AMD: Zen 4 single-core increase of 15% is just a conservative figure

AMD showed information on the next-generation Zen 4 processor architecture of the Ryzen 7000 series and AM5 motherboards at Computex this week, but some users found that AMD gave a "single-core performance improvement of 15%. %" doesn't just refer to the improvement of IPC, and the multi-core performance can't win the 12th generation Core, which is slightly disappointing.

It is reported that the AMD Ryzen 7000 processor claims to use the world's first 5nm processor core, of which the CPU core still adopts a small chip design and uses a 5nm process; the I/O core adopts a new 6nm process, integrating RDNA2 core display, DDR5, PCIe 5.0 controller with a low power architecture.

AMD's technical marketing director Robert Hallock accepted an interview with foreign media today to answer some questions about the new CPU and new platform. IT Home summarizes a few key points. Most of the Q&A at the bottom are machine flips, please help to correct errors.

• The 5800X3D is not the last AM4 processor
• AMD 16 cores and 32 threads will be the maximum core count setting when the Ryzen 7000 series processors are released.
• Socket AM5 series motherboards will not support DDR4
• AM5 is still in the early stages, new motherboards will be released in the fall, and AMD itself has no idea of ​​supporting generations of CPUs.
• Zen 4 will use 3D V-Cache technology. 
• In terms of IPC, he admitted that the 15% increase in single-threaded performance currently given is indeed a conservative figure. Regarding the increase in IPC (instruction set per clock cycle), as well as power consumption, chip area, etc., it will be announced later this summer.
• In terms of frequency, AMD's 16-core model is running at 5.5GHz with all cores, but unfortunately games don't have enough power to eat all cores, but for tasks such as rendering, it can meet full-core 5.5GHz
• Zen 4 confirms support for the AVX 512 instruction set, which is the third-generation vector acceleration instruction set (AVX3). Among them, AVX 512 VNNI is used to accelerate neural network calculations, and AVX 512 BLOAT16 is used to accelerate inference. Robert praised them for their good acceleration
• The Ryzen 7000 core is great, but it wasn't designed for gaming in the first place, so it's still only good for low-level tasks.

Anthony: I want to start now and then talk about the Zen 4 stuff, some of the questions I wanted to ask myself, and some of them were suggestions from my readers. So for now, the Ryzen 7 5800X3D is a very interesting addition to Socket AM4, perhaps even unexpected considering Zen 4 is coming. Can you tell us why AMD decided to release it now - is it to showcase 3D V-cache or to enhance gaming performance with some decent competition?

Robert: Can I answer "yes" to all of these? I think what you've seen from AMD, I'm not sure it often recognizes that we've been pushing packaging technology very hard. We're very focused on this Zen or that Zen or the process node we're in. But as we go from monolithic chips to chiplets and hybrid nodes, and now with stacking technology, AMD has made a third column of progress, which is something to come out of on the same five/six year roadmap as Zen. .

So now it's time to show what we can do with 3D stacking. Yes, for gaming performance, but also giving people an idea of ​​where we're going outside of Zen and what we can do for performance without iterating on processes or architecture. So you see now AMD has different performance control columns that are all very important and comparable in terms of revenue, all of which will play a role in our roadmap going forward. So on the one hand looking at what we can do, on the other hand we want to give Socket AM4 a really strong finish, even if it's not going anywhere. We want to end on a high note.

Antony: As far as 3D V-cache is concerned, will this be included in all Ryzen 7000 CPUs, or will there be standard models and X3D models like we have with the Ryzen 7 5800X and Ryzen 7 5800X3D.

Robert: I can't elaborate yet, but I do want people to know that 3D V-Cache is here to stay, there will be Zen 4 and 3D V-Cache, and it's not a one-off technology.

Anthony: Is the 5800X3D the last processor we'll see on Socket AM4?

Robert: Yes, you'll definitely see AM4 running in parallel with a lower price point and more mainstream options. That's not to say Zen 3 is slow. It definitely has a chance to move forward.

Anthony: Socket AM4 longevity has just gotten even better with the new backwards compatibility of Ryzen 5000 CPUs with first-gen AM4 motherboards recently -- is AMD aiming for the same longevity with socket AM5? Or will it be limited to two or three generations of CPUs?

Robert: I don't know yet -- that's the honest answer. We are still in the early stages of building AM5. It's releasing in the fall, but it's still a long way off. One thing we want to clarify is how it looks. Our users want transparency on this topic. But we just don't have the answer yet.

