Exclusive Interview With AMD’s Robert Hallock on Ryzen | Architecture, Performance & Chip Details

banner

In the next few months, AMD will release their Ryzen range of processors, threatening to a total shake-up of the gaming and enthusiast processor space. Ryzen sports a radically different design to AMD’s current range of CPUs, overhauling cache systems, introducing SMT and adding in advanced code prediction technologies.

AMD claim’s the culmination of their hardware offers at least a 40% IPC gain over their older FX range, and in the companies benchmarks (the most recent being shown during New Horizon) demonstrate an 8 Core/16 Thread Ryzen CPU keeping step with (and beating) an Intel I7-6700K processor.

In this article (and accompanying video) AMD’s Robert Hallock has agreed to answer some of our questions on the processors architecture and tell us their CPU strategy, more details about the AM4 platform and more besides.

Question 1: Can you confirm that there will indeed be multiple Ryzen CPU’s available on launch (so not just the 8c/16t models), and it won’t be a paper launch?

RH: There will be more than one Ryzen SKU available at launch. This is all I can say for now.

https://www.youtube.com/watch?v=mKttIZ3fZjM

Question 2: If you could list 3 things you’re proudest of with the creation of Ryzen, what would they be?

RH: 1. I’m a huge fan of our ITX-specific X300 chipset for Ryzen. Since the Ryzen processor is technically an SoC, it already has enough built-in PCIe Gen3 lanes and I/O controllers to support USB 3.1 10Gbps, x4 NVMe and SATA drives. That’s basically everything you’d want in a tiny system. That means Ryzen doesn’t ALSO need an I/O chipset on ITX boards—this saves area that would normally be needed to route the extra connections the chipset provides.

The X300 chipset is a tiny pinky-finger-nail-sized chip that facilitates secure boot, TPM, and other security-related features—that’s X300. X300 is connected back to the CPU with a dedicated link, freeing up four more PCIe lanes (now a total of 28) on X300-based motherboards for things like WiFi cards, GigE, and other companion chips common on the ITX form factor. I think X300 is a great answer for our fans that have asked us to facilitate more ITX solutions in the market.

2
2. I’m very proud to work for a company that unlocks all of its processors. We’ve been doing that for a long time with Black Edition and –K chips, and it’s back with Ryzen. All Ryzen processors are unlocked for overclocking in the X300, B350 and X370 chipsets. I know gamers have grown quite weary of being deliberately funneled towards a very small selection of overclockable chips with the other guy, and I know that not every user needs the highest-end processors. Ryzen can be the answer.

3. While I cannot yet disclose performance or power efficiency figures, I do want to say that what I feel in my heart: as a gamer, a PC enthusiast, I think that Ryzen is what it needs to be. The Internet hype train moves very quickly and has lots of passengers, so I do not want to be in a position where I’ve over-promised anything… but I truly believe that Ryzen is the competitive processor that people have been hoping for. And I think that our aggressive approach to the market (e.g. unlocking all of them, up to 8C16T @ 95W TDP, ITX-specific chipsets) with Ryzen will be a welcome relief to many.

Question 3: Can you comment on the clock speeds members of the press detailed at CES 2017, with Ryzen running up to 3.9Ghz turbo (though rumor has it that it will be 4Ghz Turbo for final silicone)?

RH: As we have noted during our New Horizons event, our base clockspeed for 8C16T/95W will be at least 3.4GHz. I cannot disclose more at this time.

2nd-1

Question 4: Can you explain the basic synopsis of how both Neural Net Prediction and Smart Prefetch works and their importance in Ryzen?

RH:
Smart Prefetch:
Applications have lots of data associated with them, and I don’t mean sound files, character models, or textures. The application’s underlying codebase is a type of data, plus that code can generate new data while running. This type of data has discernible patterns, because humans wrote both the programming languages and the code, and humans are creatures of habit! Ryzen’s sophisticated algorithms learn the patterns of the code and begin to predict what data will be needed in the future, then preemptively brings that data into the caches of the chip. Retaining that data in cache makes it immediately available for execution when needed, eliminating a possible source of latency penalties.

3

Neural Net Prediction: This one is super nerd-level exciting for me. Oh man. Inside the Zen architecture is an actual artificial neural network that is learning and predicting what instructions inside the CPU will be required. This learning AI is also predictive just like Smart Prefetch, but instead of predicting data, NNP is concerned with knowing what instructions and pathways inside the chip will be needed for executing an application. Staying on top of how an application can best flow through the chip also neutralizes a potential source of latency penalties.
Together, Neural Net Prediction and Smart Prefetch represent a pretty big chunk of the +40% IPC we’ve previously cited.

