EPYC, you won't get direct connection between each CPU pair, increasing latency. You need to buy a Xeon Platinum to get full CPU<->CPU connections in 4P system. And then we are talking about CPUs that cost each more than the full 32 core EPYC (which is practically a 4P system in package).
Intel actually seems to have one aggressively priced Platinum Xeon model. The 16-core Xeon 8153 at $3115. Four of them = $12460. Compare to dual socket EPYC 7601 = 8400$. Both systems have 64 cores (128 threads). AMD has 200 MHz clock advantage, but Intel has IPC advantage to tie the comparison. In this comparison Intel is only 48% more expensive. If you compare this Intel setup to the existing dual Xeon 8176 ($8719, 28 core) benchmarks, both the performance and perf/$ should be better for Intel as their quad socket 16-core gives you 8 cores (16 threads) more (but a 100 MHz lower clock rate). Price for the CPUs is 29% lower. 4P motherboard obviously costs a bit more vs 2P.
Last edited:
I think Intel's point is they have an Epyc equivalent CPU on one die. Intel's on-die mesh allows much higher bi-sectional bandwidth between cores on the same die (and lower latency).
However AMD repurposing PCIe lanes for inter-CPU communication is quite clever, IMO. Intel segments this into 48 PCIe lanes and 3 (x16) QPI links which runs slightly faster (10.4GT/s vs 8GT/s).
It allows AMD to have roughly the same inter-CPU bandwidth as Intel in the two socket case, but ridiculous IO bandwidth in the single socket case, with the same silicon.
Cheers
However AMD repurposing PCIe lanes for inter-CPU communication is quite clever, IMO. Intel segments this into 48 PCIe lanes and 3 (x16) QPI links which runs slightly faster (10.4GT/s vs 8GT/s).
It allows AMD to have roughly the same inter-CPU bandwidth as Intel in the two socket case, but ridiculous IO bandwidth in the single socket case, with the same silicon.
Cheers
Similar threads