Antony: So a popular question I see is whether DDR5 memory is going to be the only type of memory supported by all chipsets - and it is. Has support for DDR4 and DDR5 been considered? What can you say to enthusiasts who are concerned about the current high price of DDR5 compared to DDR4 and the limited benefits the former offers?

Robert: It's all a macro topic of supply and pricing, and I would say that we've been in active conversations with all of the memory-related suppliers for a few months now, and we want to make sure their supply output is in line with our supply match forecast. We don't see any challenges in supply chain advertising, and in fact, we're very excited that Socket AM5 is all DDR5. I want to be very clear - Socket AM5 will not support DDR4.

So we're all using DDR5, it's a great technology, lower power consumption, better overclocking, better density, and that's why we're totally dependent on it. One will see these economies of scale and exclusivity reduce prices and improve the overall diversity of the market.

Antony: Due to the foundation of Zen architecture and Infinity Fabric, memory speed used to be closely related to CPU performance. Can you tell us more about how Zen 4 works with DDR5 and its timing scaling? Some faster DDR5 kits are being selected which will yield significantly higher performance, like 3600MHz kits vs Zen 3 vs 2666MHz kits.

Robert: I would say Zen 4 is like Zen 3 or Zen 2 when it comes to overclocking. Same functionality, same functionality, similar benefits. This massive relearning doesn't happen with Zen 4. This is predictable. "I would say I'm excited about what DDR5 can provide from memory clocks and fabric clocks, but it's too early to talk about the details.

Antony: Does memory speed have an impact on the performance of the new integrated RDNA2 graphics card? What kind of performance can we get out of this?

Robert: That's a good question. So for us, I think I want to emphasize APUs because people think they have big graphics cores today -- that's going to continue on our roadmap. The Ryzen 7000 series doesn't show that has changed there. The Ryzen 7000 series has only a few compute units in the new I/O chip just to enable display output and video encoding and decoding directly off the CPU. In a commercial market where they don't buy discrete graphics cards at all, this will mostly help us.

Now we have a full stack of Ryzen CPUs that all have graphics and can drive displays. This is a great opportunity for us. We know that for enthusiasts, they always have discrete graphics, so we don't want to push the integrated graphics too far in terms of performance. They are lightweight, office-like graphics.

Antony: And then going to Zen 4, there's a lot of interest in how Zen 4 CPUs are going to balance frequency versus IPC and power consumption -- it looks very different from Zen 3 here. First, we saw some impressive frequency demos this week, and we knew there would be some big gains in single-threaded performance. If you could tell us anything about IPC improvements?

Robert: Not yet. I would like to say a few things about the 15% improvement figure. Note that we are talking about greater than 15%. The asterisk is that we are conservative about it in four different ways. We're very excited about the single-threaded output, both from a process and an IPC perspective, and we fully intend to give people the exact breakdown over the summer. We know this is a hot topic, and we know people expect us to have this kind of transparency. There is no change there.

From a frequency standpoint, I just want to say the 16-core prototype we demonstrated. All cores run at 5.5GHz. There aren't enough threads in the game to have all the cores running at 5.5GHz but in those cores that are running they are all running at 5.5GHz so we get a lot of new headroom from the new processes in Zen 4 so it should Offers some real bonuses.

Antony: The socket power has been increased, which I guess will allow the CPU to reach higher all-core boost frequencies. We already know that cooler compatibility will remain the same, which is great, but will we see an increase in cooling needs? For example, will an eight-core Zen 4 CPU generate more heat than a Ryzen 7 5800X?

Robert: That's a good question. So I think it will manifest itself in several ways. First I want to say that I actually want to check myself since yesterday. My wires are crossed. The 170W TDP is a new option, but that's 230W of outlet power because TDP times 1.35 is always outlet power, so that's true for 65W, 105W, and 170W TDP. So I have cleared the surrounding air.

In terms of coolers, anyone with a large Noctua cooler or a 240mm AIO liquid cooler can use the slot AM5. I use Noctua's NH-D5 myself, but I would also like to say that this new 170W option doesn't mean we've changed extensively in the way we approach TDP. So there will still be 65W and 105W. We just wanted an option to show more multi-threaded performance on higher core count CPUs, and frankly we've left a lot of this capability at lower socket power limits in the past.

So being able to expose this gives us the frequency of Zen 4 and another 50% increase in overall compute performance. It's quite big. It's worth it, although it doesn't go up and down the stack. I think people should understand that every time you shrink a chip and add more transistors, you increase thermal density. It concentrates heat more, but that doesn't mean the process runs hotter or the CPU runs hotter. It's just physics. So we said we want people to be able to use Socket AM4 coolers because we think they are very capable.

