GAA works wonders in decreasing gate leakage at these scaled channel lengths, but we'll have to wait until Samsung can get their yields up enough to ship these Socs in much higher volumes across the world.
SF3 took way too long to hit the market, and when it did it is in the form of Exynos W1000. SF2 is more of an iteration on top of SF3 and is clearly showing great potential with Exynos 2600.
The real successor to SF3 is SF2P, which sounds a little bit like Intel 18A-P that's supposed to improve clocks over normal 18A, but we shall see how that turns out.
I did. I said practical benchmarks like we insist for any other chi or platform. Did you read my comment? No one is satisfied with just timespy extreme to test GPUs so why would I only just GET Geekbench for CPUs?
Attacking me personally doesn't make the argument superior. Notice I attack only Geekbench itself.
The GPU scores are interesting as well.
Looking at the perf/pwr curve in Nomad Light, at 4W similar in score to the A19 Pro, and just a touch below the 8EG5 and D9500.
At 6W similar to the other android SoCs leading over Apple.
At 8W and above starts to lead above 8EG5 and hangs with D9500.
3D Mark Steel Nomad Light GPU
SoC
Score (wattage found in power curve)
E2600 @ 980Mhz
3231 (\~10W)
8EliteGen5 @ 1200Mhz
3147 (\~12W)
D9500 @ 1700Mhz
3402 (\~10W)
A19 Pro @ 1620Mhz
3001 (\~14W)
Laptop GPUs
Adreno X2
4331
Radeon 780M
2960
Radeon 890M
3519
Arc 140T
3613
Xe-LPG MTL 128EU
3530
Arc 140V
3358
M4
3987
3D Mark Solar Bay Extreme
SoC
Score
E2600
2050
8EliteGen5
1339
D9500
2537
A19 Pro
2355
Laptop GPUs
Radeon 890M
2255
Arc 140T
2402
Xe-LPG MTL 128EU
1861
Arc 140V
2512
M4
2957
If only Samsung actually optimizes their products. Maybe not kill OpenCL/GL ES replaced with ANGLE on Vulkan. Also maybe expose MGFX4 so that the folks at MESA could bring it over on RADV instead of some in-house derived AMDVLK (that AMD already dopped) like what they do with Turnip for Adreno.
RDNA4 is a pretty good architecture all things considered, it doesn't compete with Nvidia Blackwell favorably but that's to be expected. It is still funny to me that the first RDNA4 integrated graphics isn't in a chip from AMD but instead a phone Soc.
Why would you think RDNA4 doesn't clock well. The 9060xt that I had clocks to 2650MHz at its voltage floor, 0.75V. If Samsung spends a few more months to get their Fmax sorted out for these incredibly efficient GAA nodes, I don't see a reason why they can't get away with 25% less WGP clocking at least 33% higher.
When they get the Fmax sorted out, iso freq voltage also tend to drop for the higher VF points, thus yielding them more efficiency at the same frequency/throughput.
RDNA4 on TSMC N4P (possibly using HP libraries) clocks efficiently to 2800MHz in my experience, and is among the highest clocking GPUs to date at higher voltages.
Exynos 2600 had the area planned out for 16CUs of RDNA4, there really is no point clocking it too high when it'll be living in a phone, even if there's that HPB thermal architecture helping with hot spots.
Samsung probably has a much better node this time around
Prior to 7nm, Samsung was the fab leader for decades. Most people weren't paying attention to fabs as they are today thus its largely lost to history. There is no reason why they can't return to form.
Last year Xiaomi's Xring 01 was surprisingly better than MediaTek's D9400. Historically, MediaTek's implementations tend to be weaker vs Qualcomm/Samsung/Huawei
So the Xring 03 in a few months will be an interesting comparison point
It should explain if it's an implementation difference again, or a node difference
Having 9 middle cores seems to help, along with the node improvement. After the strong headstart, it does seem to fizzle out quicklier than 8EG5 though, especially after the 10W mark, probably because of lower clock speed throughout the board. Still, good improvement for Exynos.
No-Draw-3565@reddit (OP)
2600 is significant more efficient and stronger than 8eG5
sustained TDP of 18W for 2600 sustained TDP of 21W for 8 elite gen 5
EmptyVolition242@reddit
Almost double the perf at 2 Watts. Crazy. This chips must have insane idle scores.
Slava_Tr@reddit
There are about 2800, not 2200, according to this source, which is \~40% more - still a great result
Aleblanco1987@reddit
wow, that's great at low wattages
bunihe@reddit
GAA works wonders in decreasing gate leakage at these scaled channel lengths, but we'll have to wait until Samsung can get their yields up enough to ship these Socs in much higher volumes across the world.
SF3 took way too long to hit the market, and when it did it is in the form of Exynos W1000. SF2 is more of an iteration on top of SF3 and is clearly showing great potential with Exynos 2600.
The real successor to SF3 is SF2P, which sounds a little bit like Intel 18A-P that's supposed to improve clocks over normal 18A, but we shall see how that turns out.
