AMD Execution Thread [2022]

[BRiT]

The potent M1 Ultra chips are targeting devices *starting* at $4k. How large is that market?

It's Infinite as there's no top cap...

 

[Wesker]

It's Infinite as there's no top cap...

Ssshh! That's the secret to InfinityFabric!

 

[pTmdfx]

Wish granted early next year.

Big APU is an ancient tale that will never be. Both Kaveri GDDR5M and Greenland failed us. I can't believe you.

 

[xpea]

Big APU is an ancient tale that will never be. Both Kaveri GDDR5M and Greenland failed us. I can't believe you.

APUs, by nature, will never match dGPUs performance, so the only question is where the dGPU low-end will fall

 

[swaaye]

See consoles.

 

[Newguy]

What happened to CVML in Rembrandt? I can't believe AMD still does not take AI seriously. Even Intel got matrix acceleration now.

This is going to bite them in the butt real bad in the future.

https://hothardware.com/news/amd-reveals-ryzen-7000-dragon-range-and-phoenix-mobile-cpus

"In addition to the financial disclosures, AMD also revealed some information regarding future products. In her prepared remarks, AMD CEO Dr. Lisa Su mentions future processors with integrated Xilinx technology. “As one example, we are integrating Xilinx’s differentiated AI engine across our CPU product portfolio to enable industry-leading inference capabilities,” Dr. Su said, “with the first products expected in 2023.”"

Zen 4/5?

 

[Silent_Buddha]

https://hothardware.com/news/amd-reveals-ryzen-7000-dragon-range-and-phoenix-mobile-cpus

"In addition to the financial disclosures, AMD also revealed some information regarding future products. In her prepared remarks, AMD CEO Dr. Lisa Su mentions future processors with integrated Xilinx technology. “As one example, we are integrating Xilinx’s differentiated AI engine across our CPU product portfolio to enable industry-leading inference capabilities,” Dr. Su said, “with the first products expected in 2023.”"

Zen 4/5?

It's interesting that AMD appears to be taking the opposite approach of NV by putting hardware AI assist into the CPU rather than the GPU. Granted, NV don't currently have that option as they don't have an x86 license. If Windows on Arm takes off and there aren't future exclusivity clauses as they had with Qualcomm, I wonder if NV will potentially move in that direction?

It's an interesting way to go about it. Currently consumer CPUs are much smaller WRT die size compared to GPUs so there's more potential for adding things there without hitting the reticle limit of a given fab.

Also, considering that AMD's CPUs are already chiplet based, I wonder if this will just be an AI chiplet? If, so then there is also the potential for a future graphics product from AMD featuring the same or a similar hypothetical AI chiplet.

Regards,
SB

 

[Kaotik]

It's interesting that AMD appears to be taking the opposite approach of NV by putting hardware AI assist into the CPU rather than the GPU. Granted, NV don't currently have that option as they don't have an x86 license. If Windows on Arm takes off and there aren't future exclusivity clauses as they had with Qualcomm, I wonder if NV will potentially move in that direction?

It's an interesting way to go about it. Currently consumer CPUs are much smaller WRT die size compared to GPUs so there's more potential for adding things there without hitting the reticle limit of a given fab.

Also, considering that AMD's CPUs are already chiplet based, I wonder if this will just be an AI chiplet? If, so then there is also the potential for a future graphics product from AMD featuring the same or a similar hypothetical AI chiplet.

Regards,
SB

No they're not. They have matrix crunching hardware (this so called "hardware AI assist") in CDNA-line GPUs. This is more akin to Intel AI accelerators in their CPUs.

 

[Silent_Buddha]

Frontier supercomputer debuts as world’s fastest, breaking exascale barrier | ORNL
ORNL’s Frontier First to Break the Exaflop Ceiling | TOP500

However, a recent development to the Frontier system has allowed the machine to surpass the 1 exaflop barrier. With an exact HPL score of 1.102 Exaflop/s, Frontier is not only the most powerful supercomputer to ever exist – it’s also the first true exascale machine.

The top position was previously held for two years straight by the Fugaku system at the RIKEN Center for Computational Science (R-CCS) in Kobe, Japan. Sticking with its previous HPL benchmark score of 442 PFlop/s, Fugaku has now dropped to No. 2. Considering the fact that Fugaku’s theoretical peak is above the 1 exaflop barrier, there’s cause to also call this system an exascale machine as well. However, Frontier is the only system able to demonstrate this on the HPL benchmark test.

The world's first benchmarked exascale supercomputer hit 1.102 Exaflops on the TOP500 and has a theoretical peak of 2 Exaflops. The 2nd place Fugaku does have a theoretical peak above 1 Exaflops but scored just 442 PFlops, so it's still the first theoretical peak exascale supercomputer.

