AMD @ Computex 2022 Talks Up Ryzen 7000 Series, Announces Mendocino Budget Laptop APUs

[Teggs]

From a gaming perpective, connecting another drive straight to the CPU is great. OS on one drive, I/O intense program on the other. But it's going to be wasted effort when the drive is placed directly under a GPU heat sink like that. I pity any drive placed under a 600W 4090 Ti or such garbage. Water cooling can solve the problem, but... expense.

I also note AMD is leaning heavily into Zen 3 and even Zen 2 still for low end products. 8-core design and good yields are just killing AMD's low-end desktop offerings. I guess that AMD's response to 'Where is A620?' would be 'Why would we launch that when there are no products for it?', but I think that lack of products is a negative development. With increased cadence in development cycle, they will probably have Zen 3 as low end when Zen 5 is out, and the customer just isn't well served by that.

[ddriver]

As I mentioned just to have the board eat the message... expect those "more than 15%" to be raw compute improvement. CB ST doesn't touch most of the improvements introduced with zen 3. We should see a significantly higher boost in stuff that benefits from avx 512, memory bandwidth, ai acceleration and whatnot.

Zen 3 was already quite strong in raw compute, mostly winning or being still very competitive at sustained compute workloads against intel 12th gen. And there isn't much point to improve something you are already good at, when your competitor uses different approaches to prop up its scores.

As of why would amd want to understate its performance - who knows - catch intel unprepared, build up then crush fanbois, manipulate the stock market...

[jaxa]

That said, I don't see me jumping on the AM5 system right away if not for a long time -- unless current motherboard/cpu breaks of course. I have so much performance just 'sitting there' with my Ryzen 5000 series boxes (5900X and 5600X), I just don't see any need for upgrading. I can't help wonder if there are others out there like me that will just stick with the AM4 platform for the foreseeable future. All my systems run Linux of course. VMs run very well... Just can't see/justify upgrading.

65W TDP APUs should be enough performance for a lot of people. By the time Phoenix and Strix Point APUs come out, the AM5 socket will be mature, DDR5 prices should be down, and you're set if you don't care about anything above 1080p gaming or are willing to use FSR or other scaling algorithms.

But whether you go with graphics-oriented APUs or faster CPUs with weaker graphics, don't be an early adopter if you have performance "sitting there". Just sit back, relax, and enjoy the show.

From a gaming perpective, connecting another drive straight to the CPU is great. OS on one drive, I/O intense program on the other. But it's going to be wasted effort when the drive is placed directly under a GPU heat sink like that. I pity any drive placed under a 600W 4090 Ti or such garbage. Water cooling can solve the problem, but... expense.

I also note AMD is leaning heavily into Zen 3 and even Zen 2 still for low end products. 8-core design and good yields are just killing AMD's low-end desktop offerings. I guess that AMD's response to 'Where is A620?' would be 'Why would we launch that when there are no products for it?', but I think that lack of products is a negative development. With increased cadence in development cycle, they will probably have Zen 3 as low end when Zen 5 is out, and the customer just isn't well served by that.

It's expected that at least some PCIe 5.0 SSDs will have active cooling. It's best to stay away from PCIe 5.0 entirely, and let the early adopters take all the heat.

Here's an article suggesting that A620 will launch alongside Rembrandt desktop APUs (Zen 3+). That sounds reasonable. It's not needed while everything is high-end and DDR5 is still expensive.

[microcode]

I am wondering if there is any plans for a beefier APUs, where single compute chiplet is replaced with GPU chiplet?
Limiting CPU to 8C/16T, but still enough.

The GPU dies are generally much larger than the CPU dies, I think.

[ddriver]

Having the igpu on the io die is quite nifty - you boost your statistical market share, you guarantee yourself a certain level of functionality and compatibility for your software to leverage across the entire range, and you achieve a degree of desirable isolation between the "compute cores" of the cpu, all the while attaining much lower latency than a discrete gpu - all of which positive factors to performance.

I wonder if they are gonna find some nifty enterprise use for the igpu as well, servers could use some baseline gpu support as well, plus it can accelerate certain workloads on soc level.

Next step - stack 16 gb hbm on top of the io die, huge soc level wide fast memory pool without extra footprint - still fits am5. That plus the 64 gigs of L3 on the compute dies. Should be thermally feasible too. And they get to claim the world's first 16 gb cache cpu, even if not sram, there are still major performance and power efficiency gains, and you get to spill over to system ram, and significantly alleviate the strain on the narrower ddr5 controller.

I am wondering if there is any plans for a beefier APUs, where single compute chiplet is replaced with GPU chiplet?
Limiting CPU to 8C/16T, but still enough.

They can make a a bigger io die I suppose. Considering that AMD is already entering the era of mcm gpus, and that it already has a "gpu die" that's very much a system agent, that has IF links to which amd could easily attach a cpu compute die or two, instead of a secondary gpu die. Then can even keep the same alignment and substrate routing, simply scale the io die outward away from the cpu dies. A bigger gpu will however need a lot more memory bandwidth, which sans gddr can only be met with hbm memory. Well, I guess the 12 memory channels epyc io die actually has enough bandwidth to run a midrange gpu core. But dual channel ddr5 igpu will peak out exactly twice as fast as it did with dual channel dd4, that's kinda inevitable.

