Info 64MB V-Cache on 5XXX Zen3 Average +15% in Games

[Kedas]

Well we know now how they will bridge the long wait to Zen4 on AM5 Q4 2022.
Production start for V-cache is end this year so too early for Zen4 so this is certainly coming to AM4.
+15% Lisa said is "like an entire architectural generation"

 

[tomatosummit]

How do you balance 48GB of ram across 2 channels?

16+8GB sticks.
One of the wonders of having four ram slots is adding more later and while people saying having identical ram sticks is important, it really isn't.

 

[Insert_Nickname]

How do you balance 48GB of ram across 2 channels?

16+8GB per channel. 3 ranks across 2 DIMMs for 6 ranks total across 4 DIMMs.

 

[dr1337]

That was a strange post. I normally have close to 40-50 tabs across 2 browsers (Brave and Firefox)on a mint machine and use less than 25% of a 16 GB machine.

And here I am on windows 10 using 13gb on 32gb of ram with 5.5gb of that being chrome in 40 tabs. No games not even photoshop open, next heaviest process is oculus dash at 400mb. I know for a fact that 16gb of ram is plenty and windows adjusts allocation as necessary, but more ram isn't always a bad thing. Maybe 256gb is a bit of a joke, but depending on what you're doing I 100% believe there is logic to having more ram available specifically so that background tasks don't have to fight over it when you do want to game or 'shop a picture (as if there is a performance difference anyways lol). But on a dedicated web machine thats really never a concern.

Also I just want to say that no matter how I've done the clocks and timings on my e-die, I never see a real world performance difference. I've spent months at jedec and done every stable combo of freq and timings that are stable for my 3600, never once noticed a difference in the games I play or editing times. I see it as a meme for ryzen, its fun to tweak, tune, and bench, sure. But I really do not know who is actually bottlenecked by ram in anything other than in very specific bandwidth sensitive workloads/benchmarks.

 

[Markfw]

Getting 256GB because your browser leaks memory is crazy! Besides all that will do is delay the inevitable, you will still run out of memory it will just take longer.

At various times I've had several hundred tabs open in Firefox, and I've got "only" 16 GB on my Linux PC. Quick count right now it has 58 tabs open, and has been running since March 3. Not having any memory issues, though there have been times in the past when it slowed down and restarting Firefox fixed it - i.e. they've had memory leaks but have found/fixed them, then sometime later a new memory leak is introduced which is found/fixed in a later patch.

Not sure what you're doing that uses up memory so badly, maybe one of the sites you have running does something that leaks memory at a very high rate. If you could figure out which site that is you would file a bug report and maybe they'll fix it. Wouldn't that be worth it considering how much money you'd save not having to find a laptop that lets you stuff in 256GB?

Well, I don't know about others, I have 20-30 going. One for each forum I am moderating, one for each of 4 things like facebook, weather, bank, amazon, etc...

 

[eek2121]

I doubt there will be much OC headroom for this. First of all, Zen 3 didn't OC well to begin with. This chip has some headroom since the clocks are lower, but AMD has limited the voltage more than with regular Zen 3 so you'll be limited. Really though just getting the top single core boost a bit higher will have a big payoff since it's limited to gaming.

Zen 4 had a few ways to attack Zen 3D to either catch up or even pull ahead. We already know it will have higher clock speeds and fastest single thread still drives performance in a lot of games. There will also be IPC improvements and even if they're small they get multiplied by the faster clocks to create a compound effect. I believe the L2 cache is also a bit larger which isn't quite the same as having a massive v-cache, but does help a bit. AMD themselves said Zen 3D was 10-15% better on average than a 5900X so it's not impossible for Zen 4 to hit those numbers. It might not always win since there are some outliers where Zen 3D puts up some really crazy numbers, but it also likely has some titles where it sees almost no gains.

Even if Zen 4 isn't quite as good, AMD will eventually launch Zen 4D so anyone specifically in the market for a dedicated gaming CPU will eventually have a top choice that doesn't lock them in to an old platform. If AMD sits on their hands Intel always has the ability to come out with something and steal the crown back.

Zen 3 ported to 5nm (with no other changes) would add at least 20%-40% performance uplift in terms of frequency, excluding possible limiting factors of N5 in terms of power/frequency.