4

Question 4b: Can you also hint how Ryzen ‘learns’ the application? Does it remember what it’s learned if you leave the application on in the background, then go back to in a few hours of web surfing?

RH: The “buffer” for pattern learning is not megabytes big or anything like that, so there is not room to remember hours or days of history. The buffer is typically flushed when you move from one app to the next, or when the system is restarted. This would manifest as a game or benchmark being slightly faster on the second run through, as now its patterns and behaviors are learned. It’s important for users to know that neither Smart Prefetch nor Neural Net Prediction have any knowledge of what you are doing, only how it is being done by the application. A useful analogy might be: we’ve designed a better pencil by watching how you write, but we’ve never looked at the paper to see what you’ve written.

Question 5: Many of our users are interested in overclocking Ryzen. Can you give any information on how this works alongside both Pure Power and Precision boost? Pure example figures – if the CPU runs at 3.4Ghz turbo, and I set to 3.7Ghz turbo manually, and there’s ‘room left in the tank’ both in power draw and heat (let’s say I have a decent AIO) – will Ryzen increase clocks further?

RH: We will be providing specifics on how overclocking functions during the media review process.

8

Question 6: With the inclusion of DDR4, NVMe, USB 3.1 and so on, AMD is once again very on the forefront with motherboard features and technology. Judging from what board vendors have said in public, reception to Ryzen seems extremely positive – what’s your take on this?

RH: I’m pretty enthused about the quality and breadth of the motherboards coming out of the Ryzen project. All the latest storage and I/O support in some really premium designs. Love it!

Question 7: Intel’s Kaby Lake has seen pretty positive reception from hardcore overclockers, but a lot of disappointment from folks wanting IPC gains (aside from a few small changes for encoding/decoding for example). 

Mark Papermaster was recently quoted as saying Ryzen will be a four-year architecture, which would be TOCK-TOCK-TOCK. Would you agree AMD are being extremely aggressive in pursuing performance for desktop/power users?

2nd-6

RH: I don’t really want to speculate on what the future holds, or might hold. It’s not fair to us, not fair to the reality of silicon development, and it’s not fair to users to set expectations that could change for any number of reasons. For now, I think Ryzen will reflect our commitment to bringing a competitive (on multiple fronts) high-end desktop processor to market.

Question 8: Does AMD feel they have a lot of pressure on them given Ryzen’s immense exposure?

RH: Ryzen is very important, to AMD and the wider market. We understand the significance. We live it every day.

2nd-10
Question 9:
SMT and multicore appear to be a big feature for Ryzen. Can you elaborate as to the reason AMD decided to include SMT technology inside Ryzen, and also AMD seems to have ramped up the cache, improved branching, larger retire, load and store queue, plus several other important things. Does this mean that thread contention with Ryzen will be less of a problem than competing architectures?

RH: What you’ve cited—SMT, better cache strategies, improved prediction—are to boost instruction-level parallelism (ILP). ILP really means “reorganizing incoming instructions into work that can be done faster in parallel.” ILP is a major tent pole of the +40% IPC we’ve promised.

Question 10: Platform support for AM4 seems very impressive – 16x PCIe lanes for graphics, 4 for storage and another 4 for system/chipset. AMD have also mentioned that AM4 will be supported until DDR5 (likely 2020), and of course, AM4 will support APU’s also. Can you elaborate why AMD have decided to stay on the AM4 socket for this amount of time, do you believe it’s better for customers and choice?

1

RH: I think users are pretty weary of the socket merry-go-round from the competition. It seems like not a year goes by when   motherboards must be thrown out for a new one that’s incrementally different. AMD is historically known for providing long-term sockets, and with Ryzen we’ll also be known for the flexibility of our socket infrastructure. 7th-gen APUs, AMD Ryzen and the future Zen-based “Raven Ridge” APU all fit into socket AM4. That will allow users to grow with their system through many performance segments or use cases and over a long span of time.

Question 11: There’s a lot of discussion regarding the performance of the 16 thread Ryzen defeating the I7-6900K processor, in the few official benchmarks you’ve shown handbrake, Blender and a couple of games, this is certainly the case. Can you give any indication if we’ll see a similar scenario across a wide group of benches and scenarios – both single / multi-threaded?

RH: Not yet! Though many users think we’re being deliberately parsimonious with this type of data, the truth is that there’s simply more work to be done before this is ready for the limelight. When we’re ready, and we will be this quarter, we will share. I just ask for a little more patience!

Thanks to both Robert and AMD for agreeing to the interview, and thank you very much for reading. As usual, stick with us at RedGamingTech for more news, technology and an exclusive interview on AMD’s Vega architecture.