Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD Zen 3 is coming: +19% IPC, +200 MHz, +$50 MSRP, Nov. 5
The key points are in the thread title. The new lineup will be:
Ryzen 9 5950X: 16 cores, 4.9 GHz, $800, 105 W
Ryzen 9 5900X: 12 cores, 4.8 GHz, $550, 105 W
Ryzen 7 5800X: 8 cores, 4.7 GHz, $450, 105 W
Ryzen 5 5600X: 6 cores, 4.6 GHz, $300, 65 W
They use the same chiplet approach as third gen Ryzen. AMD has apparently decided that fourth gen is laptop only, or maybe will have desktop APUs at some point. Presumably lower end Zen 3 CPUs are coming later.
AMD is also claiming a 19% IPC increase. That is, at the same clock speed as before, programs will typically run 19% faster than on a Zen 2 CPU. Added to the clock speed increase, that would mean Zen 3 is typically about 25% faster than the analogous Zen 2 CPU, and while using about the same power as before.
AMD claims that Zen 3 typically beats a Core i9-10900K at single-threaded performance and is the fastest desktop gaming CPU there is. They also compared it to Tiger Lake at 4.8 GHz and said that Zen 3 beats it in single-threaded performance at Cinebench. That has long been a very favorable benchmark for AMD's Ryzen CPUs, but I'd interpret it as a Ryzen 9 5950X being roughly competitive in single-threaded performance to Intel's top Tiger Lake CPU that has been paper launched for laptops but might not actually be available to buy for quite some time.
We'll see what independent benchmarks show. AMD's history here is one of mild cherry-picking. That is, the average benchmark AMD chooses tends to be a few percent more favorable than the average benchmark that independent reviewers choose. Cinebench is, of course, much more cherry-picked than the rest of the benchmarks that generally showed a Ryzen 9 5950X beating a Core i9-10900K.
Zen 3 brings a lot of architectural changes as compared to Zen 2. The key one that AMD was willing to explain is the unified 8-core complexes. A Zen 2 chiplet had two complexes of four cores each, with a 16 MB L3 cache for each complex shared by the four cores in that complex. In order for one core to access the other's L3 cache, it had to hop across chiplets to go through the I/O die. Zen 3 features 8 cores with a unified 32 MB L3 cache that all cores can access. That's the same total amount of cache as before, but the cores just have better access to it.
We'll see if Intel has any answer for this. Rocket Lake is coming, but we don't yet know what it is. Sky Lake Refresh Refresh Refresh Refresh Refresh would be dead on arrival unless Intel decides to slash prices to compete at the low end. Backporting Willow Cove cores to 14 nm could be competitive in raw performance with Zen 3 if they can still clock high, but likely at a cost of double the power consumption or more.
Ryzen 9 5950X: 16 cores, 4.9 GHz, $800, 105 W
Ryzen 9 5900X: 12 cores, 4.8 GHz, $550, 105 W
Ryzen 7 5800X: 8 cores, 4.7 GHz, $450, 105 W
Ryzen 5 5600X: 6 cores, 4.6 GHz, $300, 65 W
They use the same chiplet approach as third gen Ryzen. AMD has apparently decided that fourth gen is laptop only, or maybe will have desktop APUs at some point. Presumably lower end Zen 3 CPUs are coming later.
AMD is also claiming a 19% IPC increase. That is, at the same clock speed as before, programs will typically run 19% faster than on a Zen 2 CPU. Added to the clock speed increase, that would mean Zen 3 is typically about 25% faster than the analogous Zen 2 CPU, and while using about the same power as before.
AMD claims that Zen 3 typically beats a Core i9-10900K at single-threaded performance and is the fastest desktop gaming CPU there is. They also compared it to Tiger Lake at 4.8 GHz and said that Zen 3 beats it in single-threaded performance at Cinebench. That has long been a very favorable benchmark for AMD's Ryzen CPUs, but I'd interpret it as a Ryzen 9 5950X being roughly competitive in single-threaded performance to Intel's top Tiger Lake CPU that has been paper launched for laptops but might not actually be available to buy for quite some time.
We'll see what independent benchmarks show. AMD's history here is one of mild cherry-picking. That is, the average benchmark AMD chooses tends to be a few percent more favorable than the average benchmark that independent reviewers choose. Cinebench is, of course, much more cherry-picked than the rest of the benchmarks that generally showed a Ryzen 9 5950X beating a Core i9-10900K.
Zen 3 brings a lot of architectural changes as compared to Zen 2. The key one that AMD was willing to explain is the unified 8-core complexes. A Zen 2 chiplet had two complexes of four cores each, with a 16 MB L3 cache for each complex shared by the four cores in that complex. In order for one core to access the other's L3 cache, it had to hop across chiplets to go through the I/O die. Zen 3 features 8 cores with a unified 32 MB L3 cache that all cores can access. That's the same total amount of cache as before, but the cores just have better access to it.
We'll see if Intel has any answer for this. Rocket Lake is coming, but we don't yet know what it is. Sky Lake Refresh Refresh Refresh Refresh Refresh would be dead on arrival unless Intel decides to slash prices to compete at the low end. Backporting Willow Cove cores to 14 nm could be competitive in raw performance with Zen 3 if they can still clock high, but likely at a cost of double the power consumption or more.
![[Deleted User]](http://wa.vanillicon.com/v2/a5231d9a9017cdf5f82125ccaf807f8f.svg)
Comments
AMD is hiking the price up a bit too much.
EDIT: With that said, if they really give that +19% IPC increase they're promising I think I'll be buying one of their new CPUs.
If you think Zen 3 is too expensive, then Zen 2 is still on the market, as are a variety of Intel processors. There will also surely be lower end Zen 3 parts eventually.
The 32 MB unified L3 cache is one of the ways that Zen 3 increased IPC as compared to Zen 2. Intel has had unified L3 caches since 2008, though usually much smaller, currently around 2 MB per core.
We'll see where AMD and Intel go from here. It's easier to make big improvements when you're way behind.
Each Zen 3 CPU chiplet has a 32 MB L3 cache shared between all of the cores. It's not 36 MB or 72 MB. It's 32 MB. A Zen 2 die had two separate 16 MB caches each shared between four cores. Both Zen 3 and Zen 2 have a 512 KB L2 cache for each core that is accessible only by that core.
AMD's numbers larger than 32 MB are adding multiple caches together. They add their L2 and L3 cache sizes together just because it gives bigger numbers, even though it's stupid. Furthermore, on the Ryzen 9 CPUs, even though the package has 64 MB of L3 cache, it's not a unified L3 cache directly accessible by all of the cores. Each chiplet has 32 MB of L3 cache, and to go to the other chiplet's cache, you'll have to go through the I/O die.
During the early days of multicore processors, Intel had Core 2 Duo CPUs that could clock higher than their Core 2 Quad ones. That led some people to say hey, hardly anything will use more than two cores, so the Core 2 Duo is faster than the Core 2 Quad because it clocks higher. So people who were willing to pay for the best without regard to the price tag ended up getting a cheaper CPU. Naturally, Intel was unhappy about that.
That four cores couldn't clock as high as two was really just physics back when all cores had the same clock speed. Then Intel introduced power gating and turbo with Nehalem. With Sandy Bridge, they got turbo working well, and ever since then, the CPUs that had the highest stock turbo were always the ones with the most cores and also the most expensive in the lineup, at least among the mainstream consumer models.