Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[Abwx]

It's probably not even that if it's Rembrandt. It's probably 4 cores and either 2 CUs or zero.

For under $100, you are talking about stuff headed to the garbage can otherwise.

They could release 3 SKUs, from 4 to 8C with prices in the 90-140$ range, on mobile RMB is definitly phased out by Phoenix 1 and 2, so the whole 6000 line is available as low cost DT APUs.

 

[SteinFG]

Rembrandt definitely costs more to make than phoenix 2. Even though RMB die is made on 6nm, it's 50% bigger than PHX2

 

[SteinFG]

IDK what's stopping AMD from adapting Van Gogh to AM5 - there's enough PCIe, USB, and DDR5 connectivity to do it. just 130mm² of N6 silicon. Perfect candidate for an athlon chip

 

[ryanjagtap]

IDK what's stopping AMD from adapting Van Gogh to AM5 - there's enough PCIe, USB, and DDR5 connectivity to do it. just 130mm² of N6 silicon. Perfect candidate for an athlon chip

The 8 CUs of Sepiroth (N6 variant of Van Gogh) would be too much gfx for budget parts like Athlons, according to the bean counters. AMD would most likely repurpose Mendocino for that.

 

[SteinFG]

The 8 CUs of Sepiroth (N6 variant of Van Gogh) would be too much gfx for budget parts like Athlons, according to the bean counters. AMD would most likely repurpose Mendocino for that.

Nah, mendocino is too cut down for any desktop use. Not enough PCIe (just 4 lanes) and DDR5 (64-bit).
So it's not happening unless AMD makes a new low-end socket (AM1 successor), which I highly doubt

 

[ryanjagtap]

Nah, mendocino is too cut down for any desktop use. Not enough PCIe (just 4 lanes) and DDR5 (64-bit).
So it's not happening unless AMD makes a new low-end socket (AM1 successor), which I highly doubt

Well then it may be because Van Gogh being Semi-Custom, there could be a exclusivity contract with Valve due to which they cannot use it. But IMO Zen 2 is just not snappy enough to bring it on AM5. (I'll stop now, the thread is going off-topic)

 

[Abwx]

Rembrandt definitely costs more to make than phoenix 2. Even though RMB die is made on 6nm, it's 50% bigger than PHX2

Even if manufacturing cost is higher it will cost less if you can dump unsold chips on the DIY market that is more profitable than the OEM one, beside a 5600GT FI is as big yet it s sold around 100€ in Germany, the lowest available Cezanne part wich is the 5500 is at 90€.

 

[Ghostsonplanets]

IDK what's stopping AMD from adapting Van Gogh to AM5 - there's enough PCIe, USB, and DDR5 connectivity to do it. just 130mm² of N6 silicon. Perfect candidate for an athlon chip

VGH is done. Would be great to see it on DT but doesn’t make sense given 8500G/8600G exist.

Well then it may be because Van Gogh being Semi-Custom, there could be a exclusivity contract with Valve due to which they cannot use it. But IMO Zen 2 is just not snappy enough to bring it on AM5. (I'll stop now, the thread is going off-topic)

Aerith VGH wasn't semi-custom. OEMs rejected it in favor of Renoir/Cezanne and MS cancelled Surface Neo. Only Valve and Magic Leap picked it up.

Sephiroth, however, might be a Valve exclusive SoC. So that would rule it out of DT.

 

[jpiniero]

Rembrandt definitely costs more to make than phoenix 2. Even though RMB die is made on 6nm, it's 50% bigger than PHX2

True, but (as mentioned) it could be very busted chips they are trying to unload.

 

[ryanjagtap]

VGH is done. Would be great to see it on DT but doesn’t make sense given 8500G/8600G exist.

This. The 8500G Phoenix 2 SKU is a great budget option for AM5 sans those who want to buy a dGPU later because of it's lack of PCIE lanes and being on PCIE 4.0 rather than 5.0. It outperforms the 5700G in ST while still having competitive MT. Still if they make a revision of Phoenix 2 with RDNA 3.5 instead of RDNA3, both the 2 Zen 4 +4 Zen 4c and the 1 Zen 4 + 3Zen 4c SKUs will be great ultra budget 1080p gaming setups. (Taking upscaling and frame gen into consideration, of course)

 

[LightningZ71]

Remember, it's expensive moving IP blocks between process nodes.

