AMD: Zen 3 Announced [2020-10-08]

[swaaye]

That did seem like a really nice bit of architecture giving a meaningful advantage from the iGPU for those with discrete GPUs instead of just being a bit of useless silicon only serving to help spread the thermal load.
Why did it get killed?

Hmm. Well, the EDRAM was a clear win for the IGP, but for the CPU that was not always the case. Just like here with this cache for Ryzen. I imagine the Iris Xe IGP is designed to reduce the need for an external boost like the EDRAM. I can't seem to find anything but the IGP may have more L3 cache itself and the main shared LLC has increased in size a bit. And we're supposed to be using DDR5 now and that's a big memory bandwidth upgrade.

 

[Malo]

To the extent they have mobos that do support yet haven't already upgraded to a 5000 series then yeah massive but I'm not sure how many that would be who waited this long and won't just wait a bit more/save up for a full CPU/Mobo/RAM upgrade to Zen4 &/or should just buy a better GPU.

Zen 1 motherboards have never supported 5000 series. New BIOS's are only just being released now that AMD has decided to allow it by removing support for Bulldozer. There's a big difference in price between me going from a 2600X to 5800X and going to a full new CPU/Mobo/RAM, especially when I'm planning on also upgrading my GPU.

 

[Silent_Buddha]

Wait for Zen4 and whatever next gen GPU comes from either amd or nv (or even intel). Its too close to these new launches to not wait i think.

The thing is, 5800x3D is a drop in replacement so while it'll be slower than Zen 4, it's also massively cheaper if you already have a MB that supports it (almost any AM4 MB assuming the MB maker releases an updated BIOS). DDR5 is still really expensive, not to mention having to buy a new MB for Zen 4.

Combine that with the fact that 1st generation chipsets are generally not nearly as good as 2nd generation chipsets means I'd prefer to wait for 2nd generation chipsets for Zen 4 or higher. 1st generation chipsets will often be buggier with random performance issues.

That makes 5800x3D pretty optimal, IMO. It would allow me to entirely skip the 1st generation chipsets for AMD's next platform while also giving time for DDR5 pricing to go down.

Regards,
SB

 

[Kaotik]

Combine that with the fact that 1st generation chipsets are generally not nearly as good as 2nd generation chipsets means I'd prefer to wait for 2nd generation chipsets for Zen 4 or higher. 1st generation chipsets will often be buggier with random performance issues.

Regards,
SB

"Chipsets" are nothing but bunch of extra ports and their controllers for Ryzen and I can't remember seeing any performance issues with any of them. Care to elaborate on these bugs? Personally I believe most if not all issues with 1st gen Ryzens were int he buggy AGESA and by the time 2nd gen launched it didn't matter whether you had 1st or 2nd gen mobo

 

[Silent_Buddha]

"Chipsets" are nothing but bunch of extra ports and their controllers for Ryzen and I can't remember seeing any performance issues with any of them. Care to elaborate on these bugs? Personally I believe most if not all issues with 1st gen Ryzens were int he buggy AGESA and by the time 2nd gen launched it didn't matter whether you had 1st or 2nd gen mobo

Had lots of issues with the x370 chipset that I got with my 1600x on the Taichi x370. Memory compatibility and issues were also quite bad. The Taichi x570 board I got and used with the 1600x (due to the x370 board dying) had not only better memory support with the same CPU (although still not as good as after I got the 3700x) but also less issues in Windows.

BIOS updates to address issues with memory support also weren't handled well, but that's mostly on ASROCK for how they implemented the fixes that AMD provided them, although AMD also had to provide board partners with multiple fixes before issues were at least mostly resolved.

Happens on both Intel and AMD boards in my experience and has been a thing since the 1990's. Probably since the 80's as well, but I wasn't using Intel machines with custom parts back in the 80's.

Regards,
SB

 

[Kaotik]

Had lots of issues with the x370 chipset that I got with my 1600x on the Taichi x370. Memory compatibility and issues were also quite bad. The Taichi x570 board I got and used with the 1600x (due to the x370 board dying) had not only better memory support with the same CPU (although still not as good as after I got the 3700x) but also less issues in Windows.

BIOS updates to address issues with memory support also weren't handled well, but that's mostly on ASROCK for how they implemented the fixes that AMD provided them, although AMD also had to provide board partners with multiple fixes before issues were at least mostly resolved.

