It's a failure because it secured 2 (two, dos, zwei) design wins on portable devices, of which only one is an actual work laptop.
All while lunar lake secured 50x more design wins.
So yes, unless there are more laptop design wins coming up that somehow missed the CES and Computex, Strix Halo is a massive flop. At least for the B2B Marketing teams.
N1X is
Yes you can.
View attachment 124410
View attachment 124411
View attachment 124412
Strix Halo at 15W matches Strix Point at 20W.
At 20W it beats all other APUs by 60-80%. And at 35W it's an unmatched beast.
Source:
100% this is the way, strix point not strong enough but Halo is king
- 1CCD 8core Halo with 3d cache
- RDNA4 32/40 cores
- pegged at 20-30w mobile
- dock mode for full 100w power
will be the ULTIMATE unbeatable handheld, fluid 1080p for everything
tis unfortunate medusa halo will take like 1.5 years more to release, too long
AMD should really aim for a mid release targeting handheld/mobile market but also full power thinkpads, no reason to buy any other laptop really Halo gives everything
One Chip to rule them all!
The 16C version wouldnt male sense for a handeld but there s also a 8C part with 32 CUs IIRC, this one would be releveant for such an usage.You'd be crippling the chip by limiting it to 20-30W. What would be the point? It wouldn't be useful as a handheld.
The 16C version wouldnt male sense for a handeld but there s also a 8C part with 32 CUs IIRC, this one would be releveant for such an usage.
Trying to run 8 cores and 32 CU's with 20-30W sounds like a bad time, or rather, a waste of hardware. Eight cores with 16 CU's might be doable.
It would be more efficient with more CUs.
That s the other way around, at same throughput 32 CUs are roughly 4x more efficient than 16 CUs, at same power the 32 CUs APU would have 40% better FPS and 8C/16T is way enough for a GPU that small, just look at desktops where 8C are enough for more than 64 CUs, so a handheld need no more than 4C-6C/4-6T.Trying to run 8 cores and 32 CU's with 20-30W sounds like a bad time, or rather, a waste of hardware. Eight cores with 16 CU's might be doable.
That s the other way around, at same throughput 32 CUs are roughly 4x more efficient than 16 CUs, at same power the 32 CUs APU would have 40% better FPS and 8C/16T is way enough for a GPU that small, just look at desktops where 8C are enough for more than 64 CUs, so a handheld need no more than 4C-6C/4-6T.
Sorry, that s 2x the efficency at same power.I know 8 cores are plenty. What I would like to know is where you are getting 32CU's being 4x more efficient than 16CU's. It's well known that more hardware at lower clock speeds is more efficient than less hardware and high clock speeds (looking at you 9070 XT) but 4x efficiency sounds unrealistic.
I doubt large integrated GPUs will happen unless it is a custom order for apple, valve, Microsoft etc.Anyway to increase perf/watt there s no other solution than increasing the CU count set apart using a more efficient process, so far a 8C + 32 CU Strix Halo would be perfect for a high perfs handeld, of cource it would cost quite more than the Strix Point based such devices, eventualy 1200$, but at the same time it would have a quite longer life cycle.
There shouldnt be a big price difference between a 12C/16CUs Strix Point and a 8C/32CUs Strix Halo, for a handeld 32GB are enough, so the equipement does not cost more, it s just that it will be 2x the bandwith with the same RAM quantity, overall the BOM difference is someting like 100-150$, it wouldnt even cost that much more than a Rog Ally with a Z2.I doubt large integrated GPUs will happen unless it is a custom order for apple, valve, Microsoft etc.
Maybe AI training/inferencing can change it
I doubt AMD on its own will be able to evangelize this. Maybe the steam O/S & xbox O/S can make things happen 🤔
Unless strix halo is able to clock cu and uncore separately, clocking it low would cripple its command processing, triangle rasterization, rops etc. One of the big advancements in gpus recently is how they can clock higher to boost parts of the core that do not scale with the cu count.
And GPU workloads are not as "embarrassingly parallel" as you would think. A GPU 2x the size does not have 2x fps, even with the same clocks.
lmao.There shouldnt be a big price difference between a 12C/16CUs Strix Point and a 8C/32CUs Strix Halo
That's cataclysmic for that segment.overall the BOM difference is someting like 100-150$
That would amount to 250-300$ more at the retail level, barely 35% higher price for 40% higher perfs at same power and battery life, there s nothing cataclysmic here,lmao.
That's cataclysmic for that segment.
There are 4 different things hereI do wonder if Strix Halo like products will become the core of follow on generation gaming consoles. Given how low their margins are, and how expensive development and production is, it's going to have to have usefulness in other markets to make sense. I get that the halo products are expensive, but with committed volume buys, Sony and MS could get more favorable pricing on them.
Strix Halo will not be used in any seriously marketable (read: sellable) handheld. Switch 2 probably uses 5W-10W in undocked mode, its SOC probably costs around $30 to mass produce, and it will probably sell much more units than all Zen 5 CPU types combined.There are 4 different things here
- nintendo — will fab console from a (samsung) process created a decade back
- Sony — reasonable specs for the budget. This will be 100% custom with AMD's help
- Valve — release valve o/s will run on halo type SoCs (subject purely to market demand)
- MS — will make a $1000 console plus release xbox o/s copying valve o/s. Nvidia might release a premium handheld. Surface team also could release a premium handheld (probably Intel). Asus, lenovo, Acer etc. wil use different SKUs provided by AMD to serve different markets
Strix Halo will not be used in any seriously marketable (read: sellable) handheld. Switch 2 probably uses 5W-10W in undocked mode, its SOC probably costs around $30 to mass produce, and it will probably sell much more units than all Zen 5 CPU types combined.
Everybody loves Strix Halo as a piece of tech, but not everybody is willing to pay for it. It ONLY makes sense in premium devices, period. End of story. Lenovo is pricing their new small desktop Strix Halo at $1600+. For that money, you can upgrade your existing rig to smash it all CPU and GPU workloads outside of possibly some LLMs that can make use of the unified RAM. You can probably make an entirely new build for $1600 that can still beat it handily.
Its a very niche product. If AMD would sell socketed 170W TDP versions for $600, sure, it would be super enticing and generate a ton of real interest, but thats not happening.
but it's a single data point, the difference in within margin of error, and the CLANG compilation results from the slide are showing 9950X with the lead I would expectCompiling llama.cpp
cmake -B build -S . -DGGML_HIP=ON -DAMDGPU_TARGETS="gfx1100" -DGGML_HIP_ROCWMMA_FATTN=ON -DGGML_CUDA_FA_ALL_QUANTS=ON -DGGML_HIP_GRAPHS=ON && time cmake --build build --config Release -j$(nproc)
STXH 60 watts 1m47s,
9950X 120 watts 1m50s,
STXH 40 watts 2m7s
60 watts STXH can beat 120 watts 9950X.
While I overall agree with you, switch is a bad benchmark for HaloSwitch 2 probably uses 5W-10W in undocked mode, its SOC probably costs around $30 to mass produce, and it will probably sell much more units than all Zen 5 CPU types combined.
Thats actually an awesome comparison. I'd LOVE to see a similar comparison to say, an RX 6700XT in straight 1080p and 1440p gaming.View attachment 124528
Not sure what is the orginal source got it from here https://community.frame.work/t/request-verify-dgpu-support/69392/33
Still someone also tested code compilation in there
but it's a single data point, the difference in within margin of error, and the CLANG compilation results from the slide are showing 9950X with the lead I would expect
A lot more lmao.That would amount to 250-300$ more at the retail level