Zen 4 is related to Zen 3. I wouldn't expect anything less than 50% uplift. This places it far above what the 5800X3D is capable of.

Remember that Zen 2 -> Zen 3 was on the same node and it was 30+% faster. Zen 4 is on a new node and is very likely going to push things further.

Well, I don't know about others, I have 20-30 going. One for each forum I am moderating, one for each of 4 things like facebook, weather, bank, amazon, etc...

I typically have 20-30 open as well. The browser uses around 1.5gb of RAM. Microsoft Edge also puts inactive tabs to sleep if you let it.

Side note: Don't disable swap. Disabling swap will actually HURT performance. Sometimes memory pages are swapped to disk because they are infrequently used. That memory is then reclaimed and used for other apps, or even just caching filesystem reads/writes.

 

[moinmoin]

Browsers should logically put the inactive tabs to sleep and save their data to disk.

Opera is pretty aggressive about that so it has become my first choice on slower computers with less memory.

 

[igor_kavinski]

I typically have 20-30 open as well. The browser uses around 1.5gb of RAM. Microsoft Edge also puts inactive tabs to sleep if you let it.

All sorts of websites (mostly tech. some health. some youtube tabs) on my Thinkpad. Should really look into Opera. At work, google searches lead to opening up tons of tabs and Edge reaches more than 4GB RAM. Even though that system has 16GB RAM and an 860 EVO SSD, sometimes it feels it is struggling under the load of those tabs. But it is somewhat infrequent. Haven't had that much thrashing on it recently. Could be they have improved Edge in the last update?

 

[Thibsie]

Considering we are talking about a Zen 3 part. I think it's fair to say that most with Zen 2 or Zen 3 CPU's run their memory at 3600mhz or higher. AMD said the sweet spot for Zen 2 memory was 3733mhz but the maximum memory speed without using memory dividers 1:1 ratios was 3800mhz. At the time, 3733mhz 3800mhz was not as stable with current motherboards. It's not the infinity fabric within the memory controller that is the problem. It's the motherboard. Most motherboards B450 and newer can run 3800mhz stable without any issues.

I think there are a good number of people who paid a kings ransom for old B-die sticks that can run 14-14-14 3200mhz all day but cannot run 3600-3800mhz at all or with significantly loosened timings. The newest B-die sticks can run 14-16-16 3800mhz all day.
I am thinking those who paid for the old B-die don't want to pony up the cash for the newer B-Die and stick with 3200mhz because they can't let go of their old B-die sticks.

Why the Heck would I change my 2x16GB 3200 to 3600-3800 RAM for the money asked and the little performance upgrade, having a 5600x and a 1060gtx.
There's really more urgent upgrade to apply IMO.

 

[igor_kavinski]

Memory bandwidth becomes almost a non-issue with the 5800X3D anyway.

 

[Mopetar]

I don't do anything critical. Just like to keep tons of tabs open. 256GB just seems like the amount I think I would have a better chance of not maxing out. Either that, or I need to get a really expensive Optane drive, so when the paging happens, it doesn't cause the system to crawl to a halt.

I think you're overlooking the easier solution to STOP ACTING LIKE AN ANIMAL AND CLOSE YOUR GAWSH-DERN TABS! I'll bet you have want to buy 256 dinner plates too so you never have to wash them! Next you'll be wanting a new pair of socks every day as well.

Kidding aside why not just write a script to monitor memory use and restart your browser when it gets too high? If you've got that many tabs open it'll take you longer to find anything than it would to just open a new one. If you have that many tabs a bunch are probably the same webpage for that very reason.

Zen 3 ported to 5nm (with no other changes) would add at least 20%-40% performance uplift in terms of frequency, excluding possible limiting factors of N5 in terms of power/frequency.

The fastest Zen 3 CPUs hit 4.9 GHz single core on boost out of the box. 20% of that would mean just a shade under 5.9 GHz. Maybe a golden sample with exotic cooling could hit that, but it's not happening for most chips. 40% would be ~6.85 GHz which is ridiculous.

Even if I use a more charitable interpretation and go with the base clock, your performance range still pushes it over 5 GHz for most chips. We might well see some top-end chips be able to hit that with their all-core turbo particularly with the higher 170W TDP, but a 20% frequency bump is optimistic. AMD doesn't even need 10% to get a CPU that could hit 5.3 GHz.