Happens on both Intel and AMD boards in my experience and has been a thing since the 1990's. Probably since the 80's as well, but I wasn't using Intel machines with custom parts back in the 80's.

Regards,
SB

I didn't notice any difference between my x370 (I think it was gigabyte, can't remember exact model) and x570 (Asus prime pro). I did have compatible memory from the start though

 

[Silent_Buddha]

I didn't notice any difference between my x370 (I think it was gigabyte, can't remember exact model) and x570 (Asus prime pro). I did have compatible memory from the start though

For me, I had "compatible" memory (it was on the compatible memory list), but I couldn't run it at the rated speed at the rated timings. OCing the memory was obviously out of the question. Same memory sticks OC'd quite well on an Intel board that I tried it on. With the x570 board, it would run at advertised speed at advertised timings and could be overclocked very very slightly.

Replacing the 1600x with a 3700x then allowed me to overclock the memory quite a bit more as well as allowing me to use 4 sticks of memory instead of 2 sticks of memory (something I couldn't do with the 1600x).

So, yeah, memory compatibility and stability were a combination of the chipset and CPU with the CPU having the larger impact. Oh and memory compatibility was far worse on the launch BIOS for the x370 board. The highest I could clock my 3200 DDR4 was either 2333 or 2400 (can't remember which) with the launch BIOS. It took 3 or 4 BIOS revisions (which updated the AGESA AMD provided code) before I could run it at 2933 with the 1600x and Taichi x370. 3200 was way out of reach until I got the Taichi X570.

Regards,
SB

 

[Kaotik]

So, yeah, memory compatibility and stability were a combination of the chipset and CPU with the CPU having the larger impact.
Regards,
SB

The asrock must have had horrible PCB from memory standpoint. Chipset has literally nothing to do with it though, it's all CPU (and physical wiring on mobo PCB)

 

[Silent_Buddha]

The asrock must have had horrible PCB from memory standpoint. Chipset has literally nothing to do with it though, it's all CPU (and physical wiring on mobo PCB)

All manufacturers had these issues. I chose the ASROCK x370 Taichi based on user reviews as well as professional reviews with particular attention paid to which MBs had the least amount of complaints about memory compatibility. It was one of the ones that had the least issues based on customer reviews.

A big part of why someone wouldn't experience these issues was based on what memory sticks they got, what memory dies were used on those sticks and then random variation in quality of memory dies and memory stick quality due to manufacturing variation.

Plenty of people didn't have the same issues on the x370 Taichi as I did with the exact same memory sticks. But there were some who did. Same went for all MB maker's boards. This was significantly improved with both the x470 based boards and the x570 based boards. Memory compatibility was a big focal point for AMD when they were designing those chipsets. The x570 based boards had almost no issues with memory compatibility for almost everyone, unlike the x370 based boards. However, overclocking the memory or running 4 sticks of memory was still not great on 1st gen Ryzen CPUs as the CPU was still the larger factor there.

Regards,
SB

 

[Albuquerque]

What Kaotik is trying to convey is: the memory stability issues between X370, X470 and X570 had nothing to do with "chipset" changes. Literally, the physical memory traces go from DIMM slot to CPU socket and the "chipset" is no part of the conversation. What you're describing is different factory requirements or tolerances being mandated for later system boards; the physical PCB construction itself is what made the difference in your memory stability issues.

Stories of some people having issues and others not, while working on the same identical motherboards, is a testament to poor quality control on the PCB manufacturing side. Very specifically, electrical noise creating havoc in the DIMM traces or elsewhere.

 

[Silent_Buddha]

What Kaotik is trying to convey is: the memory stability issues between X370, X470 and X570 had nothing to do with "chipset" changes. Literally, the physical memory traces go from DIMM slot to CPU socket and the "chipset" is no part of the conversation. What you're describing is different factory requirements or tolerances being mandated for later system boards; the physical PCB construction itself is what made the difference in your memory stability issues.

Stories of some people having issues and others not, while working on the same identical motherboards, is a testament to poor quality control on the PCB manufacturing side. Very specifically, electrical noise creating havoc in the DIMM traces or elsewhere.

Gotcha, OK, it would have been more accurate to say 1st gen. board designs (inclusive of chipset, PCB, and BIOS) are generally significantly worse than 2nd gen. board designs. For memory issues for first gen Zen CPUs and MBs this was a combination of PCB, BIOS and obviously the memory controller that is located on the CPU.

