• 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 250 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Isn't the 14nm Sunny Cove backport also somewhat reduced in cache sizes? Reducing the caches internal to the CPU would certainly reduce it's IPC as compared to a higher cache version on the smaller node that it was targeted at.
 
I wonder if Optane will show more relevance and performance being directly attached to the CPU instead of having to cross the chipset data link along with everything else? The access latency should be a bit quicker. It'll likely be PCIe 4.0 for the CPU link, so, if Intel bothers to update the PCIe controller on their Optane modules to support that speed, it'll have a higher transfer rate. We've got a pretty hefty amount of optane's in production at my job and they're almost completely seemless and do make a big difference for the machines that have them.
 
LightningZ71 said:
Isn't the 14nm Sunny Cove backport also somewhat reduced in cache sizes? Reducing the caches internal to the CPU would certainly reduce it's IPC as compared to a higher cache version on the smaller node that it was targeted at.
Click to expand...

Caches are the same. Same L1D of 48KB and L2 of 512KB. But cache size is just part of the picture. Like IntelUser2000 mentioned, backport process introduces additional challenges. For example L2 cache sizes might have grown in physical size and latency in clocks had to to be relaxed to meet those new constraints. Or 14nm thermals/clocking requirements forcing reduction of L2 ways from 8 to 4. Quite a few other ways to keep size the same and lower the perf -> TLB size and structure, sizes of OOP structures and queues.
The key is where that "tax" is being paid, for example Intel already "paid" for 48KB L1D in relaxing latency from "4" cycles to 5. It seems scary penalty but in fact isn't so, cause 4 cycles was best case anyway. L2 latency nerf would hurt everything, while limits of TLB coverage would be more subtle.
 
FriedMoose said:
Rocket Lake on Sisoftware:

https://twitter.com/x/status/1287655730170888193

IPC is 20% higher than Comet Lake. 729.65Mpix/s at 3.2 GHz vs 918.59Mpix/s at 4.85 GHz for the average 10700k:

ranker.sisoftware.co.uk

: SiSoftware Official Live Ranker

ranker.sisoftware.co.uk ranker.sisoftware.co.uk
Click to expand...
I know. The moment GeekBench shows an abysmal IPC increase, of all the benchmarks out there, somehow SiSoftware comes to the save with a 20% figure. Never you mind that it's somewhat silly to compare IPC at a 50% difference in clock rates.
 
lobz said:
I know. The moment GeekBench shows an abysmal IPC increase, of all the benchmarks out there, somehow SiSoftware comes to the save with a 20% figure. Never you mind that it's somewhat silly to compare IPC at a 50% difference in clock rates.
Click to expand...

Sisoftware reports only base clock - if there's a turbo-boost in use there's no way to know actual frequency.
 
LightningZ71 said:
I wonder if Optane will show more relevance and performance being directly attached to the CPU instead of having to cross the chipset data link along with everything else?
Click to expand...

It's been pointed out but its nothing new. The graph shows its an existing feature with predecessors.

Sites also did testing with CPU attached Optane SSDs. It did make a difference but not huge.
 
I tested a 32GB Optane drive on the 4x PCIe port on my son’s 2700x a while back. We put the temp directory and the swap file on that drive and then stressed the system memory wise. It wasn’t... awful, but its still not a replacement for RAM in any way.

I’ve currently got it configured in a tiered configuration with a raid 0 array of hard drives using storage spaces. It’s fairly effective.
 
Two new Ice Lake-SP benches in the "usual" 2 * 28c config.

empty empty - Geekbench

Benchmark results for an empty empty with a Genuine Intel(R) CPU $0000@ processor.
browser.geekbench.com browser.geekbench.com

empty empty - Geekbench

Benchmark results for an empty empty with a Genuine Intel(R) CPU $0000@ processor.
browser.geekbench.com browser.geekbench.com

7742P:

Generic - Geekbench

Benchmark results for a Generic with an AMD EPYC 7742 processor.
browser.geekbench.com browser.geekbench.com

These two compared:

Generic vs empty empty - Geekbench

browser.geekbench.com browser.geekbench.com

Really weird results - the frequency difference doesn't match the ST score diff and 2P Xeon strangely scaling way better in MT
 
LightningZ71 said:
I tested a 32GB Optane drive on the 4x PCIe port on my son’s 2700x a while back. We put the temp directory and the swap file on that drive and then stressed the system memory wise. It wasn’t... awful, but its still not a replacement for RAM in any way.