Antony: We're used to different CPUs hitting shelves at launch and having to wait for the entire stack to be available. Is there any information on whether AMD intends to release some or all of its key CPUs this fall?

Robert: Unfortunately, it's the last thing we decide, so we don't have any information on that right now.

Antony: The three new chipsets are an interesting choice. What's the main reason for separating the high end from the X670 and X670E? Is this purely because of PCIe 5 support, or whether you think the CPU with the most cores is mostly used by high-end gamers and professionals who might opt ​​for Threadripper.

Robert: That's a good question. The decision to split the X670 into two chipsets actually dates back to the early days of the X570, and all of these motherboards had PCIe 4 on each of them. We listened to feedback at the time, like someone said we didn't have PCIe 4 graphics cards or storage, and I'm not going to upgrade those for now, or can you give me an option without PCIe 4? at lower cost.

We think the situation is pretty much the same with PCIe 5, and now we have an opportunity to take that feedback seriously and do something different. So the X670E chipset has PCIe 5 on two PEG slots (one x16 or two x8). Then an NVMe would be PCIe 5, and then PCIe 5 is optional on the standard X670 chipset.

This offers a wider range of motherboards with the X670 chipset, but not all motherboards increase the cost of PCIe 5. There will be a wider range of price points and we think people will appreciate that.

Anthony: One of the features that more and more people are interested in Thunderbolt 4. Thunderbolt support has been limited so far, but one might expect it to be included on flagship motherboards, especially on the X670E chipset. Any news about Thunderbolt support?

Robert: As we saw at Computex, there are motherboards that support USB 4 and the interoperability with Thunderbolt is well known, so the market has closed that gap.

Antony: I know you may not be able to talk about specific CPU models, but is there any word on cache size outside of any 3D V-Cache implementation? Are these staying the same or increasing?

Robert: Well, we doubled the size of the L2 cache in Zen 4. It's always a really good balance with L2 cache because it takes up a lot of the cores themselves, but as Intel did with Alder Lake, as you transition to a smaller process node, it's cost and capacity Great opportunity to find a new balance between. L2 cache is a nice IPC improvement that increases your hit rate on L2 cache without further scaling and increasing latency. We will discuss other caches later.

Anthony: Are there any noticeable changes here in terms of tuning and overclocking? I know some enthusiasts would like to see more per-core tuning and monitoring.

Robert: We're still working on these overclocking features, but it's too early for us. We'll have to finish the CPU and then move on to overclocking as we get closer to the fall.

Anthony: AMD has been very aggressive in pushing for better acceleration technology. Will you introduce any new acceleration technology in Zen 4?

Robert: AMD will use the same Precision Boost 2 algorithm as it does today. The benefit of this is that it scales with whatever process technology or frequency margin you have. That's why we built it the way we do, because it's a one-size-fits-all solution. If we get a lot of frequency headroom or multi-threading capability or whatever, Precision Boost 2 can scale to accommodate, which is the case with Zen 4.

Antony: Since then, manual overclocking has become less and less useful on Ryzen, thanks to more automated methods like Precision Boost Overdrive 2, and it's only really worth it on specific CPUs or specific workloads. Do you think Zen 4 will be more inclined to draw more resources from the CPU at stock speed, or will manual locking still have a place in the PC enthusiast's toolbox?

Robert: So for us, it was a philosophical choice we made. Let's be clear about what overclocking is. Yes, it's fun, yes, it's exciting, and yes, it feels like you're getting something for free. But for the vast majority of users, the performance in the chip is hidden behind a breach of warranty, which is not right for us. We don't want people to have to break their warranty just to get the extra frequency a CPU is capable of, which is why we're so aggressively sorting parts to take advantage of everything out there, and Zen 4 is no exception. Users should not expect major changes in the way we handle CPU frequency overclocking. Precision Boost Overdrive is still good, Curve Optimizer is still good, memory overclocking is still good.

Anthony: Finally, one of the concerns for me and many readers is Threadripper. With the inclusion of a high-end chipset with full PCIe 5 support in the X670E, along with a new 16-core Zen 4 part (probably the Ryzen 9 7950X) and even more multi-threaded performance, it's looking increasingly unlikely we'll see Successors Threadripper 3960X, 3970X and 3990X. Does AMD have plans for a new high-end desktop CPU?

