AMD Ryzen 7 9700X

[oxrufiioxo]

I am testing at stock, without manual tweaking and addon softwares. I still feel like the 7950X3D should score higher, or maybe AMD software profile fail

And that is the way it should be tested. Anyone buying it though should not use it that way imo....

It should work out of the box though and not require user intervention but it is what it is....

 

[RandallFlagg]

I don't think you are going to see big frequency bumps on Zen 5 as you increase power.

AMD and Intel traditionally design for higher clocks, they are 4 (Zen) and 1+3 (Intel) wide decode pipelines.

Apple by comparison was at 7 wide with the A11/A12 and 8 wide back with the A14, and I believe still at 8 with M1. Samsung was at 6 a couple of years ago and generic ARM Cortex was 6. These are all designed to offer optimum performance per clock, but absolute performance is limited by lower clock speeds.

So AMD has now gone with very wide decode pipeline, 8 wide, which barring any big breakthrough in design is going to limit frequency but will get more done per clock.

This is a huge change for them. This is probably a really good, efficient design for a server. It may play out very well in the mobile / laptop space as well.

However, the way I see it, it's going to have some inherent weaknesses in the performance oriented desktop space.

 

[phints]

Higher idle power consumption and no improvement to 1% lows is a disappointment in what is otherwise a great CPU. Knowing the 9800X3D will ship ~20% faster, especially in 1% lows, is keeping me away from this for the time being. I think TPU sandbagged this review slightly by not using DDR5 6000 CL30 RAM (they used CL36?). I'm also interested to see if Intel improves power efficiency finally with Lunar Lake.

 

[x4it3n]

3 to 5% more performance versus ZEN 4 is really underwhelming... ZEN 5 is supposed to have a 15% increase in IPC on average and we're not even close to 10% in performance lol.

I remember MLID leaking that AMD were having issues with ZEN 5 last year and that they decided to fix everything on ZEN 6 because there was not enough time to fix and release ZEN 5 in time so maybe that's why those numbers are so low... Let's hope BIOS, Firmware and Chipsets updates will improve ZEN 5 performance because it's really disappointing for now.

Let's hope ZEN 5 3D will be better than that...

 

[tussinman]

This chip especially at max MSRP would have been amazing in 2022 but now in 2024/2025 ? Pretty meh

 

[oxrufiioxo]

3 to 5% more performance versus ZEN 4 is really underwhelming... ZEN 5 is supposed to have a 15% increase in IPC on average and we're not even close to 10% in performance lol.

I remember MLID leaking that AMD were having issues with ZEN 5 last year and that they decided to fix everything on ZEN 6 because there was not enough time to fix and release ZEN 5 in time so maybe that's why those numbers are so low... Let's hope BIOS, Firmware and Chipsets updates will improve ZEN 5 performance because it's really disappointing for now.

Let's hope ZEN 5 3D will be better than that...

Super disappointing for sure and while X3D will fix the gaming performance hopefully it will still come with the almost negligible mt improvements over 7000.

WTF happened is my take on this what was AMD doing for 2 years.....

9950X reviews will be interesting I guess.....

 

[oZ65]

TL;DR Applications need to switch from AVX-256 to AVX-512, then we could finally enjoy

For y-cruncher (which I'm the developer of):

• The regular Pi benchmarks and computations gain almost nothing (1-3%) on Zen5 due to being bottlenecked by memory bandwidth.
• If you run single-threaded, you remove the memory bottleneck and get a ~50% IPC improvement on Zen5 thanks to AVX512. (less than 2x due to Amdahl's Law)
• y-cruncher's BBP test (now a benchmark) shows 98% IPC improvement due to the pure AVX512 without any memory access.

In fact, the AVX512 improvement on Zen5 created a memory bottleneck so large that it became the primary reason why I promoted the BBP mini-program from a tool for verifying Pi records to a formal benchmark. The regular benchmarks wouldn't do Zen5 (and future processors) any justice. At least until someone can figure out how to get DDR5-20000 on AM5...

Article: link

 

[AnotherReader]

Super disappointing for sure and while X3D will fix the gaming performance hopefully it will still come with the almost negligible mt improvements over 7000.

WTF happened is my take on this what was AMD doing for 2 years.....