I’ve currently got it configured in a tiered configuration with a raid 0 array of hard drives using storage spaces. It’s fairly effective.
Click to expand...
Optane has several niche uses:
  1. There is no available DDR memory module at the capacity you need. Optane is slower than memory, but it is faster than not having enough memory when it is physically impossible to buy enough memory.
  2. You can't afford the DDR memory modules at the capacity you need.
  3. You have a cheap HDD that you want to make a few programs speedy.
  4. You have a good SSD, and you put Windows on your Optane drive.
If you don't belong to one of those four categories, then Optane probably won't help you much at all.
 
yuri69 said:
Two new Ice Lake-SP benches in the "usual" 2 * 28c config.

empty empty - Geekbench

Benchmark results for an empty empty with a Genuine Intel(R) CPU $0000@ processor.
browser.geekbench.com browser.geekbench.com

empty empty - Geekbench

Benchmark results for an empty empty with a Genuine Intel(R) CPU $0000@ processor.
browser.geekbench.com browser.geekbench.com

7742P:

Generic - Geekbench

Benchmark results for a Generic with an AMD EPYC 7742 processor.
browser.geekbench.com browser.geekbench.com

These two compared:

Generic vs empty empty - Geekbench

browser.geekbench.com browser.geekbench.com

Really weird results - the frequency difference doesn't match the ST score diff and 2P Xeon strangely scaling way better in MT
Click to expand...

1x 225W Epyc vs... 2x 270w Ice Lake? :smile:
 
Look at the difference in the effective ST vs. MT memory bandwidth numbers. The Ice Lake chip has double the memory bandwidth in MT, and a few of its MT wins lean heavily on memory bandwidth. The double size L2 certainly doesn’t hurt either.
 
DrMrLordX said:
I think the 38c parts are 270W. Those are probably 180W-225W depending on actual clockspeed.
Click to expand...
Even so, about 400 watt barely beating 225 watt ? And not in everything ? I wonder what a 7V12 (I think thats it) would do. 240 watt 2.45 ghz. $4200 or best offer in ebay.

www.ebay.com

AMD EPYC Rome 7V12 64-core 2.45GHz 240W (better than 7742) | eBay

240W TDP. Usable in single- and dual-socket config. - 3.3GHz boost. - 128 PCI-e 4.0 lanes.
www.ebay.com

Or wait, even better, how about 2 sockets vs 2 sockets ? Dual 7V12's ! It would be a massacre,.
 
Last edited:
LightningZ71 said:
Look at the difference in the effective ST vs. MT memory bandwidth numbers. The Ice Lake chip has double the memory bandwidth in MT, and a few of its MT wins lean heavily on memory bandwidth. The double size L2 certainly doesn’t hurt either.
Click to expand...

Unfortunately (unlike Geekbench 5) the quality of the Geekbench 4 MT results on windows are so terrible that they tell us next to nothing.

The best EPYC 7742 (64 core) results on Linux are over 60 000 points in the MT benchmark. On Windows It's all over the place but generally between 10 000 - 20 000 points.

As are the lower end models:
32 core EPYC 7502 can get 18 491 points.
24 core EPYC 7402 can get 20 683 points
Hell even the 16 core EPYC 7302 can get 14 202 points.

See where I'm gettin at? While on linux Geekbench 4 scales decently, Windows results are next to useless with CPUs that have more than 16 cores. So I wouldn't read too much into them.

I really wish we'd get Linux, or better yet, Geekbench 5 results.
 
Markfw said:
Even so, about 400 watt barely beating 225 watt ? And not in everything ? I wonder what a 7V12 (I think thats it) would do. 240 watt 2.45 ghz. $4200 or best offer in ebay.
Click to expand...

I'm not saying it's great for the IceLake-SP setup. In terms of perf/watt I don't think it's looking that good. Makes me wonder about the clocks and about how far out of the ideal clockspeed/voltage curve those chips are. IceLake-U left a lot to be desired in how much power it consumed at a given clockspeed, but IceLake-SP may actually be worse.

@Gideon

GB5 would be better, but a full review would be better still. Too bad IceLake-SP is still not ready for reviews!
 
LightningZ71 said:
Look at the difference in the effective ST vs. MT memory bandwidth numbers. The Ice Lake chip has double the memory bandwidth in MT, and a few of its MT wins lean heavily on memory bandwidth. The double size L2 certainly doesn’t hurt either.
Click to expand...
Man, that's because it's a dual-socket system, so the maximum theoretical memory bandwidth should be doubled in the MT test where both CPUs are fully utilized.
 
Well, yes, of course. I was merely trying to grasp at straws about why the ST B's MT results were showing differences that we didn't immediately expect.
 
While Charlie is obviously Charlie (buckets of salt required), if this turns out to be even remotely true, oh boy ....
semiaccurate.com

Intel should not launch Ice Lake-SP

Last week we told you about the clocks for Ice Lake-SP, now lets take a deeper look at how well it will perform.
semiaccurate.com semiaccurate.com
Charlie Demerjian said:
OK lets get this out in the open right away, Intel’s Ice Lake-SP CPU is a dog, it loses at just about everything but a few benchmarks that heavily use it’s already deprecated AI instructions. For anything real world it is badly outclassed by both AMD’s and Intel’s own offerings, at times by multiples not percentages. Intel should simply not launch it, the damage it will do to their reputation, customer relations, and investor confidence far outweigh the meager marketing halo they will attempt to grab.
Click to expand...

The full article is behind paywall, but the excerpt explains enough IMO. Anyway, can't wait to see Milan vs Ice-Lake-SP reviews
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top