News Asus unveils AI Cache Boost — claims up to 19% faster AI workloads on Ryzen 9000 series

So, can anyone confirm whether it was CPU or GPU inferencing? The article mentions a RTX 4090, but where I'd expect this to have a real impact is on CPU inferencing. Perhaps there are some CPU-based layers that need to be processed, so it's a hybrid situation?

It did seem to me like Ryzen 9000 is being held back by its IF links, so this doesn't surprise me. It'd be nice if AMD would do a 9000 refresh with a new I/O die, to give us faster memory support + faster IF, this year. However, if that were in the cards, I think the rumor mill would already be buzzing about it. I guess we'll have to wait until Zen 6 for a proper fix to these problems.
 
Setting FCLK to 2100+ is obvious tweak to boost memory subsystem and is available (and used by people) basically since day 1 on AM5.

Why doesn't Asus do some real BIOS development and, for example, correct permanently broken MCR on 7000 CPUs instead of adding pseudo-features?
Cause they’ve already sold their 600 series motherboards. They now want to sell their 800 series motherboards. They don’t care about your 7000 CPU problems because they already got your money.
 
I think 800 series motherboards have exactly the same problems as they are basically the same hardware.
It’s different hardware but yes gets the same job done. I wasn’t disputing that there are problems, I’m just emphasizing that Asus will only solve the problems that make their 800 series mobos more competitive to sell. They only care about getting your money, once they have it, they don’t care if there is more to be done when they have moved on to selling 900 series mobos.
 
"Why doesn't Asus do some real BIOS development and, for example, correct permanently broken MCR on 7000 CPUs instead of adding pseudo-features?"

You won't see much actual effort to deliver genuine performance upgrades to already released/sold hardware, as long as the media and outlets covering the industry just blindly glaze every empty "upgrade" released. They would much rather (both hardware vendors and the media who make money from affiliate links) see everyone buy brand new hardware every 18-24 months, even if the upgrades you're paying for could have been bios/software updates to your previous hardware.

For example, someone considering diving into AI/LLM for the first time, and don't know how they work, are who these updates/articles are aimed at. Anyone with half a clue understands that the bottlenecks for AI is typically GPU memory/throughput. If you're running a GPU with insufficient GDDR to hold the data needed to operate the model, the speed penalty applied to the compute is so huge using consumer grade, 2 memory channel processors. There are several generations of server/enterprise processors with 6 or 8 channels of DDR, and more PCIe lanes, that absolutely truck all consumer grade ryzen processors in AI/ML. And do it without doing things like decoupling the memory controller speed from the actual ddr speed. I have to laugh about this writer regurgitating the absolute slop about how turning off half the cores to the x3d CPU showing even better performance. This means 1 thing... the benchmark they're using is designed, or compiled, to use less than the maximum cpu threads available. Or they didn't overclock the CPU efficiently for the AI/ML load being tested, so it runs against a TDP or some other limit, and are benefiting from the lower overhead of only one CCD running, letting the remaining threads have a higher net wattage to work with.

Basically, a 2nd gen xeon scalable CPU with its 6 channels of ddr4 ram (heck, the higher performance gold xeon 62** processors are under $200 now, let's grab one of them so we can use sticks of optane dimms instead of standard ddr4 since they're $1 per GB now) will absolutely truck any brand new consumer grade CPU, even the zen5 99503xd, all else being equal (they also offer bifurcation so you can run 4 nvme as a single raid drive, doubling the write speed of pcie5 nvme drives, and you can run a bunch of those bifurcated drives.)
 
Basically, a 2nd gen xeon scalable CPU with its 6 channels of ddr4 ram (heck, the higher performance gold xeon 62** processors are under $200 now, let's grab one of them so we can use sticks of optane dimms instead of standard ddr4 since they're $1 per GB now)
If you care about memory bandwidth, then you wouldn't use Optane DIMMs. They're like 1/10th as fast as DDR4.

Anyway, at 2 DIMMs per channel, you could already fit 768 GB w/ 64 GB DIMMs, so I don't really know why you need Optane.

Getting back to memory bandwidth, the Xeon Platinum 8280 (flagship of the 2nd gen Scalables) supports only up to DDR4-2933. With 6 channels of that, your theoretical bandwidth is only 140.8 GB/s. That's not so much faster than today's gaming desktops. The equivalent DDR5 speed you'd need for a dual-channel setup would be DDR5-8800.
 
It’s different hardware but yes gets the same job done. I wasn’t disputing that there are problems, I’m just emphasizing that Asus will only solve the problems that make their 800 series mobos more competitive to sell. They only care about getting your money, once they have it, they don’t care if there is more to be done when they have moved on to selling 900 series mobos.
The 600 series motherboards and 800 series both use AMD’s Prom21 chip.

The only difference between X870E VS X670E is USB 4.

https://www.pugetsystems.com/labs/a...iB8f-92V692qm8lCth1fIPtRlonAHguhKe_AFk8qqogFc

AM5-chipset.jpg
 
The 600 series motherboards and 800 series both use AMD’s Prom21 chip.

The only difference between X870E VS X670E is USB 4.

https://www.pugetsystems.com/labs/a...iB8f-92V692qm8lCth1fIPtRlonAHguhKe_AFk8qqogFc

AM5-chipset.jpg
Plus WiFi 7, and additional support for 1-2x extra pcie5 m.2 ports. But yes the chipset silicon is the same. Thank you for clarifying that for me!

 
Last edited: