popular
Minisforum Refurb EliteMini AI370 mini PC: Ryzen AI 9 HX370, 32GB LPDDR5X-7500, 1TB Gen4 SSD, USB4, Win11H @ $670.65 + F/S
$671
$789
14% off
Minisforum
MinisforumDeal Details
Link: https://refurbished.min
Coupon: 15% OFF Via LTD15 (Make sure to replace the default applied Coupon Code with this one during Checkout)
Spec:
Coupon: 15% OFF Via LTD15 (Make sure to replace the default applied Coupon Code with this one during Checkout)
Spec:
- Windows 11 Home
- Ryzen AI 9 HX 370 12C/24T 2 GHz (5.1 GHz Boost, 36MB Cache) with 50 TOPS NPU
- AMD Radeon 890M Integrated Graphics (16 CU) @ 2900 MHz
- 32GB LPDDR5X 7500 Memory
- 1TB M.2 2280 PCIe Gen 4x4 NVMe SSD + Empty M.2 2280 Gen4 Slot
- 1x M.2 2230 Slot (populated with NIC)
- Wi-Fi 6E 2x2 + Bluetooth 5.3
- Ports:
- 1x USB4 Type-C 40 Gbps (DP 1.4, 5V 3A Power out)
- 4x USB 3.2 Gen 2 Type-A 10 Gbps (2 Back, 2 Front)
- 1x DisplayPort 2.0
- 1x HDMI 2.1
- 2x 2.5G RJ-45 Lan
- 1x 3.5mm Audio Combo Jack
- 1x Clear CMOS Hole
- 1x Power Button
- 1x Power Port (DC 19V)
- 5.12*5*1.85 inch
- 120W AC Adapter
- 1.04 lbs.
Community Notes
About the Poster
Deal Details
Community Notes
About the Poster
Link: https://refurbished.min
Coupon: 15% OFF Via LTD15 (Make sure to replace the default applied Coupon Code with this one during Checkout)
Spec:
Coupon: 15% OFF Via LTD15 (Make sure to replace the default applied Coupon Code with this one during Checkout)
Spec:
- Windows 11 Home
- Ryzen AI 9 HX 370 12C/24T 2 GHz (5.1 GHz Boost, 36MB Cache) with 50 TOPS NPU
- AMD Radeon 890M Integrated Graphics (16 CU) @ 2900 MHz
- 32GB LPDDR5X 7500 Memory
- 1TB M.2 2280 PCIe Gen 4x4 NVMe SSD + Empty M.2 2280 Gen4 Slot
- 1x M.2 2230 Slot (populated with NIC)
- Wi-Fi 6E 2x2 + Bluetooth 5.3
- Ports:
- 1x USB4 Type-C 40 Gbps (DP 1.4, 5V 3A Power out)
- 4x USB 3.2 Gen 2 Type-A 10 Gbps (2 Back, 2 Front)
- 1x DisplayPort 2.0
- 1x HDMI 2.1
- 2x 2.5G RJ-45 Lan
- 1x 3.5mm Audio Combo Jack
- 1x Clear CMOS Hole
- 1x Power Button
- 1x Power Port (DC 19V)
- 5.12*5*1.85 inch
- 120W AC Adapter
- 1.04 lbs.
Community Voting
Deal
Score
+10
Good Deal
Bad Deal
Join The Conversation
Share your experience with the Slickdeals community
Share information with the community. Please follow our Community Guidelines and be kind!
13 Comments
Sign up for a Slickdeals account to remove this ad.
I also dont think wifi 7 is really any benefit tbh yet. Most people dont have wifi 7 routers let alone have maxed out the capabilities of wifi 6 or even 6e yet.
Sign up for a Slickdeals account to remove this ad.
https://slickdeals.net/share/iphone_app/t/18534001
You want the 395+ Max to run LLMs on its unified memory like platform. It's a little deceptive on AMDs part they use the same naming schema for products that have completely different capabilities.
This also has "unified memory." That's literally how all iGPUs work since they were invented. The AI hype era has reduced tech literacy to new lows, lmao
Unified Memory Architecture is the ability for the GPU and CPU to address the same pool of memory and iGPUs have been doing this for a very long time. I suppose it is technically distinct from the concept of shared physical memory but functionally it has little impact on the operation of games or most programs. AMD and Intel iGPUs/"APUs" have had this ability for well over a decade.
Moreover, both the 370 and 395 have the same memory architecture so the OP is incorrect. Some 370 builds even have soldered RAM but the issue is the dual channel bandwidth which is the real bottleneck.
Perhaps that theoretically a game could avoid unnecessary movements to VRAM from system RAM under this architecture but no games on PCs, to the best of my knowledge, actually operate in this fashion. And they must be specifically programmed to do so, which is probably incredibly unlikely. Only video games on consoles conceivably work this way because of their extremely homogenous hardware and strong incentive to do so.