 

[Doug S]

That was a strange post. I normally have close to 40-50 tabs across 2 browsers (Brave and Firefox)on a mint machine and use less than 25% of a 16 GB machine.

???

I wasn't suggesting that Firefox uses all 16 GB, I have other stuff running as well. I also have a second Firefox profile I use occasionally for a construction project which has a couple hundred tabs open, so at times I'm running two independent instances of Firefox. I close that when I'm done with it though so it only runs for the day, rather than a month or two like my day to day instance. I use a separate profile for it because a lot of the sites I connect to require logins and/or are vendor sites that don't work properly with ublock. So rather than trying to figure out the whitelists for ublock and cookie autodelete I simply created a separate profile to run a Firefox instance without those extensions installed.

At least half the tabs are static like pdfs or spec sheets, I have them organized between windows so it is easier to keep them straight than if I bookmarked everything. Plus I've found bookmarks don't always work deep in some sites - it'll kick you back out to the login screen whereas a tab won't. Not sure why that is, but for those who are wondering why people keep lots of tabs open I'm sure that's one reason.

 

[LightningZ71]

I don't do anything critical. Just like to keep tons of tabs open. 256GB just seems like the amount I think I would have a better chance of not maxing out. Either that, or I need to get a really expensive Optane drive, so when the paging happens, it doesn't cause the system to crawl to a halt.

32 GB first gen optane drives are stupid cheap and plentiful, especially used system pulls. I have a system that has one of those four m.2 PCIe drive cards in it loaded with four of them. I configured them as a Raid 0 drive, then made the whole thing a swapfile. I've intentionally stress tested the system to overflow the system ram (4x16gb 3200) and, while it does slow down a little, the system is still plenty usable.

 

[LightningZ71]

Zen 3 ported to 5nm (with no other changes) would add at least 20%-40% performance uplift in terms of frequency, excluding possible limiting factors of N5 in terms of power/frequency.

Zen 4 is related to Zen 3. I wouldn't expect anything less than 50% uplift. This places it far above what the 5800X3D is capable of.

Remember that Zen 2 -> Zen 3 was on the same node and it was 30+% faster. Zen 4 is on a new node and is very likely going to push things further.



I typically have 20-30 open as well. The browser uses around 1.5gb of RAM. Microsoft Edge also puts inactive tabs to sleep if you let it.

Side note: Don't disable swap. Disabling swap will actually HURT performance. Sometimes memory pages are swapped to disk because they are infrequently used. That memory is then reclaimed and used for other apps, or even just caching filesystem reads/writes.

This right here. Don't disable swap in windows. If you have any open m.2 slots, grab a cheap optane drive, pop it in, and make it swap. It won't make a massive difference, but when you get in those situations where there's memory pressure, the high write speed of that Optane really comes into play.

 

[maddie]

???

I wasn't suggesting that Firefox uses all 16 GB, I have other stuff running as well. I also have a second Firefox profile I use occasionally for a construction project which has a couple hundred tabs open, so at times I'm running two independent instances of Firefox. I close that when I'm done with it though so it only runs for the day, rather than a month or two like my day to day instance. I use a separate profile for it because a lot of the sites I connect to require logins and/or are vendor sites that don't work properly with ublock. So rather than trying to figure out the whitelists for ublock and cookie autodelete I simply created a separate profile to run a Firefox instance without those extensions installed.

At least half the tabs are static like pdfs or spec sheets, I have them organized between windows so it is easier to keep them straight than if I bookmarked everything. Plus I've found bookmarks don't always work deep in some sites - it'll kick you back out to the login screen whereas a tab won't. Not sure why that is, but for those who are wondering why people keep lots of tabs open I'm sure that's one reason.

I was not referring to your post, but the one you were answering. That's why" I wrote that post instead of this post".

I phrased it badly.

 

[Mopetar]

32 GB first gen optane drives are stupid cheap and plentiful, especially used system pulls. I have a system that has one of those four m.2 PCIe drive cards in it loaded with four of them. I configured them as a Raid 0 drive, then made the whole thing a swapfile. I've intentionally stress tested the system to overflow the system ram (4x16gb 3200) and, while it does slow down a little, the system is still plenty usable.

