News AMD's integrated graphics performance shines in AI tests with extreme memory — Ryzen 7 8700G with DDR5-9000 up to 15% better

[Admin]

Enthusiasts have demonstrated the prowess of AMD's Ryzen 7 8700G with DDR5-9000 memory and how fast memory affects AI performance.

AMD's integrated graphics performance shines in AI tests with extreme memory — Ryzen 7 8700G with DDR5-9000 up to 15% better : Read more

 

[Amdlova]

15% better for combo CPU + RAM our you get a i3 and a 4070 super for 350% better performance

 

[ThomasKinsley]

I'm going to make myself sound very old, but I remember the DDR spec generally doubled with every edition (DDR1, 2, 3, etc.). But since DDR4 we went from 2133MHz to 9000MHz. That's a significant change of pace and hopefully it will continue.

 

[d0x360]

I've been playing with ai video upscaling (7950x3d @ 5.4ghz ccd0 & 5.9 GHz ccd1), rtx 4090, 64 gigs 7000mt ddr5 @ CL28-28-30-70 oc'd from 6000mt) and with a combination of memory, cpu and GPU overclocking I've got the upscale time down from 16 hours to 9... on a 4 min 1080p to 4k video using a Vulkan based scaler lol.

Sigh.. someday deep space 9 some day! I wonder if a ai accelerated CPU if I'd get gains assuming the tool supported it. I had my 4090 pegged at +350 core+1400 mem and it was chugging all the vram with a tile size of 400 so I had to lower it to 375 which left me 2.4 gigs of wiggle room so no crashing but it increased my final output time from 6.9 per frame to 9.1s. booooo hissssss! I'm g testing out fastencode on ffmpeg tonight on a 6 sec clip to see how the quality is because that's another reduction of like 30%. Also need to enable my igpu for mgpu and see if that lil boi can help any lol. Only an hour to go!

 

[TerryLaze]

Enthusiasts have demonstrated the prowess of AMD's Ryzen 7 8700G with DDR5-9000 memory and how fast memory affects AI performance.

AMD's integrated graphics performance shines in AI tests with extreme memory — Ryzen 7 8700G with DDR5-9000 up to 15% better : Read more

The memory timings (36-51-49-55) weren't so bad, but he had to push the DRAM voltage to 1.65V, which is ok for demonstration purposes but not a value we would recommend for daily usage.

So did AMD fix the mem controller issue? Do they use a new one? Did anybody report on that?! Are bioses still being restricted to 1.35V on these and is it safer now to overclock the ram?

 

[strobolt]

It didn't take long for AMD's Ryzen 8000G (Phoenix) processors to find their way onto the list of best CPUs.

It's just that that they didn't. There's no single 8000G CPU on the list of best CPUs because the value proposition is not there.

 

[Eximo]

I'm going to make myself sound very old, but I remember the DDR spec generally doubled with every edition (DDR1, 2, 3, etc.). But since DDR4 we went from 2133MHz to 9000MHz. That's a significant change of pace and hopefully it will continue.

Well, if you are going to use 2133 for DDR4, then you have to use 4800 for DDR5.

Or you could do a comparison of mainstream speed of 3200/3600 vs 5600/6000/6400. Of course DDR5 hasn't run its course and we can expect 7200 or something becoming a norm for desktop ram (already seeing such speeds in embedded systems)

 

[jp7189]

I'm going to make myself sound very old, but I remember the DDR spec generally doubled with every edition (DDR1, 2, 3, etc.). But since DDR4 we went from 2133MHz to 9000MHz. That's a significant change of pace and hopefully it will continue.

Clocks are going up, but absolute latency has been worse so far this gen. Notwithstanding this OC attempt as 9000/36 is pretty good for CAS latency. The secondary timings aren't nearly as good though.

Generally speaking, absolute latency has stayed pretty flat to slightly improved over the years; DDR5 bucked that trend, but there's still hope it'll get back on track by the end of the gen.

 

[jp7189]

Excuse me, but did y

I've been playing with ai video upscaling (7950x3d @ 5.4ghz ccd0 & 5.9 GHz ccd1), rtx 4090, 64 gigs 7000mt ddr5 @ CL28-28-30-70 oc'd from 6000mt) and with a combination of memory, cpu and GPU overclocking I've got the upscale time down from 16 hours to 9... on a 4 min 1080p to 4k video using a Vulkan based scaler lol.