Robert: All I can say is that Threadripper is not going anywhere.

Is 16 cores and 32 threads the maximum core configuration of the Ryzen 7000 when it was launched?

right.

Will Zen 4 focus solely on the high end?

Do not

At Computex, you showed a 15% increase in single-thread performance over the Ryzen 9 5950X. Wouldn't that just put the gaming performance on a par with the 5800X3D?

I think it's too early to say actually. We are intentionally conservative with our single-threaded performance numbers. We do intend to publish an exact breakdown of IPC versus frequency contribution later in the summer, as well as performance, power and area for the new process. As for what's on par with what, I think it's too early to say, we're still in the chip development phase.

What do you think of Zen 4's 3D Vertical Cache (3DV Cache)?

3DV Cache will definitely be part of our roadmap. It's not a one-off technology. We strongly believe that packaging is AMD's competitive advantage, which can significantly improve people's performance, but we haven't announced anything specific for Zen 4.

Your presentation mentioned "Artificial Intelligence Acceleration". Is it AVX-512 or something more exotic like Intel GNA?

yes. Specifically, AVX 512 VNNI for neural networks and AVX 512 BLOAT16 for inference. Both are pretty decent acceleration, we're not using fixed-function acceleration, which is probably something we can do with the acquisition of Xilinx. We're starting to see more consumer applicability of AI workloads, such as video upscaling, growing a lot over the past two years. I think it's a general trend for the average hobbyist to take on more AI-type workloads. Given our move to smaller process nodes with better performance, power and area capabilities, the time is ripe to bring these features into chips.

Are most SKUs integrated graphics standard?

IGPs are standard. It is included on all 6nm IO dies with a small number of compute units built in specifically to enable video encoding and decoding and multi-display output. Integrated graphics are very relevant to the commercial market, most of our CPUs don't have graphics, and that's a big appetite for customers who don't buy discrete GPUs. Now, we have a richer portfolio of processors that can make a difference in the commercial space. For enthusiasts, it's helpful to diagnose a bad graphics card to get your system up and running while you're still waiting for the GPU to show up. The iGPU configuration [specification] is the same, and all CPUs will have it.

Does this mean the end of the APU on the AM5 desktop platform?

Not at all. We don't actually consider the Ryzen 7000 series to be APUs. It's a processor with graphics, and I know that's a subtle difference. To us, when we say "APU", it actually means that the product has powerful graphics, is capable of playing games, has video encoding, display, drivers, everything. The IGP in Ryzen 7000 is designed to light up displays, handle video encoding/decoding, run home theater PCs, and be more productive, but it's not gaming-grade graphics. APUs with big graphics are definitely part of our roadmap, and you'll see more.

Does IGP support AV1 decoding?

Yes, it functions similarly to the Ryzen 6000 series. It's the same RDNA2 compute unit, same VCN [video] and DCN [show] IP.

Why the new heatsink design? Why are there holes on the side?

This is actually how we achieve cooler compatibility. If you flip one of the AM4 processors over, you'll find a blank spot in the middle with no pins, with room for a capacitor. This empty space is not present on Socket AM5, which has LGA pads on the entire bottom surface of the chip. We had to move those capacitors elsewhere. Because of the thermal challenge, they don't go under the heatsink, so we have to put them on top of the package, which requires us to make a cutout on the IHS to make room. Thanks to these changes, we were able to keep the same package size, length and width, same z-height, same socket inhibit mode, which is what makes the cooler AM4 compatible.

Will the AM4 cooler provide the best experience or is it just "compatible"? IE,

I think it's both. We still intend to offer 65 W and 105 W CPUs, even though the socket power limit has been raised to 170 W. Not every processor will use that power envelope. Coolers designed for 65 W and 105 W parts are also available for AM5. For 170 W socket power CPUs, I expect existing high-end air and liquid cooling units to be very good, but there may be a wave of new solutions positioned to be "designed for" these CPUs. Personally, I have a 5950X in my system on a Noctua NH-D15 and I fully intend to reuse that cooler on a Socket AM5.

We're seeing several exciting new power management features on the Ryzen 6000 mobile processor. Are some of these included in the new Ryzen 7000 CPUs?

Yes, but not as you would expect. In fact, these technologies have actually made their way into the IO die, which is a new design for the platform. We'll talk more about these technologies this summer. You should expect the greatest energy savings and power improvements from the new IODs and the 6000 Series technology we've integrated into them.

In the photo, the computing chip looks gold-plated?

