Wednesday, May 29th 2024
AMD Ryzen 9000 Zen 5 Single Thread Performance at 5.80 GHz Found 19% Over Zen 4
An AMD Ryzen 9000 "Granite Ridge" desktop processor engineering sample with a maximum boost frequency of 5.80 GHz was found to offer an astonishing 19% higher single-threaded performance increase over an AMD Ryzen 9 7950X. "Granite Ridge" is codename for the Socket AM5 desktop processor family that implements the new "Zen 5" CPU microarchitecture. The unnamed "Granite Ridge" processor comes with an OPN code of 100-0000001290. Its CPU core count is irrelevant, as the single-threaded performance is in question here. The processor boosts up to 5.80 GHz, which means the core handling the single-threaded benchmark workload is achieving this speed. This speed is 100 MHz higher than the 5.70 GHz that the Ryzen 9 7950X processor based on the "Zen 4" architecture, boosts up to.
The single-threaded benchmark in question is the CPU-Z Bench. The mostly blurred out CPU-Z screenshot that reveals the OPN also mentions a processor TDP of 170 W, which means this engineering sample chip is either 12-core or 16-core. The chip posts a CPU-Z Bench single-thread score of 910 points, which matches that of the Intel Core i9-14900K with its 908 points. You've to understand that the i9-14900K boosts one of its P-cores to 6.00 GHz, to yield the 908 points that's part CPU-Z's reference scores. So straight off the bat, we see that "Zen 5" has a higher IPC than the "Raptor Cove" P-core powering the i9-14900K. Its gaming performance might end up higher than the Ryzen 7000 X3D family.
Many Thanks to TumbleGeorge for the tip.
Source:
Wccftech
The single-threaded benchmark in question is the CPU-Z Bench. The mostly blurred out CPU-Z screenshot that reveals the OPN also mentions a processor TDP of 170 W, which means this engineering sample chip is either 12-core or 16-core. The chip posts a CPU-Z Bench single-thread score of 910 points, which matches that of the Intel Core i9-14900K with its 908 points. You've to understand that the i9-14900K boosts one of its P-cores to 6.00 GHz, to yield the 908 points that's part CPU-Z's reference scores. So straight off the bat, we see that "Zen 5" has a higher IPC than the "Raptor Cove" P-core powering the i9-14900K. Its gaming performance might end up higher than the Ryzen 7000 X3D family.
132 Comments on AMD Ryzen 9000 Zen 5 Single Thread Performance at 5.80 GHz Found 19% Over Zen 4
For those who don't know, clock-to-clock Zen4 offered a 1% improvement over Zen3 in CPU-Z. So to bring 19% gains in such a shallow benchmark shows major design changes.
But wasn't this leak declared a fake?
With AM5 AMD decided to give users what they where cheering for. Some extra performance for much higher power consumption.
It runs entirely from L1I amd has a bunch of things that are synthetic and easy for modern CPUs.
Op claimed that amd's launching zen5 in august, then october. Said that the source has proofs and can confirm it
Did a 180 turn 1 day later and claimed june launch, july availability (was known long before he made his post)
A user put up a screenshot of alleged zen5 cpuz bench, deleted it after a couple hours
He claimed that he did it just for fun and didn't expect people to repost and take it seriously
Chinese users laughing at wccf reposting his shit
A number of chinese tech forums have already started to warn users against sharing baidu bs and threatened bans or infrator points. Your choice on whether ya wanna believe in them"
"Also, the CPU-Z screenshot doesn't list AVX-VNNI in ISA extensions."
Amd/comments/1d2od9j
Moreover, i still remember the story about how developers of this utility revised their tests after, irc, Zen1 showed better results (and its result was of course downgraded).
Didn't someone from Intel have contact with these developers back then?
I know better than to believe those claims until I see them. It's usually too good to be true.
I however wouldn't say that is the same for AMD who has showed us double digit increases in IPC in the last 10 years easily.
AMD Excavator lineup to Zen 1: There was roughly 52% IPC Increase
Now if it's 19% increase in a real world application, that would be progress. Sort of similar to Zen 2 - Zen 3. Based on the architectural changes, it should at least be more than Zen 3 - Zen 4.
Note 2: Neither AMD nor Intel is FORCING users to run the CPU at 250 Watts, and neither AMD nor Intel nor Nvidia is FORCING gamers to run GPUs at 400 Watts. Running a CPU or GPU at high wattage is an OPTION offered to consumers. Another OPTION offered is to limit CPU's max temperature to 75℃ in the BIOS (single-threaded performance stays the same, while multi-threaded performance is reduced). 144Hz 4K HDR gaming is just an OPTION offered by high-end displays. Complaining about 250 Watt CPU consumption, while multiple options of how to limit/optimize power consumption and temperatures do exist and are fairly obvious, is a sign of incompetence+misunderstading on the side of the user of the desktop machine.
For long years there were regular huge improvements in manufacturing processes that enabled huge increases of transistor budgets and huge efficiency increases. These manufacturing process increases have slowed down a lot in recent years but the industry and consumer expectation is for the performance of end product to keep increasing.This is about cost to the consumer. If consumer would prioritize buying - and paying for - low power consumption and efficiency the products offered would reflect that. Basically for a CPU or GPU it means going larger-wider (more cores, more shader) and lower frequencies. This is exactly what enterprise and data center are doing - they feel the power requirements, cooling requirements and initial investment of buying the thing is relatively smaller. Thus, products offered there are more efficient.
Nothing stops me or you from buying an RTX 4090 and running it at half power limit - at 225W that will become a very efficient GPU with surprising bit of its performance intact. The problem - this will bring its performance down to lets say RTX 4080 level. RTX 4080 would be much cheaper to buy.
Although if I remember correctly 4090 is most efficient somewhere around 300W where it does not lose as much performance and would still be faster and more effcient than RTX4080. More costly, still.
It has been clear for a while that frequencies will no longer increase considerably which has an effect on how microarchitectures need to evolve. Some - if not most - of the evolution has happened and will have to happen on different levels. Multi-/manycore CPUs and their consenquences in the system and software level has been significant and will go on.
Purely on microarchitecture there seem to be two cardinally different directions being attempted - going small simple like RISC-V or some of ARM, alternatively going wide and complex for which the Apple M is probably the best mainstream example. I think the problem with simple is that it will eventually have to rely on either clock speed or parallelism to improve. Clock speeds are not expected to improve considerably these days and parallelism works well with cores of any size or complexity. Plus of course ASICs for specific tasks for efficiency improvements.
Interesting times either way :D
Just a note: JIT compilation arrives to CPython with version 3.13, although it might be initially disabled by default and might noticeably improve performance only after version 3.14+ (peps.python.org/pep-0744/)The RISC-V standard will have (but I have no idea when it will happen, it is taking a long time) an extension "J" for accelerating dynamic programming languages (github.com/riscv/riscv-j-extension). With it in place, competition between ARM/x86 and RISC-V might become quite interesting.
In my own experience, I could run a 4070 Ti Super at 200W instead of the rated 285W, and lose maybe 10-15% framerate or compute throughput doing so. The only significantly affected benchmark I've seen so far was FurMark. It also appeared that significant power - on the scale of 50W for this card - would be consumed by the memory bus and VRAM when they run at rated frequency, a consumption that won't be significantly reduced by reducing board power limit, amplifying actual GPU power reduction by percentage in such scheme.
So...yes, it aligns with expectations.