Zen 5: Dude, where's my IPC?

The first two desktop SKUs of Zen 5 are now available and TLDR, AMD's new generation is not much of a boost over Zen 4 for most games but it’s a lot better for other applications. So you have a situation where most non-gaming-oriented reviews confirm the promised IPC uplift and more, but gaming channels like Hardware Unboxed call Zen 5 a “flop” (web version). Why can’t we get decent gen-over-gen gains specifically in games?

The kinds of games used in these reviews — high-profile "AAA" titles — are famously GPU-bound. Reviewers attempt to fix that by testing with very low output resolutions, like this:

This table illustrates the “Zen 5%” scandal, literally the new CPUs score only 5%–6% above their direct Zen 4 equivalents. But let’s also look at a more complete selection of the CPUs in that chart: not all of them yet, I kept only the CPUs that can be fairly compared in terms of IPC. That means no DDR4, no X3D Cache, no overclock, and no mobile parts or below 6 cores (6P for Intel). What remains are the CPUs that can be compared directly and fairly, isolating factors like external memory bandwidth or the performance tradeoffs for many non-game apps of X3D.

The advantage of Intel’s 13th gen over 12th for gaming is 14%, comparing for example the 13900K to 12900K or 13600K to 12600K. This is what counts as “not a flop”. The full spread for all SKUs is 30%, but that’s comparing a 24-core 13900K to a 10-core (and only 6 good cores) 12600K.

I adapted these charts from TechPowerup’s review because they include many CPUs, equalize all DDR5 platforms to 6000MT, and test at the deep end of 720p. And quite frankly 720p is an embarrassing choice, at that point you might as well be just running good ol’ Time Spy. Most major reviewers stop at 1080p as their minimum. With TPU’s averages, the 13900K’s gen-over-gen lift is +10% at 1080p and only +5% at 1440p.

In summary, Zen 5’s +5% at 720p doesn’t look good. Some of that comes from benchmark choices; the test suites from some reviewers fail to include a single game that’s truly CPU-limited. You can find some in Eurogamer / DigitalFoundry’s review: +11% in both Microsoft Flight Simulator and Crysis 3 Remastered. Also HUB gets a very convincing +18% in Assetto Corsa Competizione. Games like these are outliers… except that old games like Crysis 3, even in recent Remastered versions, tend to be very single-threaded and Zen 5’s single/few-thread performance is better than the multi-threaded scenario of most modern games. TPU also shows +11% for RPCS3, a PlayStation emulator that takes advantage of AVX-512 — despite the fact that RPCS3 currently does not use that above 128-bit operations, but Zen 5’s AVX-512 overhaul goes beyond the full 512-bit datapath.

Note: Apologies to Hardware Unboxed, I previously missed their ACC result and cited their review — one of dozens I have read for this blog, obviously failing in my note-taking — in a more negative way because of that.

Here, I found it!

If you want to make Zen 5 look good, check Phoronix’s great suite of tests that are much better CPU benchmarks and do not require contrived settings to remove noise from GPUs or intense file I/O.

The geomean of all Phoronix tests (chart edited) puts the 9700X +20% over the same-TDP 7700, and the 9600X +25% ahead of both 7600 and 7600X with the TDP not making a difference. These averages over-deliver on AMD's +16% IPC promise. The individual tests show many big gains gen-over-gen (9600X x 7600, same TDP): +38% PHPBench, +33% SVT-AV1, +40% PSPDFKit WASM, +83% Open Image Denoise, +108% Apache HTTPD, +68% OpenSSL, +131% Memcached, +40% PyBench, +51% TensorFlow, +57% PyTorch. I cherry-picked tests and variants but not too much, in particular only three of these tests are likely to rely a lot on AVX-512. (But if you care, check out y-cruncher’s AVX-512 heavy benchmarks.)

That's with all CPUs at stock and with the same DDR5/6000MT; there’s room for more with overclocking and faster RAM but I believe it’s still early for conclusions. We may need some extra AGESA updates, EXPO kits, plus the high-end Ryzen 9 parts, to find Zen 5’s “sweet spots” of CPU & DDR5 tuning that don’t depend on silicon lottery.

Phronix's benchmarks are Linux based and with a bias for servers but for a desktop / Windows friendly review with great selection of synthetic, productivity and gaming benchmarks, check AnandTech.

For one thing, unlike other big outlets they’re not too cheap for SPEC2017. That is the gold standard of CPU IPC benchmarking, and it rules Zen 5 improves +13% in integer code and +25% in floating point versus Zen 4. Among other tests, +37% in JetStream confirms Phoronix’s tests of VM-based language runtimes. If you run fat JS apps like Discord, Slack or VSCode, or also .NET or JVM-based applications, Zen 5 looks great.

In summary, we found the big IPC gains. They are just not in games.

A Balanced View

I don’t use my PC to run any servers, but it’s also not just a gaming system. I also run productivity apps, some light software development, video editing, dabbling with ML… should I buy Zen 5?

Most gaming-oriented reviews reviews don’t help with that question. And many dismiss Zen 5’s huge gains in areas like AVX-512 as niche, but then they have a section of “productivity” testing that’s full of apps like Blender, Photoshop, Premiere, and Handbrake. Earth to YouTube creators: your average viewer never uses that. You use those “creative” apps so you think they are relevant for PC users. Tbh I expect the average subscriber of GamersNexus to have Photoshop — pirated and only used as a fancier MS Paint; and Blender — only the runtime downloaded by BlenderBench.

AnandTech’s review is again much better including the UL Procyon tests for Office apps and others for web, databases, compilers. The Office tests average +4% gen-over-gen, but here this is OK: it’s what you get in these benchmarks for good CPU updates. We can go back to their review of the 13900K and compare Zen 5’s to Intel 13th’s gen-over-gen gains and they are the same in average. Exactly same +10% in Word, the Office test that’s most CPU-bound. If Zen 5 is a flop for productivity, so was Intel 13th gen.

Notice I had to compare data from two reviews 2 years apart with different setups since the Zen 5 test does not include Intel’s 12th or even 13th gen.

Conclusions

Zen 5 seems disappointing for game-only usage, but a normal upgrade for productivity and a strong upgrade for a number of areas from engineering to server-like tasks. For one thing if you are a software developer that needs to run bloated IDEs (all of them these days written in VM-based languages), run local databases and other servers for testing etc., Zen 5 is probably a good upgrade. In fact there’s a fair number of “niches” where Zen 5 flies that will be relevant for the typical PC enthusiast, from testing the last ML novelty to emulating the PS3 games that you definitely own to ranking higher in distributed computing platforms like BOINC.

Does that mean AMD’s CPU architecture has abandoned gamers? I’ve seen some comments about Zen becoming more server-centric each generation: these are EPYC dies reused for client SKUs, and maybe the modularity of chiplets is not anymore sufficient to adequately tune for client / gaming. But I think the answer is simply that AMD’s gaming-optimized solution is the X3D parts, so if that’s really all you need to maximize, just wait for Zen 5’s V-Cache-strapped SKUs. Those never interested me personally because I’m using my build for all that other fun stuff.