- Mar 3, 2017
- 1,687
- 6,243
- 136
I don't expect StxH to be in the PS6. The drop on memory bandwidth will be too significant. I do think that the CPU side of StxH will be vastly better than 8 Zen2 cores in a pair of CCXs. 40CUs of RDNA3.5 should be notably better than 36CUs from 90% of RDNA2. There should be at least twice the L3 cache. There may be an iGPU cache. It'll support PCIe 5 for better SSD performance.Perf uplift would not be good enough. PS5 has 36 CUs.
Everyone else at the moment seems to have tunnel vision into the ST ES results when there isn't a MT performance ES result yetSo far the MT uplift doesn't look good for the desktop part, and it becomes clearer what the rumors about “prioritized ST performance” meant )
Look at what lenovo and asus are getting for their steamdeck competitors. The top end versions are pushing $800+. Take away the screen, the battery, and relax the minaturization while throwing out the controllers and you can easily get below the bill of materials for a generic console with an APU that costs over twice what the Z1 extreme costs.The problem is that Strix Halo is more expensive, so this Valve Steam Box will have to cost a lot more than the current consoles, which are already heavily subsidized.
Would there be a big enough market for such a console?
Current LPDDR5 specs are getting better at 9600 MT/s even without the full benefit of full fat DDR5 latencies.
Combined with the benefits of CAMM for future smol systems it should be enough to do OK with 40 CUs, especially if SH also has a little ♾ cache too.
Halo isn't a DT part, it's not socketable.My guess is that sometime after the release of Strix Halo, AMD is going to define a new socket that will have memory internally to it (making it larger than AM5), and it will probably ditch local DIMM memory alltogether.
I wasn't comparing It to handhelds but consoles.Look at what lenovo and asus are getting for their steamdeck competitors. The top end versions are pushing $800+. Take away the screen, the battery, and relax the minaturization while throwing out the controllers and you can easily get below the bill of materials for a generic console with an APU that costs over twice what the Z1 extreme costs.
It all depends on what clock can RDNA3.5 reach compared to RDNA3.40CUs of RDNA3.5 should be notably better than 36CUs from 90% of RDNA2.
Even 2000 points would mean 41% better score than desktop 12C R9 7900, still very high, but at least more believable score.Except in the vid that table says 2K and not 2400 for 8950HS (35W). 2400 is supposed 8950x.
Its still curious anyway. So 4 Zen5 + 8 Zen5c cores (thus not even 12 full fat Zen5 cores) provide performance equal to 83 percent of 16 full fat Zen5 cores in CB24 MT. I mean, you would expect 12 full cores to provide 3/4 of that performance. This is not even that and still performs better.
An unfortunate yet fascinating side-effect of overclocking a dis-order design, IPC increases dramatically but so does the BS (branch stalling).Unlike CPUs however, RGT IPC is susceptible to much more branch mispredicts
Beliving RGT is like thinking Fox news is goodMassive increase of IPC (Information per Clock) from RGT is the main surprise here.
Normally RGT IPC is pretty poor, 10 seconds of useful info in 10 mins video.
Unlike CPUs however, RGT IPC is susceptible to much more branch mispredicts, "not final BIOS", "last minute design changes", "I have been told", "I could be wrong", etc
Well I don't know if you're living in the US or not but fox is much better than the rest of the msm. They actually invite leftists on to debate while the left is actively censoring news an constantly pushing their own fake news. Can' even call a man a man and a woman a woman with them.Beliving RGT is like thinking Fox news is good
Massive increase of IPC (Information per Clock) from RGT is the main surprise here.
Normally RGT IPC is pretty poor, 10 seconds of useful info in 10 mins video.
Unlike CPUs however, RGT IPC is susceptible to much more branch mispredicts, "not final BIOS", "last minute design changes", "I have been told", "I could be wrong", etc
they claim its result at 35w, dont they? so it does score 41 percent better than previous gen 65w cpu, while not even having 12 of those new full fat cores like that one did? could zen5c cores be significantly more performant than regular zen4 ones?Even 2000 points would mean 41% better score than desktop 12C R9 7900, still very high, but at least more believable score.
If that Strix Point was <=55W, then I am also very skeptical about It being true, because It would mean almost 2x better efficiency(+92% perf/W)!