9950X reviews will be interesting I guess.....

From Phoronix's tests, it's clear that gamers weren't a priority when they were designing Zen 5. I expect Zen 5 based EPYCs will dominate both Intel and the ARM competition.

 

[oxrufiioxo]

From Phoronix's tests, it's clear that gamers weren't a priority when they were designing Zen 5.

I get that it's ok in some specialized use cases but in general both gaming ant MT uplifts aren't very impressive.

Again if this was after maybe 12 months sure but it's been 2 years and sometimes loses to the 7700X which should never happen....

I feel like this is the 4060ti of CPU's sure it isn't bad at everything but price still sucks and generationally overall it isn't very good.

 

[Lesha]

der8auer shows +20% improvement using max PBO settings.

 

[R0H1T]

So basically wait for 512bit TR or get Turin?

I get that it's ok in some specialized use cases but in general both gaming ant MT uplifts aren't very impressive.

Looks like memory is more of a bottleneck than ever! The x3d chips could help a lot in that sense.

 

[oxrufiioxo]

So basically wait for 512bit TR or get Turin?


Looks like memory is more of a bottleneck than ever! The x3d chips could help a lot in that sense.

Hopefully because depending on the games tested this can lose by as much as 20% vs the 7800X3D while using more power.....

 

[SalemTheKit]

Tbh it almost feels like the 9700X is the perfect advertisement for the 7900 non-x with how similar they are in price and wattage. In efficiency and workstation tasks the 7900 blows it out of the water and it's not that far behind in gaming.

I'm a bit shocked they didn't include a cooler with the 9000 series so far if they were planning on keeping wattage this low given how the 7700/7900 non-x can run at full tilt without breaking 80c on stock. Now if they only included a decent cooler with the 7600...

 

[oZ65]

So basically wait for 512bit TR or get Turin?


Looks like memory is more of a bottleneck than ever! The x3d chips could help a lot in that sense.

I wonder if trying to push RAM to 6400 Mhz gives tangible results

 

[R0H1T]

Should be fairly easy to check, like if 5-10% faster memory gives close to 5% better results then it is heavily memory bound assuming latency is similar.

 

[kapone32]

Unrealistic settings. Such advanced options aren't available on most motherboards, and he's doing both asynchronous eCLK + meticulously tuned curve shaper on what's no doubt an excellent sample. Most chips won't go that far.

How do you know that? They were just released today. Did yours not do it?

 

[efikkan]

TL;DR Applications need to switch from AVX-256 to AVX-512, then we could finally enjoy
<snip>
Article: link

I agree that many applications need to switch to AVX-512 for some great performance improvements, but like most, people confuse performance and IPC.
IPC means Instructions Per Clock, it's the CPU's average amount of instructions per clock. It's architectural, not application specific. While we usually use benchmarks to approximate relative IPC, we must never forget that performance never equates IPC.

Even moreso, if a workload exists in AVX-512 and AVX2 variants, the amount of instructions per workload actually goes down. So when the performance from AVX512 is almost double, the amount of instructions per clock of that workload is actually lower, not higher, so if anyone uses that to claim the IPC is higher then they've just proven their lack of knowledge.

When we approximate IPC we usually use a number of benchmarks to eliminate statistical outliers, and most important of all; make sure all of them execute the same instructions. Using different code or even different instruction sets (e.g. x86 vs. ARM) to compare "IPC" is utterly meaningless.

Ultimately, understanding IPC doesn't matter to the consumer, as real world metrics like performance, power draw, etc. is the only thing that should matter. IPC is one of the things we discuss to explain the architectural differences, but unfortunately most of you have no idea what it means. This has only become worse in recent years with IPC becoming marketing BS (thanks AMD!).

That being said, and as I've been saying for years, when AVX-512 eventually becomes widespread in productive applications, you're not going to look back. Unfortunately, we should have been there 5 years ago (thanks Intel!).

 

[DemonicRyzen666]

I think you all have unrealistc expactations on the cpu side, or have just been spoiled by 3D V-cache cpus.

Yeah, unless the 9800x3D is going to scale better with cache or has other tweaks it's going to be a whopping 4% faster than the 7800x3D as the 9700x is only 4% faster than the 7700x in games.