I looked and a 1 TB Optane drive on Newegg is around $500. That's going to cost less than 256 GB of RAM. Even DDR4 is expensive if you're buying 64 GB sticks.

 

[Insert_Nickname]

I looked and a 1 TB Optane drive on Newegg is around $500. That's going to cost less than 256 GB of RAM. Even DDR4 is expensive if you're buying 64 GB sticks.

It's actually not a bad time to buy 64GBs worth of DDR4 right now. Price is actually below what I paid for 32GB a couple of years ago. 64GB DIMMs are probably going to be expensive for a while yet of course.

I was tempted to pick up a matching 64GB set for one of my 32GB kits. For 96GB total. Not that I need it, but 22 years ago my PC had 96MB RAM. Seems fitting.

 

[LightningZ71]

I looked and a 1 TB Optane drive on Newegg is around $500. That's going to cost less than 256 GB of RAM. Even DDR4 is expensive if you're buying 64 GB sticks.

You still run into a bit of an interface bottleneck on those big Optane drives. Late last year, I had a friend ask me for advice on what he could do with four used 1TB SATA SSDs that he had gotten used for really cheap. He had a board that supported the correct bufurcation for his PCIe x16 slot, so we got a pass through quad M.2 card, mounted 4 cheap 32GB optane drives on it, attached the SATA drives to the motherboard ports, then used storage spaces to build a tiered storage array with the optanes caching the SATA drives in a raid 0. Its super fast. Granted, he doesn't game, so loosing the x16 for storage was acceptable and his video card is hanging off a chipset slot.

It all really depends on your use case.

 

[JoeRambo]

Even if I use a more charitable interpretation and go with the base clock, your performance range still pushes it over 5 GHz for most chips. We might well see some top-end chips be able to hit that with their all-core turbo particularly with the higher 170W TDP, but a 20% frequency bump is optimistic. AMD doesn't even need 10% to get a CPU that could hit 5.3 GHz.

What if we go with all-core clock instead? Current all core workload clocks for AMD CPUs are anemic, maybe 3.9Ghz or so for 5950x. Does not take anything exotic to take that to 20% higher clock of 4.8Ghz? It's not something today's 5950x can't do with tuning and cooling for 300W for 16C, but it takes new process node and new arch to cut power usage to more palatable region.

 

[szrpx]

What if we go with all-core clock instead? Current all core workload clocks for AMD CPUs are anemic, maybe 3.9Ghz or so for 5950x. Does not take anything exotic to take that to 20% higher clock of 4.8Ghz? It's not something today's 5950x can't do with tuning and cooling for 300W for 16C, but it takes new process node and new arch to cut power usage to more palatable region.

Well, 5950x's @ stock hit EDC limits before power limits so that holds back the all core frequency there. Meanwhile the 5800x with the same power budget hits like 4.4 ~ 4.5GHz all core.

I could definitely see some 5 GHz all cores skus for Zen 4.

 

[Mopetar]

What if we go with all-core clock instead? Current all core workload clocks for AMD CPUs are anemic, maybe 3.9Ghz or so for 5950x. Does not take anything exotic to take that to 20% higher clock of 4.8Ghz? It's not something today's 5950x can't do with tuning and cooling for 300W for 16C, but it takes new process node and new arch to cut power usage to more palatable region.

I think the 3.9 GHz all core speed for the 5950X is due to the TDP of the chip more than anything else. The 5900X has a higher all-core turbo and both of the single chiplet CPUs can hit around 4.5 GHz all core. There's some data from an AT article that shows the 5800X, 5900X, and 5950X all hitting the same power levels when boosting, with the only difference being the clock speed that all of the cores hit.

The details we have on TSMC 5nm don't suggest that provides either enough of a frequency or a power uplift for a Zen 3 shrink to hit something like a 5 GHz all core turbo within the same 125 W TDP. The transistor density improves a lot, but it's a 15% frequency boost or a 20-30% power improvement. I'm not sure where the person who originally made the post got 20% - 40% improvement from, but it doesn't line up with other published figures. Some articles do talk about other possibilities for additional performance gain, but those parts are light on details or specifically indicate that they come at the expense of density.