Hour_Firefighter_707@reddit
Interesting stuff. That's a big difference
marknor1983@reddit
i wonder how many devices will use this soc
i think it is just the fold and the big S
ResponsibleJudge3172@reddit
The importance of Geekbench on this sub has exploded in recent years
DerpSenpai@reddit
Because it correlates very nicely to real workloads and it's ISA independent and OS too
dagmx@reddit
Notice that these doubters will never actually suggest a better alternative beyond very specific cherry picked examples
DerpSenpai@reddit
Gaming brain rot. Where their 500-600$ CPU is destroyed by a phone feels wrong so it must be fake
ResponsibleJudge3172@reddit
You know, purely from geekbench, Rocketlake 11th gen from Intel should be faster than Zen 3 for example.
dagmx@reddit
Yeah and then they’ll inevitably point to the gaming performance not holding up and ignore that a lot of that is down to the graphics side of things.
ResponsibleJudge3172@reddit
I did. I said practical benchmarks like we insist for any other chi or platform. Did you read my comment? No one is satisfied with just timespy extreme to test GPUs so why would I only just GET Geekbench for CPUs?
Attacking me personally doesn't make the argument superior. Notice I attack only Geekbench itself.
ResponsibleJudge3172@reddit
I genuinely need proof in the form of app benchmarks with the expected 50% IPC deficit vs Intel among others
DerpSenpai@reddit
go check SPEC? In SPEC, the smartphone chip from 2024, the 8 Elite. Is faster than Panther Lake and Zen 5
ResponsibleJudge3172@reddit
OK, where can I find this SPECINT and SPECFP score? Any Google search so far circles back to Geekbench
Noble00_@reddit
The GPU scores are interesting as well.
Looking at the perf/pwr curve in Nomad Light, at 4W similar in score to the A19 Pro, and just a touch below the 8EG5 and D9500.
At 6W similar to the other android SoCs leading over Apple.
At 8W and above starts to lead above 8EG5 and hangs with D9500.
3D Mark Steel Nomad Light GPU
3D Mark Solar Bay Extreme
If only Samsung actually optimizes their products. Maybe not kill OpenCL/GL ES replaced with ANGLE on Vulkan. Also maybe expose MGFX4 so that the folks at MESA could bring it over on RADV instead of some in-house derived AMDVLK (that AMD already dopped) like what they do with Turnip for Adreno.
bunihe@reddit
RDNA4 is a pretty good architecture all things considered, it doesn't compete with Nvidia Blackwell favorably but that's to be expected. It is still funny to me that the first RDNA4 integrated graphics isn't in a chip from AMD but instead a phone Soc.
DerpSenpai@reddit
It doesn't compete all that well here though, they have to be wider and clocked slower for the same performance vs Qualcomm and ARM
bunihe@reddit
Why would you think RDNA4 doesn't clock well. The 9060xt that I had clocks to 2650MHz at its voltage floor, 0.75V. If Samsung spends a few more months to get their Fmax sorted out for these incredibly efficient GAA nodes, I don't see a reason why they can't get away with 25% less WGP clocking at least 33% higher.
DerpSenpai@reddit
I mean in power. At 10W it has to be wider and clocked lower to have the same performance. ARM Mali doing that at 1700Mhz is seriously impressive IMO
bunihe@reddit
When they get the Fmax sorted out, iso freq voltage also tend to drop for the higher VF points, thus yielding them more efficiency at the same frequency/throughput.
RDNA4 on TSMC N4P (possibly using HP libraries) clocks efficiently to 2800MHz in my experience, and is among the highest clocking GPUs to date at higher voltages.
Exynos 2600 had the area planned out for 16CUs of RDNA4, there really is no point clocking it too high when it'll be living in a phone, even if there's that HPB thermal architecture helping with hot spots.
No-Draw-3565@reddit (OP)
the drivers are unoptimised for Xclipse 960 a few uodates will put the score in similar to 9500 also g1 Uktra in 9500 is very unstable
BlueSwordM@reddit
Oh my, that is quite the difference.
It also obliterates the D9400/D9500, which is not surprising.
andreif@reddit
It's literally got 2 more cores than the competition, I would not read into process node too much because of that.
ML7777777@reddit
Prior to 7nm, Samsung was the fab leader for decades. Most people weren't paying attention to fabs as they are today thus its largely lost to history. There is no reason why they can't return to form.
Vince789@reddit
Last year Xiaomi's Xring 01 was surprisingly better than MediaTek's D9400. Historically, MediaTek's implementations tend to be weaker vs Qualcomm/Samsung/Huawei
So the Xring 03 in a few months will be an interesting comparison point
It should explain if it's an implementation difference again, or a node difference
BlueSwordM@reddit
Yep. Even the XRing O1 was on par/slightly better than the D9500.
Front_Expression_367@reddit
Having 9 middle cores seems to help, along with the node improvement. After the strong headstart, it does seem to fizzle out quicklier than 8EG5 though, especially after the 10W mark, probably because of lower clock speed throughout the board. Still, good improvement for Exynos.
DerpSenpai@reddit
Give me this chip in the Tab S12 please. It will be insanely good to play PC games on it
floodassistant@reddit
Hello, your submission has been removed for the following reason:
Please do not flood the subreddit with posts
^(Click here to message the moderators if you have any questions or concerns)