AMD Zen4 is Coming

[Nick Name]

I really haven't kept up on tech in great detail the last couple years, but this sounds pretty good.

The biggest news for the 5nm Zen 4 chips comes as a 15% or better improvement in single-threaded performance over the previous-gen Zen 3-powered Ryzen 5000 processors. AMD tells us that this comes as a mixture of instruction per cycle (IPC) and frequency improvements, but won't share the specific percentage each factor contributes until later. AMD says the chips will reach 'significantly above' a 5 GHz peak frequency and even demoed a 16-core model hitting 5.5 GHz. However, that comes with the standard caveat that this frequency only applies to a single core during a light bursty workload, just as we've seen with Zen 2 and Zen 3 processors.

That means the 15%+ figure isn't based solely on IPC improvements, but improved single-thread performance does boost performance across the board as it is amplified as workloads spread across the cores. To that effect, AMD also boosted the maximum power delivery of the AM5 socket (PPT) that will house the Ryzen 7000 chips to 170W, a 28W increase over the previous-gen AM4 socket's 142W peak. If AMD sticks to its standard use of PPT x 1.35x = TDP, that means we'll see AM5 socket chips top out at a 125W TDP. We'll dive into performance comparisons a bit further below, but it looks like we'll see a closer-than-expected race between Ryzen 7000 and Intel's Alder and Raptor Lake chips.

I'm not sure how I feel about this though.

Finally, AMD has confirmed that the AM5 socket will only support DDR5 memory. The company says that DDR5 provides the extra performance to justify the cost, but we'll have to watch pricing closely. As we've reported, DDR5 pricing remains prohibitively expensive, but Micron expects PMIC/VRM supply to rebound in the second half of 2022, so we could see the situation improve as AMD's Zen 4 Ryzen 7000 processors come to market.

However, DDR5's more complex power circuitry and design mean that these modules will continue to command a premium over DDR4. DDR5 also has in-built ECC mechanisms for data at rest, which requires additional dies to provide the same memory capacity as DDR4. Though the pricing differences will become smaller over time, DDR5 will remain more expensive than DDR4, regardless of supply. We'll have to wait until closer to the launch to see how this impacts Ryzen 7000's platform costs.

My initial thought is I doubt DDR5 is going to benefit crunching, at least in a meaningful way, and almost certainly not enough to justify the cost. I guess it will be one of those pay to play situations.

 

[doneske]

I have never considered memory speed to be that critical to system performance. There are too many bottlenecks between the processor and external memory to negate the memory speed (cache being one of them, memory bus being another). They did increase the cache size for the new processor so that would help some crunching tasks as we have seen in the past. They are also using the PCI5e and the transfers per lane are impressive which may help GPU crunching. They have also included RDNA2 as part of their integrated graphics on the processor. However, if they aren't going to allow any other memory, I won't be buying the processor. The processor will be pricey and the memory will be pricey and I would contend you won't get the benefit to justify the extra dollars.

 

[Nick Name]

The processor will be pricey and the memory will be pricey and I would contend you won't get the benefit to justify the extra dollars.

My thought was it might be better to look for deals on Zen3, from a value perspective.

 

[Vester]

Intel Alder Lake RAM Guide: Picking Between DDR4 and DDR5​

Our Key Takeaways​

• Memory ranks have a minor role with DDR5. Four memory ranks outperform two memory ranks by roughly 3%. The caveat with DDR5 is that only the 32GB memory modules come with a dual-rank design. Therefore, four memory ranks entail buying a 2x32GB memory kit or a 4x8GB memory kit - either alternative is costly.
• DDR5 adopters don't need to worry too much about the frequency for now. DDR5 is still in its early stages, so pickings are slim, and there's much room for improvement. High-end DDR5-6200 delivered a 3% higher performance than baseline DDR5-4800 with the exact timings. If you've picked DDR5 for your Alder Lake build, you should prioritize capacity over frequency. Unless you're a performance enthusiast, it makes little sense to go up to DDR5-6200.
• Memory ranks matter a little more for DDR4. Four memory ranks beat two memory ranks by 5%. Furthermore, DDR4 users have the luxury of picking up dual-rank memory kits without having to pay ridiculous sums of money. It's essential to remember that single-rank 16GB DDR4 memory modules are more common now, so you should double-check the specifications for your 32GB (2x16GB) memory kit before pulling the trigger.
• If you want to play it safe, stick with DDR4-3200. DDR4-3200 should always be your baseline for Alder Lake since it's the guaranteed data rate for Gear 1. Don't settle for anything less. DDR4-3600 is doable on most processors, while DDR4-4000 is for the fearless. Aim for the tightest timings that your budget allows, such as C14.

 

[Nick Name]

Local DDR5 kits are roughly 2x the price of DDR4 at the same density, e.g. 32gb which is about the minimum I want these days. Certainly not worth the minimal performance improvement.

 

[doneske]

From AMD Analyst Day:

- AMD notes 5nm and 4nm products in the Zen 4 space and 4nm and 3nm for Zen 4 processors, without noting the breakdown of SKUs or chiplets between the different processes.

- AMD confirmed AVX-512 and AI acceleration instructions for Zen 4.

- 9~10% instructions per clock (IPC) increase with Zen 4 over Zen 3. <-- This is significant IMO

- Up to 125% memory bandwidth per core improvement with Zen 4, thanks to DDR5.

- Zen 5 in 2024 will indeed be a new "grounds-up microarchitecture" redesign. More AI and machine learning optimizations are expected for Zen 5.

