AMD’s Zen 2 architecture isn’t expected in-market until 2019, and possibly not until the back half of the year (AMD has said nothing public about its launch plan alignment). That hasn’t stopped people from speculating about the new CPUs, and there’s a fresh round of rumors out today.
Chiphell, a Chinese forum known for its CPU leaks (sometimes accurate, sometimes not). This time around, the claim is that AMD could field mainstream desktop CPUs with up to 16 cores and a 10-15 percent IPC (instructions per cycle) uplift when the company moves to 7nm. Core counts would remain the same for Threadripper (TR4) at up to 32 cores, while server CPUs would similarly bump up to 64 cores.
Is this accurate? Maybe. It’s certainly not impossible. But there are some good reasons to watch this news with caution.
First, let’s talk about IPC. A 10-15 percent IPC improvement is easily believable. Ryzen’s overall IPC is roughly comparable to Skylake, which it still lags Intel in some areas. Because Zen is the first iteration of an architecture that’s quite a departure from what AMD has previously shipped, it’s also believable that there’s some low-hanging fruit that the company can address. Our verdict? Not crazy.
The idea of pushing up to 16 cores in a dual-channel memory configuration, however, is a little nuts. Most wfoojjaec readers will be aware of the gap between memory latency and CPU latency (sometimes referred to as the memory wall). One of the most important reason we use CPU caches is to avoid the latency of a main memory search as often as possible, and packing more CPU cores into an SoC means more competition for limited DRAM bandwidth. This is why both Intel and AMD often offer quad-channel or higher memory configurations with their highest-end CPUs — while the exact characteristics of the workload will vary depending on the test in question, more CPU cores generally requires higher amounts of memory bandwidth to keep applications appropriately fed.
That said, there are ways AMD could potentially deal with this challenge. Right now, AMD’s Infinity Fabric clock is tied directly to its RAM clock and RAM itself functions as a last level cache in Zen (Zen has no unified pool of cache common to all CPU cores the way Intel does). AMD could overhaul its CCX design to adjust how it partitions cache between CPU cores. It could also simply increase the amount of cache it allocates per die.
Today, an eight-core Zen die has 16MB of L3. If AMD were to double up the core count per die to 16, it would presumably at least double the L3 to 32MB — but it could conceivably go larger or use a different type of cache allocation scheme. Pushing the L3 cache up to 64MB per die (effectively 4MB per core) could decrease pressure on the memory bus. There’s not a ton of headroom in the DDR4 spec for higher clocks (at least not as far as formal JEDEC specifications are concerned), but a Ryzen 2 on 7nm with formal support for, say, DDR4-3666 would represent a 1.35x improvement in DRAM bandwidth as well.
The other point to consider, however, is that adding more CPU cores simply isn’t a free way to keep boosting performance. Not only does it come with its own costs (higher die sizes, higher intrinsic power usage), it also inevitably limits frequency scaling. There’s a reason why Intel’s 28-core 5GHz CPU needed a massive water chiller to help it hit that clock. The more CPU cores you stack alongside each other, the lower the maximum sustained clock is going to be within a constant TDP envelope.
AMD’s modular approach with Ryzen gives it some freedom to experiment with hitting different core counts, but jumping up to 16 cores as a baseline consumer product is quite a leap. It’s possible that AMD won’t change the baseline core count of a Ryzen 2 die, but might offer CPUs with up to 16 cores as part of a lower-cost Threadripper alternative on its conventional desktop socket. This isn’t even speculation on our part — it’s just an observation of what the company could theoretically do given its socket designs — but it speaks to the flexibility of the company’s options. One way or another, Ryzen 2 should be a significant improvement on previous chips.
RISC-V Tiptoes Towards Mainstream With SiFive Dev Board, High-Performance CPU
RISC V continues to make inroads across the market, this time with a cheaper and more fully-featured test motherboard.
Nvidia Goes All-In On G-Sync With New ‘BFGD’ Ultra-High-End Displays
Nvidia is bringing some of the highest-end displays imaginable to market in 2018, with 4K panels, 120Hz refresh rates, low latency displays, integrated Nvidia Shields, and support for 1,000 nits of brightness in HDR. Yowza.
Samsung Announces High-Speed HBM2 Breakthrough, Codenamed Aquabolt
Samsung has found a way to turn up the speed on HBM2 dramatically without needing a voltage increase. Could this be the turning point for the standard?
Nvidia Calls for Limits as Crypto Hysteria Pushes GPU Prices Sky High
You can thank the surge of interest in cryptocurrency for the increase in graphics card prices, but Nvidia is trying to do something about it. "Trying" is the operative word here.