AMD can do a Rembrandt2 with 4 cores and half an iGPU to shave roughly 35-40% off the size of the N6 Rembrandt die. That would be cheap and performant enough for a "value" line AM5 product. They wouldn't have to do anything to the I/O section, except maybe cut down the PCIe lanes a bit (maybe 8 instead of 16 for the GPU lanes).

The other possibility is doing a 4 core CCD for the existing AM5 Zen4 package done on N4C, TSMC's advertised "lower cost" N4 node that's set to be a long-life commodity node. It's a more relaxed recipe that can be built on the same gear, but with fewer patterning passes. Taking the Zen4 CCD down to 60% of it's size and moving to the cheaper node should notably reduce the cost of the most expensive part of the processor.

If they are willing, they can do a reduced cost package that includes reduce I/O support and a ~half sized N6 IOD too.

 

[SteinFG]

Honestly this whole report seems more baseless the more I think about it, especially considering that AMD still talks about AM4 as a low-cost solution in almost every presentation - during Ryzen 9000 announcement several weeks ago, and during recent tech day. There's no room for something more powerful than 5600G/5600 but less powerful than 8500G/8300G/8400F.

 

[jpiniero]

Honestly this whole report seems more baseless the more I think about it, especially considering that AMD still talks about AM4 as a low-cost solution in almost every presentation

True, but if you are just dumping garbage... there's no better time than the present.

 

[igor_kavinski]

There's no room for something more powerful than 5600G/5600 but less powerful than 8500G/8300G/8400F.

There could be if they decided to stop 5600G/5600 production. If those are getting replaced, better to replace them with something a bit more powerful for essentially the same or even better wafer price (maybe from Samsung?).

 

[LightningZ71]

Remember, AMD wants to move away from VEGA. They are still pushing Renoire/Lucien in the low end market by the bucket loads, even renaming it to the new standard last year. I would have thought that Mendocino would have picked up most of that volume by now, but the 5500U/5700U (and their renames) are hanging on tenaciously on the bottom end. I would think that a half-rembrandt would make some long term volume there giving better cores and more iGPU than Mendocino at the cost of a slightly larger die, plus, it would play better on a desktop as well.

 

[Asterox]

This. The 8500G Phoenix 2 SKU is a great budget option for AM5 sans those who want to buy a dGPU later because of it's lack of PCIE lanes and being on PCIE 4.0 rather than 5.0. It outperforms the 5700G in ST while still having competitive MT. Still if they make a revision of Phoenix 2 with RDNA 3.5 instead of RDNA3, both the 2 Zen 4 +4 Zen 4c and the 1 Zen 4 + 3Zen 4c SKUs will be great ultra budget 1080p gaming setups. (Taking upscaling and frame gen into consideration, of course)

R5 8500G has faster iGPU, or more FPS in gaming vs R7 5700G. As we know, red has both AVX512 and AV1 decoder.

Vega iGPU is old, or in general it is hard to compare old Zen 3 APU with the new Zen 4 APU models in the AM5 socket. And finally, red has an excellent performance per watt or power efficiency.

AMD Ryzen 5 8500G: A Surprisingly Fascinating Sub-$200 CPU Review - Phoronix

www.phoronix.com

 

[Asterox]

It was about time, that someone finally did a classic detailed test on Windows.

AMD Ryzen 5 8500G Review - Zen 4 + Zen 4c Tested

The AMD Ryzen 5 8500G is a highly affordable Socket AM5 processor. It features the Phoenix 2 core, which combines the Zen 4 and Zen 4c architecture, for improved energy efficiency. In our review we test gaming performance with both integrated graphics and a discrete graphics card.

www.techpowerup.com

Blue, i don't remember the last time we saw that prominent detail in any another CPU test. But that's to be expected, and of course it matches the Phoronix test on Linux.
Green, that's why we need detailed tests. Very interesting, considering that this is the first time I'm hearing about that detail.