There are plently of benchmarls already for zen3 v-cache showing 15% increase in gaming.
Hardware unboxed found that zen4 3D v-cache models were up to 22% faster than non 3DV-cache models of zen 4 in gaming.
Amd has been increasing how good each generation is using the 3D V-cache. To me it would not be unreasonable to expect anywhere from 20% to 35%(maybe over 50% in some games) increase with just 3D V-cache

That fact that single thread IPC is so high, but multi-thread isn't suggests that even with the additional efficiencies in design and new process node, they are still hitting the power / thermal limits of the silicon.
I guess if that constraint was taken away, i.e. free-up the package power limits and use better cooling then it would look a lot better all around... but then you start getting into more exotic requirements.

Wendel mentioned this its a memory bandwidth bottleneck/ latency issue, even threadrippers are having that issue it gets harder to feed all the cores the memory needs they want in mulit-thread. The infinity fabric can only do so much until it gets a new revision or increased speed.

In every area of the core, Zen5 has been redesigned. including a unified schedule for 6 ALU ports. Rebuilt and expanded FPU block. The frontend has dual prediction and prefetching with 2x 4 (8-wide) instruction decoding instead of 4-wide as in Zen 4. These are just first examples, without going into further details.

Dual decoder and predictor can be a problem because it behaves like a Zen4 width (wide decoder 4) in the worst case.

I wonder how a single 8-Wide decoder will behave in LionCove with ArrowLake-S.

With the daigram you point me a small issue when increasing the front end of a cpu it takes about 4 times as much die area normally to do this and increases power usage. We can this in older fx bulldozer designs. This however is in the same original size of zen 4 while still getting lower tempuretures, alobg with lower power consumption.

I keep talking about the Zen5 x86 core. It's safe to say that Zen4 is an improvement over Zen3. But not about Zen5. Zen5's x86 core is a new microarchitecture design from the ground up.

View attachment 357876




View attachment 357877


View attachment 357878

The daigram points to a new starting base with many more improves to come later on. As it can concluded that the dual issue predition is not totally keeping up with the need to feed the cores currenlty. That most likely amd will have to make it a qaud issue predictor to feed it or hexa. Though changing to hexa would mean it probably have other be heavily reliant on feeding the ALU's over the FPU's or switch on the fly.

In terms of cpu design of what was done 15 years ago with Bulldozer this is a revoultionary ground up design. As there in minal loss in any ipc in most areas.

 

[Visible Noise]

It is no Raptor Lake killer, that's for sure. I wouldn't call it a disappointment, it's just not the flawless, supremely performant architecture that many hyped it up to be. It seems most of the improvements are derived from its architectural improvements regarding cache associativity and execution width. Tasks which do not leverage these (which seem to be few right now, even in the niche of video encoding given that Raptor is still beating it in this workload) should perform roughly the same as Zen 4. For the price, I think AMD delivers on an unassuming upgrade after two years of Zen 4.

In other words... idk, buy it if you're building from scratch, otherwise don't bother, your Zen 4 chip is still great, especially if gaming is all you do. Raptor Lake 8P+12E chips such as the 14700K are still leading benchmarks, and Bartlett Lake's purported 10P+0E Core 7 configuration would smoke this chip. The Arrow Lake Core Ultra 7 should be significantly faster even at 6P+8E, due to Lion Cove and Skymont gains.

Overall, I am satisfied with what was delivered. Hope folks manage to buy them soon. More people with great machines the merrier.

When has an AMD release not been overhyped? It’s the same story every time.

 

[Craptacular]

The way I'm looking at it is from an architectural standpoint. Comparing the best perf per watt a given architecture can produce gives you a true sense of it's efficiency. The 7700X isn't indicative of the 7000 series's efficiency, it's essentially a 7700 pushed outside it's sweetspot for very minimal performance gain. It's essentially dead on with the 9700X OC results in that chart.

Saying the 9700X reduces power consumption by 29% is only correct is the very literal sense but completely misses the point that the architecture itself does not appear to be any more efficient. When you compare both architectures within their ideal efficiency range there appears to be zero to potentially negative efficiency improvement. Maybe the 9000 series can be tuned even further for better efficiency but there is not much of a performance gain over the 7000 series for that to happen. Overall it does not paint a good picture.

