AMD: our Zen 4 numbers are conservative in a couple of very important ways

There's plenty more left in the tank, according to AMD.

AMD Ryzen 7000

AMD took to the Computex 2022 stage and drove up anticipation for next-generation Ryzen 7000 series processors. Chief amongst announcements was confirmation of a 170W TDP for highest-performing parts, doubling of L2 cache, and smooth sailing past 5GHz frequency, as seen in the demonstration.

Having whetted our appetite for more information, and with due knowledge Zen 4-based Ryzen 7000 won’t be out until at least the September timeframe, Club386 chatted to Robert Hallock, Director of Technical Marketing, to learn more about AMD’s thought processes regarding next-generation CPUs. Presented in a Q&A style for easy digestion, let’s dig right in.

Club386: Can you explain the reasoning behind increasing the maximum chip TDP from 105W on the present generation to 170W for Zen 4 when core-and-thread count remain unchanged and manufacturing process drops to an energy-efficient 5nm?

Robert Hallock: To be clear, Ryzen 7000 series' maximum TDP is 170W, but the socket power limit is 230W, or 1.35x TDP.

We left quite a bit of performance on the table for AM4 (Ryzen 5000 series), from the 142W socket power limit coming from a 105W TDP. The new 170W TDP is not a frequency thing. We noted many multi-threaded compute workloads, especially on high-core-count chips, were socket power limited above anything else.

When we thought about what people are demanding from high-core-count CPU and what they are most likely to be used for, it felt like a good opportunity to offer the higher TDP. I do want to clarify that doesn't mean we're broadly transitioning to a 170W TDP; we still intend to offer 65W and 105W chips, and those will make up the bulk of the stack.

Club386: How much more performance does going from 142W socket power to 230W give high-end chips?

Robert Hallock: What we ran at Computex 2022 was a prototype part. We have sandbagged that [TDP] number in four different ways. There is still work left to do to optimise the part... so what happens in terms of all-core frequency could change a lot over the summer as we head towards release.

What I can say right now, is that the demo you saw on Ghostwire: Tokyo, almost all the cores were hitting 5.5GHz on a regular basis. That part was configured somewhere between 105W and 170W TDP - in the middle, because it is not a final part yet. We've been very conservative about the single- and multi-thread performance.

Club386: Ryzen 7000 series will have integrated graphics across the entire stack, which is a first. Is the IGP there solely for diagnostic purposes, or is there a modicum of gaming to be had?

Robert Hallock: We want game-capable APUs to be part of our stack moving forwards, and that will continue. Our intent with the Ryzen 7000 series, even though they have graphics built in, is that we still think of them as CPUs. The integrated graphics are light duty - the exact count we will get into later - but primarily we're bringing the graphics in for video encode/decode and display output capabilities. You need a couple of Compute Units in the architecture to activate that IP.

Chiefly, this will help AMD go after the commercial market, where they don't buy discrete, and there is a strong appetite for integrated graphics. But no, to answer the question, these graphics are not intended to play games; they help with productivity.

Club386: Is there a possibility people buying Ryzen 7000 series CPUs in the autumn will be left disappointed that the IGP is merely light duty? When users think of AMD graphics, they think of something more capable.

Robert Hallock: I think there is an ethical responsibility for us to be clear about what the IGP capabilities are. From a messaging and packaging point of view, do we put Radeon graphics or just graphics? We are thinking through that and definitely have to deal with it because APUs will continue to live on, and we need to be clear about what each respective IGP can do.

Club386: Will Ryzen 7000 series’ IGP be exactly the same across the entire stack?

RH: Yes, there is one I/O die with one configuration, and we'll deploy that up and down the stack. The capabilities will mirror Ryzen 6000 series mobile, so four displays, HDMI 2.1, and DisplayPort 2.x.

Club386: The presentation referenced an advanced low-power state for Ryzen 7000 series CPUs. Can you go into any further detail on what this entails?