Sigh.. someday deep space 9 some day! I wonder if a ai accelerated CPU if I'd get gains assuming the tool supported it. I had my 4090 pegged at +350 core+1400 mem and it was chugging all the vram with a tile size of 400 so I had to lower it to 375 which left me 2.4 gigs of wiggle room so no crashing but it increased my final output time from 6.9 per frame to 9.1s. booooo hissssss! I'm g testing out fastencode on ffmpeg tonight on a 6 sec clip to see how the quality is because that's another reduction of like 30%. Also need to enable my igpu for mgpu and see if that lil boi can help any lol. Only an hour to go!

Excuse me, but did you say 9 hours for a 4 minute clip?! ..while using all 24GB on a 4090!? Yikes! I'm curious what tools you're using and what you're doing exactly. I don't do a lot of upscaling, but have been tinkering with AI transforms on video frames. Speed varies a lot, but normally I see frames per second, not seconds per frame.

 

[ThomasKinsley]

Well, if you are going to use 2133 for DDR4, then you have to use 4800 for DDR5.

Or you could do a comparison of mainstream speed of 3200/3600 vs 5600/6000/6400. Of course DDR5 hasn't run its course and we can expect 7200 or something becoming a norm for desktop ram (already seeing such speeds in embedded systems)

I see your point. 2133Mhz was the original "mainstream" for DDR4 back in 2015 or 16 with mobos and chips restricted to that speed (certainly obsolete now). But even comparing the current mainstream of 3600Mhz, we're seeing DDR5 speeds in excess of 7200Mhz. It appears that DDR5 has a lot more headroom left in it before we move on to DDR6.

 

[Eximo]

I see your point. 2133Mhz was the original "mainstream" for DDR4 back in 2015 or 16 with mobos and chips restricted to that speed (certainly obsolete now). But even comparing the current mainstream of 3600Mhz, we're seeing DDR5 speeds in excess of 7200Mhz. It appears that DDR5 has a lot more headroom left in it before we move on to DDR6.

Yes, though DDR4 4800 and even 5133 were sold. 4400 is where most enthusiasts hovered around.

DDR5 already has official specs for 8000, which I believe was their target at launch.

There is a problem though with going much faster with DIMM modules, and why DDR5 has built in ECC. Which is why we see the fastest speeds on embedded solutions and very much tweaked solutions like in this article.

Faster speeds may not be achievable with the distance between the memory slots and CPUs. I expect onboard DDR6 or HBM to start makings its way into CPU packages and for socketed memory to become another memory layer.

 

[Alpha_Lyrae]

So did AMD fix the mem controller issue? Do they use a new one? Did anybody report on that?! Are bioses still being restricted to 1.35V on these and is it safer now to overclock the ram?

The issue on chiplet Zen 4 parts wasn't memory voltage. It was SoC voltage, which was being set too high by motherboard auto settings when using EXPO and frying chips; SoC voltage, AFAIK, still has a hard 1.3v cap. An AGESA release a little while ago improved memory OC.

However, AMD's monolithic chips have always overclocked better when it comes to both FCLK and UCLK:MCLK. Chiplet parts have external wires going through package substrate for CCD to access IOD/mem controllers. In these monolithic APUs, it's all contained within the same die, so there's tighter integration of Infinity Fabric (no external wires and PHYs) and better signal quality too.

 

[TerryLaze]

The issue on chiplet Zen 4 parts wasn't memory voltage. It was SoC voltage, which was being set too high by motherboard auto settings when using EXPO and frying chips; SoC voltage, AFAIK, still has a hard 1.3v cap. An AGESA release a little while ago improved memory OC.

However, AMD's monolithic chips have always overclocked better when it comes to both FCLK and UCLK:MCLK. Chiplet parts have external wires going through package substrate for CCD to access IOD/mem controllers. In these monolithic APUs, it's all contained within the same die, so there's tighter integration of Infinity Fabric (no external wires and PHYs) and better signal quality too.

Yes, as far as I know SoC includes the mem controller part.
So, APUs are build different.