The only reason Apple's "Unified memory" offers so much more performance, is ironically, due to all the efforts they put in the physical design of the memory and the interconnection with the CPU/GPU to offer as much bandwidth as possible which dramatically accelerates memory bound tasks. The only way you could exceed this is to put a bunch of HBM next to the CPU/GPU. And then there are plenty of other very advanced techniques that are far beyond my understanding that Apple used to make their processors so fast and efficient, not just the huge memory bandwidth.
In short, the OP going on about the "unified memory" of the 395 being different is incorrect and that unified memory access has been implemented for decades. What separates the 395 from the 370, other than the huge iGPU is really the quad channel bandwidth and to some degree the soldered RAM.
No they don't. Sharing same physical memory for system and integrated graphics is not UMA or Unified Memory Architecture, it's quite a bit complex than that.
Unified Memory Architecture is the ability for the GPU and CPU to address the same pool of memory and iGPUs have been doing this for a very long time. I suppose it is technically distinct from the concept of shared physical memory but functionally it has little impact on the operation of games or most programs. AMD and Intel iGPUs/"APUs" have had this ability for well over a decade.
Moreover, both the 370 and 395 have the same memory architecture so the OP is incorrect. Some 370 builds even have soldered RAM but the issue is the dual channel bandwidth which is the real bottleneck.
Perhaps that theoretically a game could avoid unnecessary movements to VRAM from system RAM under this architecture but no games on PCs, to the best of my knowledge, actually operate in this fashion. And they must be specifically programmed to do so, which is probably incredibly unlikely. Only video games on consoles conceivably work this way because of their extremely homogenous hardware and strong incentive to do so.
The only reason Apple's "Unified memory" offers so much more performance, is ironically, due to all the efforts they put in the physical design of the memory and the interconnection with the CPU/GPU to offer as much bandwidth as possible which dramatically accelerates memory bound tasks. The only way you could exceed this is to put a bunch of HBM next to the CPU/GPU. And then there are plenty of other very advanced techniques that are far beyond my understanding that Apple used to make their processors so fast and efficient, not just the huge memory bandwidth.
In short, the OP going on about the "unified memory" of the 395 being different is incorrect and that unified memory access has been implemented for decades. What separates the 395 from the 370, other than the huge iGPU is really the quad channel bandwidth and to some degree the soldered RAM.
Now the term Strix Halo is called Unified memory Access, not Unified memory Architecture. What it does is enables the CPU and the iGPU to access any location of the memory from the single Address Space which is 1st in x86 Processor, introduced by Strix Halo.
In any other Intel or AMD Consumer Processors with integrated Graphics, when a specified amount is allocated to the iGPU, it creates a separate address space, accessible to the GPU only. When any data is needs to be loaded to the GPU Buffer, 1st it has to be loaded to the system memory currently available, and then it gets copied to the GPU allocated memory and the GPU memory address space gets updated. After that only, the GPU can start processing the data. Once the processing is done, it again needs to come from the shared GPU memory to the system memory and then gets stored in the storage. Because of that huge number of to and from memory operations, it bottlenecks the performance as memory operations waste 100s of thousands CPU and GPU Clock Cycles. On top of that, it is an inefficient way of memory allocation because if say 8GB is allocated to the GPU and system has 24GB of memory and currently only using 10GB, the raste can't be used by the iGPU, even if that is free.
You can literally see that by looking at the task monitor if you load a big LLM model to the iGPU where first system memory will be filled with it and then it will be start filling the shared GPU memory.
In Unified memory Access, once the data is loaded and directed to the CPU or iGPU, they can directly pick it, without any separate address space required by CPU and iGPU. Currently Windows still does not support this feature fully and that's why even in Strix Halo, people need to allocate a fixed amount to it.
I am sure you probably thought since integrated Graphics in Windows PCs are called "UMA" Graphics for decades, it must be Unified memory Access since the beginning of integrated graphics in CPU. And I don't think you also understood how apple has achieved better performance either and thrown some mumbo jumbo from here and there. Apple uses Unified memory Access, since its SOC, and OS and Frameworks, all are designed by them, it uses an extremely optimized Close to metal Software-hardware combination for running its instruction set, the main strong feature of any RISC processor, where the OS/Compiler should have very low-level view of the actual architecture and can do a lot of heavy lifting and generate optimize code paths. In hardware front, they simply uses Unified memory Access + Higher memory Bus + Higher speed LPDDR5X Memory to offer a high memory bandwidth for the whole package, not some magical optimization.
An another thing, since here UMA is implemented, for both Apple Silicon and AMD Strix Halo, between Heterogeneous Compute Units, CPU and iGPU and NPU, this is also called hUMA. UMA normally refers to classical Multi-Processor or MultiCore systems where the whole package of the similar Processing Cores see the memory as same.
Join The Conversation
Share your experience with the Slickdeals community
Share information with the community. Please follow our Community Guidelines and be kind!