Monday, March 28th 2022
AMD's Upcoming Zen 4 Based Genoa CPUs Confirmed to Have 1 MB L2 Cache per Core
As unreliable as Geekbench can be as a comparative benchmark, it's also an excellent source for upcoming hardware leaks and in this case more details about AMD's upcoming Zen 4 based Genoa server and workstation processors has leaked. Someone with access to a 32-core engineering sample thought it was a good idea to run geekbench on it and upload the results. As the engineering sample CPU is locked at 1.2 GHz, the actual benchmark numbers aren't particularly interesting, but the one interesting titbit we get is that AMD has increased the L2 cache to 1 MB per core, or twice as much as its predecessor.
What seems to be missing from this engineering sample is any kind of 3D V-Cache, as it only has a total of 128 MB L3 cache. Despite the gimped clock speed, the Genoa CPU is close to an EPYC 7513 in the single core tests and that CPU has a 2.6 GHz base clock and a 3.65 GHz boost clock, both system running Ubuntu 20.04 LTS. It manages to beat it in a couple of the sub-tests, such as Navigation, SQLite, HTML5, gaussian blur and face detection and it's within a few points in things like speech recognition and rigid body physics. This is quite impressive considering the Genoa engineering sample is operating at less than half the clock speed, or possibly even at a third of the clock speed of the EPYC 7513. AMD is said to be launching its Zen 4 based Genoa CPUs later this year and models with up to 96 core and 192 threads, with 12-channel DDR5 memory and PCIe 5.0 support are expected.
Sources:
Geekbench, via VideoCardz
What seems to be missing from this engineering sample is any kind of 3D V-Cache, as it only has a total of 128 MB L3 cache. Despite the gimped clock speed, the Genoa CPU is close to an EPYC 7513 in the single core tests and that CPU has a 2.6 GHz base clock and a 3.65 GHz boost clock, both system running Ubuntu 20.04 LTS. It manages to beat it in a couple of the sub-tests, such as Navigation, SQLite, HTML5, gaussian blur and face detection and it's within a few points in things like speech recognition and rigid body physics. This is quite impressive considering the Genoa engineering sample is operating at less than half the clock speed, or possibly even at a third of the clock speed of the EPYC 7513. AMD is said to be launching its Zen 4 based Genoa CPUs later this year and models with up to 96 core and 192 threads, with 12-channel DDR5 memory and PCIe 5.0 support are expected.
30 Comments on AMD's Upcoming Zen 4 Based Genoa CPUs Confirmed to Have 1 MB L2 Cache per Core
But more adding cache into the die is not the cure all and to me it just feels like a bandaid approach to things.
One mb L2 really says something to its architecture, maybe we'll see more smaller L2 and Larger L3 caches. meh, what do I know. :rolleyes:
When I was putting together the news post, I was thinking it would be impossible to do any kind of sensible comparisons due to the low clocks, but apparently I was wrong.
AMD seems to have figured out some serious ways of improving the performance, although some of it might be down to the larger L2 cache.
The more AMD gains market share, the more money it makes to throw in R&D.
And the more AMD is having the upper hand, the more Intel will push itself to become again the number one company.
The thing with cache is you need to be able to use it. You want the data you put in it to be accessed sooner than it would have been in the next level of cache.
In this example, let say your L3 have 80 cycles of latency. If you preload frequently data into your L2 that are not accessed after 160 cycles, if would have been better into a L3, and the cache real estate would have been put to better use into ALU or other things like that.
To me, this larger cache do not increase IPC but it got larger due to the increased IPC.
the smaller cache would just have starved the executions ports.
As for 3D-Vcache, i think we will only see it on Zen 4 refresh and or top of the line SKU. It still very useful for certain load but most apps aren't made to utilize such a large L3.
Regarding Raphael the vibe I get from the leaks lead me to expect around up to +30% more performance in single thread and up to 38% for the high core count models (or more in 16c case) in multi thread performance (with the DDR5 uplift and if 170W TDP is true etc) while on the Intel platform I'm expecting 12900KS to be just -10% vs 13900KS in single thread and up to -25% in multi thread (16 E- cores vs 8)
That would make 13900KS competitive with 16core Raphael in both gaming and multi threading ups like Rendering.
For example in the academic 720p test which the +30% Raphael single thread performance uplift might manifest will give us the below results:
101,5%:16c Raphael(+30% vs 5900X)
100% : 13900KS
90% : 12900KS (-10% vs 13th gen)
86.4% : 12900K (just 4% slower vs KS)
78.1% : 5900X (highest 720p Vermeer)
tpucdn.com/review/intel-core-i9-12900k-alder-lake-12th-gen/images/relative-performance-games-1280-720.png
Like always in 4K the difference will be nonexistent (and the 4K experience will trickle down to lower segments if navi 31 is 2.5X vs navi 21...)
4 cycles on top of 80 cycles isn't that bad, 4 cycles on top of 12 isn't good at all.
Wanted to get a current gen platform for a while, ddr4+ryzen prices are pretty good (5700X incoming too) but ipc improvement from zen 4 is too substantial to ignore, especially considering it would come along with the feeling of a shiny new platform while also being more future proof with AM5. Oh, and pcie 5.0, which i mean it should be beneficial for ssd related stuff.
Im with a legendary 4790k, for 2160p it should be able to get the job done for a little while.
Something im still very undecided is what to do about gpu, im on desperate need of one due to unfortunate events.
So far my candidates are 6800XT and 3070ti
The thing is, i have many concerns:
-shit is overpriced, especially mid-low range (6700 xt, 3060ti, 6600 xt and such)
-amd is supposed to release a refresh soon...
-prices have started to drop...
-intel gpus coming relatively soon, which should further lower prices worst case scenario, best it outperformens 6800xt and 3080 with a lower price point
-well, if all that wasnt enough, the next gen is dropping this year and should be a massive improvement
So wtf should i do? I basically have no gpu atm, my morals completely refuse to even get a 6700 xt because i consider its way overpriced already at msrp. Yes im poor.
6800xt would be my choice if there was avaliability and/or priced adequately. 3070ti would be a perfect candidate atm if it was on msrp, currently 800eur where i live.
Im being this scrupulous because they are almost 2 year old gpu's so overpaying too much would make me feel even dumber when rdna3 drops.
I guess at the very least i have to wait for the AMD refresh (april right?), and probably get a 6850 xt or a 6070ti if the msrp is dropped (assuming there is avaliability which im relatively confident there will be). I have no hope for Intel to release anything before june and there is a chance the wait would been a waste of time forcing me to wait again but for rdna3 or it could be the other way around, im completely lost.
I am in need of hdmi 2.1 and impatient to play 4k :(
Ive been enduring for so so long, the last 3 years ive been with a partially working r9 280x that is almost completely dead now.
I actually had an Gigabyte rx 580 aorus that blew up the vrm (shortly after like 2.5 years from purchase). Garbage brand, only negative experiences with it, never again. My first ever motherboard was gigabyte (back in 2011 lga1155) that died literally like a week after 2 years from purchase, it was a p67a 100€ motherboard paired with an i3 so there is no excuse. My current motherboard is also gigabyte and has given me countless headaches, originally i had an MSI Gaming 5 Z97 that i had to swap for a Gigabyte G1 Z97 due to ram incompatibility issues, other than that the MSI board was great, to this day i regret doing the swap cause i could surely have figured something out with the ram.
sorry i had to rant about it
The rumors also state that this zen4 small core would be the little core of a big-little architecture in Zen 5.
I dont recall if anything have been confirmed (except the small zen4 core for Bergamo (epyc for up to 128 cores for cloud providers).