It uses Epyc processors combined with Instinct MI250X accelerators.

Regards,
SB

 

[Kaotik]

Frontier supercomputer debuts as world’s fastest, breaking exascale barrier | ORNL
ORNL’s Frontier First to Break the Exaflop Ceiling | TOP500



The world's first benchmarked exascale supercomputer hit 1.102 Exaflops on the TOP500 and has a theoretical peak of 2 Exaflops. The 2nd place Fugaku does have a theoretical peak above 1 Exaflops but scored just 442 PFlops, so it's still the first theoretical peak exascale supercomputer.

It uses Epyc processors combined with Instinct MI250X accelerators.

Regards,
SB

Also LUMI, based on same HPE Cray nodes as Frontier, got third spot in top500, green500 and HPCG lists, worth promise for notable improvements especially in HPCG results by November update to the lists

 

[pcchen]

The world's first benchmarked exascale supercomputer hit 1.102 Exaflops on the TOP500 and has a theoretical peak of 2 Exaflops. The 2nd place Fugaku does have a theoretical peak above 1 Exaflops but scored just 442 PFlops, so it's still the first theoretical peak exascale supercomputer.

Frontier's peak is actually ~ 1.68 Exaflops. I guess they just round up to 2 exaflops in the press release to make it "simpler." Fugaku's peak is similarly ~ 0.53 Exaflops.
One interesting info in the press release is the HPL-AI results, which uses mixed precision units to achieve full precision results. Frontier managed to get 6.88 Exaflops in HPL-AI, while Fugaku did 2 Exaflops.

 

[Kaotik]

Ryzen 7000 launch event (unveiling, not actual sales date) 29th August

AMD to Host Livestream Event to Unveil Next Generation Ryzen Processors

Browse AMD’s company-wide and financial press releases.

ir.amd.com

 

[digitalwanderer]

I can't get excited, DDR5 still costs way too much.

 

[Malo]

Good time to buy a Zen3 for by X370 board though

 

[Putas]

Good time to buy a Zen3 for by X370 board though

Or B350 board. What a value after five years.

 

[Kaotik]

We actually did a video on that

It's in finnish, but to cut it short, using RTX 3080 Ti at FullHD high:
Old BIOS:
Ryzen 7 1800X: CBr20 3732, Witcher 3 99.5 FPS, Cyberpunk 2077 58 FPS
New BIOS:
Ryzen 7 1800X: CBr20 3610, Witcher 3 107.8 FPS, Cyberpunk 2077 64 FPS
Ryzen 7 5800X3D: CBr20 5587, Witcher 3 201 FPS, Cyberpunk 2077 112 FPS

(for CBr20, old BIOS had consumption at 230W and CPU clocking 3,7 GHz, new 205W and CPU clocking 3,6 GHz)

edit: for clarification, both CPUs used DDR4-3200 at same settings, so there's easily more room for 5800X3D

 

[DavidGraham]

The prices of current AMD GPUs are crashing way harder than NVIDIA GPUs, seems the demand for AMD GPUs is much lower than the demand of RTX GPUs.

Earlier this month, AMD stated their Graphics devision suffered a decline in Q2, and will sustain another decline in Q3.

Gaming graphics declined in the quarter as macro conditions impacted discretionary spending.

While we expect the gaming graphics market to be down in the third quarter, we remain focused on executing our GPU roadmap, including launching our high-end RDNA three GPUs later this year.

GPU Price Crash Is Making It Hard For AIBs To Offload AMD Radeon Graphics Cards Too, RX 6700 XT Drops Below $400 US, RX 6600 Below $260 US

It's not just NVIDIA that has been hurt by the GPU price crash but AIBs are having a difficult time with AMD Radeon RX 6000 graphics cards.

wccftech.com

Should be interesting to see the Q2 marketshare numbers this quarter, will AMD fall below 17% (as per their Q1 share?) or manage to sustain/exceed it?

 

[PSman1700]

Good this is happening, gpu prices have been too high for awhile now. Ihvs have no choice but lowering msrp. Things are much more affordable now but still too high.

 

[Kaotik]

Should be interesting to see the Q2 marketshare numbers this quarter, will AMD fall below 17% (as per their Q1 share?) or manage to sustain/exceed it?

Probably exceed. Their quarters don't line up perfectly but NVs GeForce division crashed 44% (expected)last Q

 

[pharma]

AMD 400G Adaptive Exotic SmartNIC Architecture at HC34

The new AMD 400G Adaptive "Exotic" SmartNIC architecture was detailed at HC34 for the PCIe Gen5 and CXL 2.0 era

www.servethehome.com