- For EPYC 7004 "Genoa", AMD is talking greater than 75% improvement for enterprise Java performance compared to EPYC 7003 series. 12 channel DDR5 confirmed for Genoa as well as PCIe Gen 5 and CXL support. Genoa is on track for Q4 launch.

- EPYC "Bergamo" with Zen 4C cores for cloud computing will be able to offer twice the cloud density as 3rd Gen EPYC. Bergamo CPUs will launch in H1'2023.

- Genoa-X was confirmed as Genoa with 3D V-Cache. Genoa-X will offer more than 1GB of 3D V-Cache per CPU/socket.

- "Siena" was announced as a lower-cost Zen 4 platform intended for edge computing and telecommunications. Will be quite interesting to see how this Siena play is for lower-cost EPYC.

 

[Nick Name]

- 9~10% instructions per clock (IPC) increase with Zen 4 over Zen 3. <-- This is significant IMO

Seems like a nice bump.

For EPYC 7004 "Genoa", AMD is talking greater than 75% improvement for enterprise Java performance compared to EPYC 7003 series.

To me this is shocking. 75% boost in any use case is amazing. I'm also a little surprised to know Java is still that important.

 

[Nick Name]

This can't be right.....

 

[Jason Jung]

The 7900 just came out and includes a heatsink and the 7900X3D is right around the corner. Based on reviews I've seen the 7900 is identical to the 7900X but with a different voltage/boost curve. Turn on PBO and they are the same. Out of the box I'd prefer the non-X version as the X version seems to use a lot more power for only a small gain in frequency.

 

[Nick Name]

The 7900 just came out and includes a heatsink and the 7900X3D is right around the corner. Based on reviews I've seen the 7900 is identical to the 7900X but with a different voltage/boost curve. Turn on PBO and they are the same. Out of the box I'd prefer the non-X version as the X version seems to use a lot more power for only a small gain in frequency.

This is what I've seen too, but it's still surprising to me the X version - which is supposed to be the premium chip - is priced lower. This is the first I've heard of the X3D version.

Developed for breakthrough performance on targeted critical product design and technical computing workloads, these processors will be game changers for electronic design automation (EDA), computational fluid dynamics (CFD), and finite element analysis (FEA). In these processors, stacking chiplets deliver TRIPLE the L3 cache of other 3rd Gen AMD EPYC™ processors, from 256MB to 768MB.1 It also locates logic units and corresponding memory on top of each other instead of spreading them out, yielding significantly lower latency. Tripling the L3 cache enables these processors to deliver up to 66% more EDA RTL simulation jobs per day for technical computing applications compared to 3rd Gen AMD EPYC™ processors with the same number of cores but no 3D V-Cache

I wonder what pricing is going to be for this beauty.

 

[Vester]

This is what I've seen too, but it's still surprising to me the X version - which is supposed to be the premium chip - is priced lower.

The X version does not have a cooler but the 7900 includes the Wraith Prism cooler.

 

[Nick Name]

The X version does not have a cooler but the 7900 includes the Wraith Prism cooler.

The 7700X doesn't have a cooler either but is priced higher than the 7700, and the same is true for the 7600 series. X versions usually are priced higher because they're supposed to be higher performing version although it doesn't usually pan out that way. The pricing on the 7900 is a bit unusual based on what I've seen in the past.

 

[Jason Jung]

7800X3D: $450
7900X3D: $600
7950X3D: $700

 

[Nick Name]

7800X3D: $450
7900X3D: $600
7950X3D: $700

Not really much of a premium, but I wonder if it will have much benefit. I was looking through Rosetta 5800 stats and the benchmark for the 5800X3D is actually a bit lower than the other 2. I included the 5700 and 5600 since I think those are both great values at current pricing.

CPU model​	Cores​	GFLOPS/core​	GFLOPs/computer​
AMD Ryzen 7 5800X3D [Family 25 Model 33 Stepping 2]​	16​	5.34​	85.37​
AMD Ryzen 7 5800X [Family 25 Model 33 Stepping 2]​	16​	6.26​	100.21​
AMD Ryzen 7 5800X [Family 25 Model 33 Stepping 0]​	15.9​	6.02​	95.63​
AMD Ryzen 7 5700X [Family 25 Model 33 Stepping 2]​	16​	5.79​	92.59​
AMD Ryzen 5 5600 [Family 25 Model 33 Stepping 2]​	12​	5.41​	64.91​

I figured if there was any project that would benefit from the extra L3 cache it would be Rosetta but I haven't been able to find a 5800X3D with any completed tasks for comparison.

 

[Jason Jung]

The 5800X3D has more cache but also has a lower base and boost clock speed than the 5800X. With 8-16 instances of Rosetta tasks running the cache is nowhere near large enough to not have to access system memory constantly so the lower clock speed hurts. I think that will be the case when running many single threaded tasks from any project.

With the 7900X3D and the 7950X3D one CCD will have the extra cache with the accompanying lower clock speeds while the other CCD will match their non-3D counterpart. All L3 cache can be accessed by any core. Should be more interesting to see how it actually affects BOINC performance.

Multiple PrimeGrid sub-projects should benefit tremendously from the extra cache if you run tasks with multiple threads. The 7950X3D will have 128MB of L3 cache letting it hold four SoB tasks completely in cache. The 7950X has 64MB of L3 cache which can only hold two SoB tasks and performance does not scale well past four threads.