Application Performance

"With this review we're introducing our new 2024 CPU Test Suite, which runs the newest versions of our apps and includes new workloads, mostly for AI-related tasks, which are becoming more and more important every day. Averaged over these 49 tests, the Ryzen 5 8500G offers performance roughly matching the higher positioned Ryzen 7 5700G. It's also able to outperform the Intel 12400F by a small margin—a very important win. AMD's AM5 Ryzen 7000 "Raphael" CPUs are considerably faster though, the Ryzen 5 7600 is 20% faster in applications. Compared to older Socket AM4 processors, the Ryzen 5 8500G sits roughly between the Ryzen 5 5600X and the Ryzen 7 5700X. Overall, this is very respectable performance, especially for such an affordable processor that's targeted at lighter tasks like office, productivity and internet browsing.

Power Consumption

The Phoenix 2 chip design used on the Ryzen 5 8500G was originally designed for thin and light laptops where energy-efficient operation is one of the most important capabilities. On top of that, AMD is fabricating their chip using TSMC's 4 nanometer node and it's a monolithic design, which further helps with power usage, because there's no IO die that's off-silicon that you need to push data over long distances, which is an energy-hungry task. The result is that Ryzen 5 8500G is the most power-efficient processor we've ever tested—by a big margin in many tests. Even when loaded with demanding tasks its power usage will stay at under 40 W, gaming with a discrete GPU is actually closer to just 20 W. Gaming with integrated GPU reaches between 20 and 40 W, depending on the game. Very impressive numbers, only the Intel 12100F can get somewhat close to these results and everything else is much higher. For our 2024 Test Suite upgrade we've added full system idle power measurements, which are very important, because many PCs spend most of their life in that state. Here the 8500G sits roughly in the middle of the pack with 64 W, which is just a few W more than competing Intel setups, but considerably lower than the "big" Ryzen 7000 CPUs, which reach around 80 W, because they have to power the IO die as well."

 

[gdansk]

I'm not sure which Zen 4 thread this would be relevant to. But here is an example of AVX-512 optimization targeting Zen 4:

tolower() with AVX-512 – Tony Finch

dotat.at

And it is interesting to see how Clang's autovec actually did OK. Not as good as hand-written version but still not bad for a compiler flag.

 

[igor_kavinski]

I'm not sure which Zen 4 thread this would be relevant to. But here is an example of AVX-512 optimization targeting Zen 4:

tolower() with AVX-512 – Tony Finch

dotat.at

And it is interesting to see how Clang's autovec actually did OK. Not as good as hand-written version but still not bad for a compiler flag.

An article like that would interest our resident CPU lovers @SarahKerrigan and @Nothingness

 

[SarahKerrigan]

An article like that would interest our resident CPU lovers @SarahKerrigan and @Nothingness

I don't love CPUs. I barely tolerate them.

Also, nothing in that post is particularly surprising. It's vectorization of a vectorizable function.

 

[Nothingness]

An article like that would interest our resident CPU lovers @SarahKerrigan and @Nothingness

No breaking news here. Beyond FP, AVX-512 has been designed to also be useful at managing strings. There are many examples on Daniel Lemire (the guy behind simdjson) blog: https://lemire.me/blog/

 

[Joe NYC]

There is a CPU launch today after all. Zen 3 "XT" launch, where XT is apparently + 100 MHz.

https://twitter.com/x/status/1818339061465010331

 

[RnR_au]

edit: sheesh... wrong thread...

 

[igor_kavinski]

New SKU invented!

 

[Abwx]

Zen 4 X3D for the masses :

AMD Ryzen 5 7600X3D Reportedly Launches In Early September, 6 Core "Zen 4" With 3D V-Cache

AMD is reportedly preparing a brand new 3D V-Cache CPU under its "Zen 4" lineup, the Ryzen 5 7600X3D, which will cover the budget market.

wccftech.com