If we use the 5N shrink for a 30% power reduction and the figures from the AT article, we'd only expect an all-core turbo of around 4.5 GHz for a 16-core chip. The 8-core chip could be pushed so that the all-core turbo is above 5 GHz, but I question the usefulness of that since games largely still rely on having a single fast thread as opposed to being able to boost all threads to some amount and anything that's massively parallel like Cinebench is going to do far better with twice the cores even if they're not as fast.

Even if we could get a magic 40% frequency boost, at that point the chip is probably outrunning the memory system and won't realize a 40% performance gain from those clocks. We've already seen how much Alder Lake benefits from DDR5 memory in some benchmarks, so it's likely that AMD is going to see a lot of the performance uplift from Zen 4 come from faster memory as well.

 

[LightningZ71]

One of the things that you have to keep in mind is that the frequency boost possible quoted in process improvements are the absolute maximum achievable clocks assuming sufficient power available AND proper thermal dissipation. If they just aim for the maximum power draw reduction while taking whatever frequency cap remains, for a product like the 5950x, it would significantly improve the "all core" frequencies while making very little improvements in the single core boost clocks. Why? As you nudge closer to the frequency maximum that the process can achieve, the power draw, and subsequent heat generation, increases dramatically, in an asymptotic fashion. So, just a 10% improvement in power draw alone would make a several hundred Mhz improvement in all-core frequency for the chip assuming that the heatsink and heat spreader configuration are held consistent.

To give a real world example, one need look no farther then Rembrandt. The N6 process node allows modest frequency, density and power improvements, depending on how you implement it. AMD appears to have chosen to improve density and power draw at the expense of maximum frequency. You can see this in the size of the actual chip not increasing greatly despite a significant increase in iGPU transistors. When you look at benchmarks, the single thread throughput barely moves the dial, but, multi-threaded throughput improves notably.

I was hoping that AMD would see fit to produce a 5900XT and 5950XT based on a straight N6 die shrink of the existing CCD. With no other changes, it is likely that they could have had products that had a significant improvement in MT throughput. I had originally hoped that they would try for stacked cache versions, but, with the significant reduction in All core clocks that would have brought, it's likely that there would be very few situations where it would really help.

 

[Hans Gruber]

I am curious. What is the strike price on a Zen 3 CPU including the 5800x3D. The 5600 (non X) is $183. I think the 5700 is probably the sweet spot (detuned voltage vs. 5800x) no price yet on that besides MSRP. The 5900x is $399. I know we have a couple of weeks until the full refresh of chips is out. Any takes on what the ideal Zen 3 chip is? I am talking price/performance not just performance. For example, if the 5900x hits $300-350, would that change it's value.

I think the 5800X3d will be a niche part but the stacked cache will be a staple for Zen4 CPU's.

 

[Ranulf]

For example, if the 5900x hits $300-350, would that change it's value.

I'll let you know if the deal I got comes through on a 5900x for $300. Sure as heck beats the $450 I paid a month ago and then returned it after the 5700x announcement. I was just going to either skip zen3 or pick something up cheap in a year or two due to frustration.

 

[RnR_au]

The details we have on TSMC 5nm don't suggest that provides either enough of a frequency or a power uplift for a Zen 3 shrink to hit something like a 5 GHz all core turbo within the same 125 W TDP. The transistor density improves a lot, but it's a 15% frequency boost or a 20-30% power improvement. I'm not sure where the person who originally made the post got 20% - 40% improvement from, but it doesn't line up with other published figures. Some articles do talk about other possibilities for additional performance gain, but those parts are light on details or specifically indicate that they come at the expense of density.

On the ~40% performance improvements for Zen 4... AMD is not using straight N5. They are using a custom variant of N5p which affords a ~20% frequency clock speed improvement at the same iso power (over N7). And then there is an expectation of ~20% IPC improvement with all the additional transistors that the brainy AMD engineers might find a use for.

But yeah just guesses in the end.

 

[gruffi]

The details we have on TSMC 5nm don't suggest that provides either enough of a frequency or a power uplift for a Zen 3 shrink to hit something like a 5 GHz all core turbo within the same 125 W TDP. The transistor density improves a lot, but it's a 15% frequency boost or a 20-30% power improvement. I'm not sure where the person who originally made the post got 20% - 40% improvement from, but it doesn't line up with other published figures.

AMD themselves talked about 50% lower power at the same performance or >25% higher frequency at the same power. They don't use regular N5 process but an improved N5 process.