They're not gold plated, it's a process called "backside metallization" that we use to solder the die to the heatsink. Depending on how it's made, it can refract light of different colors (like the surface of a DVD), in this case, it's gold.

in this case, it's gold.

Do low-end chipsets support CPU overclocking? How to overclock the memory? Or will the memory run against some JEDEC benchmark on these chipsets

Yes, low-end chipsets support CPU overclocking. Our chipset overclocking strategy has not changed. Memory, multipliers, voltages, all of which will be available on the B650 and X670.

What can we expect from processors when it comes to CPU overclocking?

I'm not going to make a promise on frequency, but I'm going to say 5.5 GHz is pretty easy for us. The Ghostwire demo was one of many games that achieved that frequency on an early silicon prototype 16-core part with an off-the-shelf liquid cooler. We're very excited about Zen 4's frequency capabilities at 5nm; it looks really good, and there's more to come.

FCLK to memory clock 1:2 for best performance (= ~1500 MHz FCLK), or is 1:1 possible (3000 MHz FCLK)?

We'll have more discussions over the summer. We are encouraged by the architecture and memory overclocking capabilities.

How do you see the transition from DDR4 to DDR5?

AMD is betting on DDR5, Zen 4 has no DDR4 support. Over the past few months, we've spoken to a number of component suppliers, module manufacturers, etc. to confirm their supply roadmaps to confirm timing and avoid shortages. Everyone came back with very optimistic answers. DDR5 will be abundant in the life cycle of Socket AM5. The abundance and new demand for Socket AM5 will help lower the price.

At this point, the demand for DDR5 is limited because our competitors have people skipping DDR5 for DDR4. We think Socket AM5 will bring cost parity or very close to parity. We're very excited about DDR5 because it's a great frequency, which is what Ryzen likes. Like Zen 3, Zen 4 is fabric based, and as we move to Zen 4 in terms of memory and fabric, we see better overclocking characteristics. Memory verification is one of the last things we do because we need a nearly finished platform to do this kind of work. Even at this early stage, we were able to hit DDR5-6400 and more, which is very encouraging.

The PCIe 5.0 configuration of the CPU is a bit confusing. We've seen 24 and 28 lanes mentioned. Can you clarify?

The CPU has a total of 28 lanes, all 5th generation, with 4 stripped for downlink to the chipset and the remaining 24 available to the user. On the X670 Extreme, that means the graphics run on x16 Gen 5 or x8/x8 Gen 5, and there's an M.2 NVMe x4 Gen 5. On the X670 (non-Extreme), only the M.2 NVMe slot needs to be Gen 5, the top slot of the graphics will optionally be Gen 5. On the B650, only the M.2 storage will be 5th generation. Of course, other components, such as companion controllers or additional NVMe devices, can be connected to Gen 5 on the CPU.

All X670E boards use PCIe 5.0 PEG? Or can motherboard manufacturers downgrade to 4th generation for free?

Yes, this is a requirement. It has to be the first two slots for Gen 5; the top slot is x16 if you have one GPU, and x8 for both if you have two cards installed.

The CPU has 28 Gen 5 channels, 16 PEG channels and 4 chipset channels, leaving 8 channels. Can a motherboard have two M.2 sockets connected to the processor?

Yes, it is a possibility.

Why Introduce Chipset E SKU? Such as "X670E" and "X670"

This is actually a direct response to consumer feedback when we launched the X570. We have some users who are still using 3rd generation graphics cards or storage. The extra PCI-Express generation increases the cost of the motherboard, which must have retimer and redriver components to extend the signals from the CPU. Customer feedback is that they like the features and design, but they don't really need the Gen 4 stuff and would rather save on cost. When we consider how to make motherboards more accessible across a wider price range, it makes sense to offer quality motherboards with and without Gen 5 requirements. It will provide a more affordable option for users who don't need the graphics capabilities of Gen 5.

Is the X670E chipset fanless?

It is fanless.

You announced USB4 at Rembrandt, but we didn't see anything about it on the Zen 4 slides.

We'll be talking about USB configurations later in the summer, but you should have seen some motherboards announce USB4 support, I think from Gigabyte this week.

AMD and MediaTek recently announced the development of new Wi-Fi 6E modules, is this a requirement for all AM5 motherboards?

Motherboard vendors are free to choose modules, and they will choose the ones that make the most sense from a cost and functionality standpoint. We don't have any requirements on what they should use, but of course we have AVL. [List of approved suppliers, which components are recommended for use in the design]