If 2400 is for Halo, then the difference would be only 20%, despite Halo having 33% more cores.
If they were tested at pretty similar power limits, then I can understand the small difference.
Umm... dude...... but fox is much better than the rest of the msm.
How is this massive ipc increase? ... How is this even possible, am I dreaming?
I've stopped watching msm since 2020 so your point is mootUmm... dude... don't prove the point about FoxNews viewers in your own post
It was a joke post that DisEnchantment madeI've stopped watching msm since 2020 so your point is moot
8950X has obviously similar Fmax as the mobile Zen 5 part that scored 3400pts.Then how do you explain 7950x w/ 2919 vs 8950x w/ 3500 ?
I see your point, albeit a bit off the mark 3400 vs 3500. Still according to rgt this is late ES B0 silicon, would you agree that is as close as it get to retail?8950X has obviously similar Fmax as the mobile Zen 5 part that scored 3400pts.
The score difference is a measly 3% between the two, so I think that Fmax for a desktop Zen 4 (top SKU) is likely 5.35Ghz. So yep, my guess is that the max clock we might see is around 5.35 to 5.4Ghz for AM5 top SKU. That is round 5-6% clock deficit versus Zen 4, not bad at all.I see your point, albeit a bit off the mark 3400 vs 3500. Still according to rgt this is late ES B0 silicon, would you agree that is as close as it get to retail?
At bit of topic, but how come that LPDRAM can get these high speeds and the memory controllers can handle them, while standard DDR cannot? I guess they have a lot higher latency but besides that?Agreed. But socket AM5 does not have the pins for 4 LPDDR5 channels. Which is why there is not going to be a socket AM5 Strix Halo.
As you mention, LPDDR5 is moving on from 8533 to 9600 speeds, leaving regular DDR far behind. The whole idea of buying your own DRAM is slowly becoming meaningless.
The only reason to have a socket will be down to being able to pair your pick of motherboard and CPU, and perhaps upgrade in the future.
My guess is that sometime after the release of Strix Halo, AMD is going to define a new socket that will have memory internally to it (making it larger than AM5), and it will probably ditch local DIMM memory alltogether.
The downside might be a spam of SKUs (as @adroc_thurston keeps reminding us) but probably a better ecosystem. With cheaper motherboards and faster memories.
Trace length , and the loss of the physical connection to the dimm.At bit of topic, but how come that LPDRAM can get these high speeds and the memory controllers can handle them, while standard DDR cannot? I guess they have a lot higher latency but besides that?
If by desktop you mean DIY, sure, but SFFPCs are insanely popular and nearly all of AMD’s mobile offerings have appeared in one. Strix Halo won’t be an exception.Halo was never monolithic. Strix halo is not coming to desktop ever, unless AMD caves in to the demand of all 10 people who'd actually pay premium price for this chip on AM5.
RDNA 3.x vs RDNA 2 does have an uplift. Also, faster memory/IO will help.Perf uplift would not be good enough. PS5 has 36 CUs.
To be fair, AMD pushed Zen 4 well past its sweet spot. Take a 7950X/X3D and limit it to 65/88W and it will sing.Its 35W ES of 4 Zen5+8 Zen5C with 16 CU design.
35W ES, supposedly.
IF this is true its running circles around Zen 4 in terms of perf/watt.
Not addressing the video I didn’t watch, but as to your claim that much higher scores for nT workloads “is impossible “, that is not true.Seriously?
4*Zen5+8*Zen5c 8950HS(>>35W) ES: 130/2400 pts
12*Zen4 R9 7900(65/88W): 107/1419 pts
Result: 21.5% better ST and 69% better MT.
I can still accept ST gain, but that MT is impossible considering It has the same amount of cores.
I love the 'last minute design changes' caveat. So like before a year ago - when the design was complete (whatever it is these days). You just found out something changed over a year ago - and you are pretending it just happened.Unlike CPUs however, RGT IPC is susceptible to much more branch mispredicts, "not final BIOS", "last minute design changes", "I have been told", "I could be wrong", etc
Mendocino is still not released in any mini pc form, you can't be so sure. Mini PC manufacturers only adopt the mass market chips.Strix Halo won’t be an exception