Robert Hallock: Sure, one of the areas ripe for opportunity in power management was our I/O die. Doing a node (14nm to 6nm) shrink and redesign was a great opportunity to bring in new power architectures. The power IP from Ryzen 6000 series mobile, particularly in terms of controlling graphics power, uncore power, display controllers, is being leveraged in Ryzen 7000 series. We'll go into the specifics in the summer. The I/O die is a great case in borrowing mobile-first power technology and porting it over.

Club386: Ryzen 7000 series will keep to a 16-core, 32-thread architecture, which is the same as the last two Ryzen generations. Is this a slowing of innovation from your side?

Robert Hallock: It's a conscious choice. 16-core vs. 16-core [Intel], we're getting 30-45 per cent more performance out of this SoC. That feels like a very meaningful jump without taking on a lot of extra complexity and cost that would otherwise be passed on to the consumer. 16 cores still feels right, and we're getting a lot of extra performance out of it because of the new node and Zen 4 architecture.

Club386: On to a more practical question, can enthusiasts re-use their cooling for Ryzen 7000 series?

Robert Hallock: Yes, the mounting holes are in the same position and CPU z-height [thickness] is also the same. This means you can use existing AM4 brackets without issue. Technically speaking, the AM5 CPU package size is the same as AM4. Dimensionally, everything is identical, but instead of PGA it's LGA, and this is what dictates the unusual appearance of the heatspreader.

Club386: Switching gears to supporting chipsets, can you explain why AMD will have X670E and X670 this time around?

Robert Hallock: This is a direct response to user feedback from early X570 boards. Many enthusiasts didn't need the compulsory PCIe Gen 4 technology we mandated in that chipset and wanted a cheaper board. Fair point. As we thought about our chipsets in X670, it was a good opportunity to make premium boards available at a wider range of prices by making PCIe Gen 5 optional.

On X670E, we require the top two graphics slots to be Gen 5 enabled, as well as at least one NVMe storage slot. On X670, however, Gen 5 is optional, and if it is enabled, it will be on the top graphics slot only. X670 and B650 must also have Gen 5 storage enabled, as well. Furthermore, no chipset will be required to have a fan.

Club386: Smart Access Storage appears interesting. Can you provide some colour on that?

Robert Hallock: With Microsoft's DirectStorage, the flow of information is NVMe to RAM to framebuffer. Smart Access Storage (SAS) is NVMe direct to framebuffer; we skip the step to RAM in the middle. SAS is cross-compatible with DirectStorage. It'll be available on all three upcoming chipsets.

Club386: Wrapping it up by looking ahead, is AMD confident against Intel’s also-upcoming Raptor Lake CPUs?

Robert Hallock: Yes! We have always tried to be conservative about the numbers we show about Ryzen, intentionally. We don't like overpromising; we don't like saying things we're not sure of. The numbers we have shown thus far have been made conservative in a couple of very important ways.

We'll talk about those ways in the summer. Zen 4 and 5nm puts us in a very healthy spot for 2022 and 2023. It's going to be a real fight. It's going to be good.

The desktop CPU landscape will change hugely this autumn. AMD puts its best foot forward with Zen 4-infused Ryzen 7000 series and rival Intel is launching 13th Gen Core ‘Raptor Lake’ chips armed with more cores and higher IPC.

We feel as if AMD is deliberately holding performance back because there isn’t great indication, as yet, of Raptor Lake’s capabilities. It wouldn’t surprise us in the least if AMD ‘found’ an extra 10 per cent single- and multi-thread performance over the summer.

Another reason for sandbagging may lie with ensuring existing Ryzen 5000 series continue to sell well. Should final-product Ryzen 7000 series hammer present chips, there’s a financial argument to be made for obfuscating true next-gen performance.

Nevertheless, the upcoming battle between AMD and Intel’s finest consumer chips to date has the makings of a blockbuster. We can’t wait.