So, the question then becomes is how do we know that the same isn't true for the 9700x? How do we know if they come out with a 9700 non x that there won't be even more power savings? We don't, not until they come out with one.

If we look at the 5.3 ghz OC power consumption that you are referencing it is within one watt of the 7700x in the 47 application average.

If we look at the performance of the 5.3 ghz we can see in the cinebench single thread that the 5.3 ghz OC gets 7.1 points per watt and the 7700x gets 5.5. In Cinebench that is a 29% better performance per watt for single thread. If we do multi thread the advantage is 10.2% in favor of the 9700x. If we do gaming the 5.3 ghz OC of the 9700x is 3.01 fps per watt, the 7700x is 2.66 fps, that means the 9700x gets 13% more fps per watt when overclocked to 5.3 ghz when compared to a stock 7700x. If the architecture wasn't any more efficient then we wouldn't be seeing these results. Maybe the architecture only shows it efficiency when you go up the wattage curve when compared to Zen 4.

Again, we won't have final confirmation until AMD decides, if ever, to release a 9700 non x.

 

[regs]

Interesting. But it's a very nominal improvement over stock considered that in Roman's testing, it hit around 170 watts at the PBO Max configuration running R23. However, even in this situation, looks like it still loses to Raptor Lake's P-cores as long as AVX-512 is not involved. In Roman's testing he got 22026 multi and 2057 points single, well, just for you, BenchMate'd so you have insurance I'm not BSing you. This is the stock clock configuration with Intel's latest power spec, with my cooling on automatic (quiet). I can probably extract more out of it if I do some careful configuration.

It doesn't matter who it losing to. It wasn't the case. The case was is that it pointless to go over 88 W, well it's not. It's clearly giving advantage, who those who need more multithreaded performance.

As of single threaded performance, I stated it that it needs more single core boost ultimately.

 

[AMDK11]

I don't think you are going to see big frequency bumps on Zen 5 as you increase power.

AMD and Intel traditionally design for higher clocks, they are 4 (Zen) and 1+3 (Intel) wide decode pipelines.

Apple by comparison was at 7 wide with the A11/A12 and 8 wide back with the A14, and I believe still at 8 with M1. Samsung was at 6 a couple of years ago and generic ARM Cortex was 6. These are all designed to offer optimum performance per clock, but absolute performance is limited by lower clock speeds.

So AMD has now gone with very wide decode pipeline, 8 wide, which barring any big breakthrough in design is going to limit frequency but will get more done per clock.

This is a huge change for them. This is probably a really good, efficient design for a server. It may play out very well in the mobile / laptop space as well.

However, the way I see it, it's going to have some inherent weaknesses in the performance oriented desktop space.

View attachment 357898

Zen 4
4-Wide decoder

Sunny Cove
4-Wide decoder (1 complex + 3 simple)

GoldenCove
6-Wide decoder (1+5?)

Zen5
2x 4-Wide decoder

LionCove
8-Wide decoder

 

[regs]

Let's not take things too far so it becomes unstable, I'd rather have an "XT" model based on golden samples with a few hundred MHz more.

It won't become unstable from 95 or 120 W. It's still lower than PBO. And it's just should be available as a single click option for those who needs it. Many, especially gamers, wouldn't need that.

 

[Godrilla]

Just in case some were claiming higher power usage on am5 than previous platforms, I just checked my 7800X3D with -25 on pbo offset at idle at its sitting at 28 watts.

 

[Cheeseball]

Hmm... would've been nice if they integrated USB4 support with the CPU, but I understand there's still not enough room. It's one of the reasons why the APUs are limited to PCI-E 4.0 x8 (and x4 for the 8500G and lower) since USB4 needs at least it for 40 Gbps, and also why these CPUs are limited to just the 610M.

I'm guessing the X870E/X870 is delayed because they're probably deciding what USB4 controllers to use on them since they're mandatory. Hopefully they won't fall back on the old Intel JHL8540 TB4 (because of PCI-E 3.0 x4) controller and use something better like ASM2464PD or even ASM2464PDX (PCI-E 4.0 x4).