Regards,
SB

 

[pharma]

AMD Ryzen 7 5800X3D Available for 489 EUR in the EU but 589 USD in the USA (guru3d.com)
April 21, 2022

The new Ryzen 7 5800X3D has started selling, a day after availability the CPU already has dropped to 489 EUR here in the EU. The price was exorbitant high yesterday though.

There are many articles on the web right now indicating vulgar price hikes, however as it always goes, never purchase on the first day of availability guys. At an MSRP of 449 USD the lowest price we could find in the EU was 489 EUR at caseking. In the US however we see a strong price hike at Newegg, they are listing it at a nauseating 589 USD. At that money, you'd be better off with the Core i9-12900 (non-K) and unlocking its power constraints in the motherboard BIOS. "

 

[no-X]

I saw 5800X3D at Amazon for MSRP ($449) yesterday. It's sold out at the moment.

 

[Clukos]

I got one

I was trying the same scene in Cyberpunk with all settings maxed out and as low res as it can go (720p + Ultra Performance DLSS)

https://abload.de/img/rt_smt_fix85kn7.jpg
https://abload.de/img/5800x3d_smt_fixbgkg3.jpg

35% uplift, even with a clock deficit. I'm also using the SMT mod to fix the scheduling for these CPUs.

 

[hoom]

Nice, did you have the 5800XT OCed/Curve Optimisered at all?
And does it show improvements at higher res?

 

[Clukos]

Nice, did you have the 5800XT OCed/Curve Optimisered at all?
And does it show improvements at higher res?

Yeah they are using the same exact settings (in fact I simply loaded the same profile I had on my 5800X )

The 5800X3D I got seems to have a much better IMC though, now I'm using 4000CL16 1:1 fclk

If you are GPU bound you won't see much improvement, but there's some cases at 1440p that I either benefit a lot or simply get better 1% lows, plus I like tweaking hardware so I got it to try out the tech.

 

[Jawed]

1% and 0.1% lows need to be the only criteria for gaming performance!

Cache should help with both of these.

 

[Albuquerque]

1% and 0.1% lows need to be the only criteria for gaming performance!

Cache should help with both of these.

100% agreed. it's not the incremental gain of 120fps peak to 150fps peak that you really care about, it's the move from 40fps minimum to 80fps minimum that really makes the difference.

 

[BRiT]

Twitter thread from Sebbbi:

https://twitter.com/i/web/status/1517801311009390592

Unrolled at https://threadreaderapp.com/thread/1517801311009390592.html
-----

A modern gaming CPU such as the 5950X has achievable memory bandwidth of 35 GB/s. At 144 fps that's 35 GB/s / 144 fps = 243 MB/frame. That's how much unique memory you can access in 1 frame.

5800X3D has 96 MB of LLC. Almost the whole working set fits to the cache.

That 243 MB/frame figure is highly optimistic. It assumes that everything your run on the CPU is bandwidth bound, and it assumes you never access the same data twice. It's common to produce the data first and consume later in the frame. You access the same cache lines twice.

So it's likely that we already have games that fit their whole working set in the 96 MB cache of 5800X3D, assuming the game runs at 144 fps of course. Since games are highly temporally coherent, there's only a few percent change in the working set between the frames

I think this is the reason why 5800X3D is so fast in many games. The LLC is finally large enough to contain the whole working set, especially for games designed to run at 144 fps or higher. All data accessed in previous frame are already in the cache, improving the latency A LOT.

Since 120Hz and 144Hz seem to be the new normal for gaming, it's iteresting to see whether the CPU vendors will continue on this path. Aiming to provide caches large enough to contain the whole working set, being able to provide cache hit for all last frame data.

Some games double buffer their data, some games mutate the current data directly. Mutating one set of data results in smaller working set, but is trickier to parallelize.

Frame temp allocators (bump allocators with frame life time) can also increase the working set, as there's a conservative assumption that all temp data lives until end of the frame, and those memory regions are not reused during the frame.

If CPUs with massive LLCs become common, we might need to reconsider the way we implement our data processing. Minimizing the working set and ensuring the working set is temporally coherent (next frame) becomes crucial for best performance.​

 

[swaaye]

It looks like a catalyst for a new form of gaming CPU arms race. Maybe Intel will bring in the EDRAM again. Maybe it will also help maintain a lower-priced tier of gaming CPUs that are pretty awesome but don't have the fancy halo aura of tons of extra cache.