• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

Discussion Intel current and future Lakes & Rapids thread

Page 363 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Bouowmx said:
How many cores in one die?
Click to expand...

56 cores for the whole SKU according to NBC.

www.notebookcheck.net

First pictures leak of Intel's upcoming Sapphire Rapids Xeon CPUs, but 2022 release date could make it a tough sell in the face of Zen 4 EPYC Genoa

Chinese tipster @YuuKi_AnS released what appear to be the first leaked images of Intel's upcoming Sapphire Rapids Xeon server processor. Featuring up to 56 Golden Cove cores, Sapphire Rapids would be Intel's first server processor built on 10nm lithography.
www.notebookcheck.net www.notebookcheck.net
 
Regarding the die size, keep in mind the die covering the top SKUs will feature a massive amount of next-gen IO - HBM2 & DDR5/DDR4 memory controllers and PCIe 5 with CXL 1.1 functionality. Also the amount of required additional "MCM glue logic" is unknown at this point.

The current-gen 4 Willow Cove client cores including L3 are around 43mm2 on the Intel Super 10nm. 16 of them should be over 170mm2. That's excluding the whole IO, probably larger L3 and Golden Cove size increases.

Looking at the XCC Cascade Lake, the "uncore" occupies roughly 40% of the die size. So 16 cores seems a bit low. 20 of them would take a bit over 210mm2. That would fit that 418mm2 chiplet a bit better.
 
yuri69 said:
Regarding the die size, keep in mind the die covering the top SKUs will feature a massive amount of next-gen IO - HBM2 & DDR5/DDR4 memory controllers and PCIe 5 with CXL 1.1 functionality. Also the amount of required additional "MCM glue logic" is unknown at this point.

The current-gen 4 Willow Cove client cores including L3 are around 43mm2 on the Intel Super 10nm. 16 of them should be over 170mm2. That's excluding the whole IO, probably larger L3 and Golden Cove size increases.

Looking at the XCC Cascade Lake, the "uncore" occupies roughly 40% of the die size. So 16 cores seems a bit low. 20 of them would take a bit over 210mm2. That would fit that 418mm2 chiplet a bit better.
Click to expand...
There's no DDR4 IMCs but otherwise yes. Most of the die area for each of these chiplets are going to be filled with I/O.

Interested to see exactly how many cores are actually on each die though, because I don't think it's 15...
 
yuri69 said:
The current-gen 4 Willow Cove client cores including L3 are around 43mm2 on the Intel Super 10nm. 16 of them should be over 170mm2. That's excluding the whole IO, probably larger L3 and Golden Cove size increases.
Click to expand...

Saphhire is using Golden Cove, and surely that is going to be a lot bigger than a Willow Cove Core. Plus the second AVX-512 unit, the mesh logic, the server exclusive features, any additional cache... Skylake Server's tile is close to 18 mm2 so only 15 cores and the one IMC tile is not crazy.

I suppose that Saphhire is going to end up being a HPC only product.
 
uzzi38 said:
There's no DDR4 IMCs but otherwise yes. Most of the die area for each of these chiplets are going to be filled with I/O.
Click to expand...
My bad, DDR4 is for ADL. The server platform makes sense to be DDR5 only.

The HBM SKUs will surely be aimed towards HPC, although the conventional DDR ones might be fine for general purpose.
 
jpiniero said:
I suppose that Saphhire is going to end up being a HPC only product.
Click to expand...

Nah, it's not HPC only. The HBM2 ones might be for all I know, but SKUs that don't use HBM2 should be generally available at the very least on the high end of the market.
 
jpiniero said:

Gigabyte Technology Co., Ltd. Z490 AORUS MASTER - Geekbench

Benchmark results for a Gigabyte Technology Co., Ltd. Z490 AORUS MASTER with a 11th Gen Intel Core i9-11900K processor.
browser.geekbench.com browser.geekbench.com

New GB5 score of what looks like a retail 11900K. 1894 is really good for ST.
Click to expand...

Yeah, that would be a pretty big increase. A 10900k scores ~1400.

Intel Core i9-10900K Benchmarks - Geekbench

browser.geekbench.com browser.geekbench.com

That would be a 35% increase. That's... higher than my reasonable expectations, I guess. That would be sweet if true 🙂
 
Actually it broken the 1900 barrier: https://browser.geekbench.com/v5/cpu/6348421

The AES score is a small part, the IPC at 5300 Mhz is higher than on Tigerlake-U 4700 Mhz. This RKL-S sample uses DDR4-3600 and tighter latencies. Tigerlake-U runs with DDR4-3200 CL22 or LPDDR4-4266 CL36 and this might have run with DDR4-3600 CL15. Bandwidth isn't really important for ST workloads unlike latencies.
 
AES is about 6% of the final score. Keep that in mind for anyone trying to do IPC calculations from Geekbench.
How much that helps in other workloads you may guess based on Intel's claimed IPC gains being much lower.
 
mikk said:
Actually it broken the 1900 barrier: https://browser.geekbench.com/v5/cpu/6348421

The AES score is a small part, the IPC at 5300 Mhz is higher than on Tigerlake-U 4700 Mhz. This RKL-S sample uses DDR4-3600 and tighter latencies. Tigerlake-U runs with DDR4-3200 CL22 or LPDDR4-4266 CL36 and this might have run with DDR4-3600 CL15. Bandwidth isn't really important for ST workloads unlike latencies.
Click to expand...
Memory has little to no effect on GB5 at all. On my 5950X I saw no improvement in performance between JEDEC DDR4-2133 and DDR4-3200cl16.
 
Last edited:
uzzi38 said:
Memory has little to no effect on GB5 at all. On my 5950X I saw no improvement in performance between JEDEC DDR4-2133 and DDR4-3200cl16.
Click to expand...

It has quite a big effect on the AES score and unless Cypress Cove has a bigger IPC than Willow or Sunny Cove it seems to have an effect on RKL-S as well.
 
So by the end of this year or next one we'll have stock CPUs pushing 2000+ points. From both vendors probably.
Looking good, in spite of x86 decadence and limitations talks.
Sure power is another matter (we'll see how big little on desktop helps with that) but when it comes to max performance they aren't stopping.

jpiniero said:
I suppose that could work if they keep the tile structure in the core. 15 cores plus one IMC one. 16 tiles does seem low for 418 mm2 but I wouldn't say it's wrong either.
Click to expand...

I also keep hearing the 72 number for Sapphire. Maybe it's indeed 18 cores each tile and several disabled for yield, or there's two variants with different layouts...
If they can peek inside that naked chip we'll have a quick answer.

18 each tile would also work nicely as HEDT successor to the aged i9-10980XE. Same core count but huge generation gains and platform changes.
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top