Pretty soon they will also have "zen 3+" 16 core chiplet at 5nm for am5, which I guess will do pretty well against zen 4 in most use cases. So technically, amd could do as much as 32 zen3+ or 16 zen cores plus more of a mid-ranger igpu to am5. But again - only with hbm. And it will be a bit constrained at 170 watts, they will have to balance the power focus the same way they do on 17 watt power contained mobile apus.

[Drago]

The GPU dies are generally much larger than the CPU dies, I think.

Not if you want to put the GPU on laptop for example. Rurrently all AMD APUs are laptop design chips, retrofitted on AM4 socket, without any I/O die, but a monolithic.

[Csokis]

Intel Mendocino

[HyperDrive]

Mendocino… now there's a code name I haven't heard in a long time…!

[ddriver]

Kids are posting this stuff around, as proof that "15% aint gonna be enough", which I think is quite curious:

4f1b351990b35f0a8ea50389bfb6311a.png


Curious because if we increase the amd result by 15%, we will get exactly the intel 12 gen ddr4 result.

The thing is that CB ST has exactly zero benefit from ddr5. How do I know? Because it is already maxed out at single channel ddr4 - there is no performance difference single vs dual channel ddr4. In fact it barely has any effect (5-7%) on the multicore test as well. So extra bandwidth could not possibly help.

Thankfully, we have the intel result to indeed verify that this particular game, unlike CB ST, benefits from ddr5 significantly. So by simple logic, amd's final performance with the added advantage of ddr5 will be in the same ballpark. There's no reason why amd would not benefit the same way in a clearly memory constrained scenario. Those results also prove that CB ST sees no gain from ddr4 to ddr5:

126995.png

Ironically, the same scores and titles that became unpopular because they made "slightly better value" ddr4 based 12 gen look inferior are now touted to demonstrate how much amd has to catch up, and ironically - ultimately suggest that zen 4 will end up competitive in what's essentially intel's best case scenarios, and in all likelihood, significantly better for compute productivity. Intel's about to lose a significant synthetic bench advantage due to the incorporation of avx 512 in zen 4 as well, another thing that CB does not reflect.

[Anux]

It's not IPC but Cinebench performance. That's the reason they are marketing "single-thread uplift" not IPC like they were previously.

You're right, i totally missed that. And seeing 5800X3D being slower in Cinebench therefore having no gains from cache it leaves room for other code to be substantially faster.

Intel has been on target with their performance numbers.

Yeah if you didn't miss any of their footnotes and knew about the stuff they "forgot to mention" like comparing mobile chips with double the power consumption that will get heavily throttled in real world notebooks. Or compairing it to a 4 year old chip while having already 4 iterations between it.
Always trying to misslead their customers.

I also like the idea of on-board graphics too rather than have special 'g' chips for the purpose.

Yes, having a small GPU on die is really helpfull it just needs enough power to drive 2 4k Monitors + a little head room and provide video acceleration. But the G chips are also great, I'm using 2400G and its amazing how many indie or older games you can play on it. Having a powerful graphics card on board opens many doors in the silent PC, SFF and HTPC area.

That said, I don't see me jumping on the AM5 system right away if not for a long time -- unless current motherboard/cpu breaks of course. I have so much performance just 'sitting there' with my Ryzen 5000 series boxes (5900X and 5600X), I just don't see any need for upgrading.

If you're not extremly rich you shouldn't buy a new plattform at release, especially not in the current market situation. The new boards will hover around 200 €, DDR5 is double the cost of DDR4 with no performance gains and the new CPUs won't be on the cheap side either.
I always buy one generation older or atleast wait till the sucsessor is anounced. The performance difference isn't that big between generations and you can easily get a higher grade model of the older generation to make up for that. Meanwhile prices are lower and you get better linux support.
Having no pressure to upgrade I'm waiting for 5700G to drop to around 200 € or get a decent used deal. That might last me for another 7 to 8 years and saves me from buying 32 GB DDR5 and an expensive mini-ITX board while at least doubling the MT performance to the 2400G.

I also note AMD is leaning heavily into Zen 3 and even Zen 2 still for low end products. 8-core design and good yields are just killing AMD's low-end desktop offerings. I guess that AMD's response to 'Where is A620?' would be 'Why would we launch that when there are no products for it?', but I think that lack of products is a negative development. With increased cadence in development cycle, they will probably have Zen 3 as low end when Zen 5 is out, and the customer just isn't well served by that.

I think they make use of the older designs, to have more capacity for their newer designs. If everything they sell is comming out of 5 and 6 nm fabs only the output would be much less. You also don't buy a 50 € CPU to pair it with 200 € boards and 100 € RAM.