Filter GPU choices by chassis bay width and validate quantities

src/derive/validate.test.ts

@@ -244,24 +244,32 @@ describe('validatePlan GPUs and vendor availability', () => {
         modelId: serverModelId,
         count: 8,
         uplink: '2x25G',
+
         gpu: { modelId, perNode },
       },
     ]
     return plan
   }
 
   it('accepts a GPU within the per-node limit', () => {
-    const issues = validatePlan(planWithGpu('gpu-rtx-6000-ada', 1, 'server-microcloud-x13'))
+    const issues = validatePlan(planWithGpu('gpu-rtx-6000-ada', 1, 'server-superserver-tn12'))
     expect(issues.filter((i) => i.message.includes('GPU'))).toEqual([])
   })
 
-  it('rejects more GPUs than a node accepts', () => {
-    const issues = validatePlan(planWithGpu('gpu-rtx-6000-ada', 2, 'server-microcloud-x13'))
-    expect(issues.some((i) => i.severity === 'error' && i.message.includes('1 GPU per node'))).toBe(
-      true,
+  it('rejects a GPU too wide for the node', () => {
+    const issues = validatePlan(planWithGpu('gpu-h100-pcie', 1, 'server-microcloud-x13'))
+    expect(issues.map((i) => i.message)).toContain(
+      'H100 PCIe is a double-width GPU, but SYS-531MC-H8TNR takes single-width cards.',
     )
   })
 
+  it('rejects more GPUs than a node accepts', () => {
+    const issues = validatePlan(planWithGpu('gpu-rtx-6000-ada', 3, 'server-superserver-tn12'))
+    expect(
+      issues.some((i) => i.severity === 'error' && i.message.includes('2 GPUs per node')),
+    ).toBe(true)
+  })
+
   it('rejects GPUs on a server model that takes none', () => {
     const issues = validatePlan(planWithGpu('gpu-rtx-6000-ada', 1, 'server-bigtwin-x11'))
     expect(issues.some((i) => i.severity === 'error' && i.message.includes('takes no GPUs'))).toBe(

src/derive/validate.ts

@@ -1,4 +1,4 @@
-import { catalog, itemLabel, portCount, type SwitchRole } from '../model/catalog'
+import { catalog, gpuFits, itemLabel, portCount, type SwitchRole } from '../model/catalog'
 import {
   mgmtDeviceCount,
   type Partition,
@@ -242,6 +242,15 @@ function checkGpu(issues: Issue[], scope: Scope, group: ServerGroup): void {
     )
     return
   }
+  if (server && !gpuFits(item, server)) {
+    report(
+      issues,
+      scope,
+      'error',
+      `${itemLabel(item.id)} is a ${item.gpuWidth}-width GPU, but ` +
+        `${itemLabel(group.modelId)} takes ${server.gpuWidth}-width cards.`,
+    )
+  }
   if (gpu.perNode > capacity) {
     report(
       issues,

src/model/catalog.test.ts

@@ -149,12 +149,20 @@ describe('category invariants', () => {
     }
   })
 
-  it('gives every GPU a power draw and a part number', () => {
+  it('gives every GPU a power draw, a part number and a slot width', () => {
     const gpus = entries.filter(([, i]) => i.category === 'gpu')
     expect(gpus.length).toBeGreaterThan(0)
     for (const [id, item] of gpus) {
       expect(item.powerWatts, `${id}.powerWatts`).toBeGreaterThan(0)
       expect(item.partNumber, `${id}.partNumber`).toBeDefined()
+      expect(item.gpuWidth, `${id}.gpuWidth`).toBeDefined()
+    }
+  })
+
+  it('gives every GPU-capable server a bay width', () => {
+    for (const [id, item] of entries) {
+      if (item.category !== 'server' || !item.gpuCapable) continue
+      expect(item.gpuWidth, `${id}.gpuWidth`).toBeDefined()
     }
   })
 
@@ -185,7 +193,7 @@ describe('reachability', () => {
       ...serversForUsage('worker'),
       ...serversForUsage('management'),
       ...serversForUsage('storage'),
-      ...gpusForServer('server-microcloud-x13'),
+      ...gpusForServer('server-superserver-tn12'),
     ].map((i) => i.id),
   )
 
@@ -295,8 +303,10 @@ describe('helpers', () => {
     }
   })
 
-  it('offers GPUs only for GPU-capable servers', () => {
-    expect(gpusForServer('server-microcloud-x13').length).toBeGreaterThan(0)
+  it('offers GPUs only for GPU-capable servers, filtered by bay width', () => {
+    expect(gpusForServer('server-superserver-tn12').length).toBeGreaterThan(0)
+    // The MicroCloud bays are single-width; no cataloged GPU fits.
+    expect(gpusForServer('server-microcloud-x13')).toEqual([])
     expect(gpusForServer('server-bigtwin-x11')).toEqual([])
     expect(gpusForServer('nope')).toEqual([])
   })

src/model/catalog.ts

@@ -67,6 +67,8 @@ export type ServerUsage = 'worker' | 'management' | 'storage'
 
 export type PortSpeed = '1G' | '10G' | '25G' | '100G'
 
+export type GpuWidth = 'single' | 'double'
+
 /** Platform tier a NOS support license is priced by. The vendors name the
  *  tiers differently — Edgecore sells "40/100G" and "10/25G", Broadcom
  *  "100G" and "10G" — so the license table maps this per NOS. */
@@ -105,6 +107,8 @@ export interface CatalogItem {
   serverUsages?: ServerUsage[]
   /** Max GPUs a single server node of this model accepts; absent = none. */
   gpuCapable?: number
+  /** Card width on GPUs; maximum bay width on servers. */
+  gpuWidth?: GpuWidth
   /** No plan can select this item, so it never reaches a BOM: it is kept
    *  because the official compatibility list names it. Say why. Every
    *  otherwise-unreachable entry must set this (catalog.test.ts), so that
@@ -212,6 +216,7 @@ export const catalog: Record<string, CatalogItem> = {
     status: 'stable',
     serverUsages: ['worker'],
     gpuCapable: 1,
+    gpuWidth: 'single',
   },
   'server-microcloud-h13': {
     id: 'server-microcloud-h13',
@@ -226,6 +231,7 @@ export const catalog: Record<string, CatalogItem> = {
     status: 'stable',
     serverUsages: ['worker'],
     gpuCapable: 1,
+    gpuWidth: 'single',
   },
   'server-bigtwin-x11': {
     id: 'server-bigtwin-x11',
@@ -280,6 +286,8 @@ export const catalog: Record<string, CatalogItem> = {
     nodesPerChassis: 1,
     status: 'stable',
     serverUsages: ['worker', 'storage'],
+    gpuCapable: 2,
+    gpuWidth: 'double',
   },
   'server-superserver-tr12p': {
     id: 'server-superserver-tr12p',
@@ -409,6 +417,7 @@ export const catalog: Record<string, CatalogItem> = {
     description: '48 GB GDDR6 ECC, PCIe 4.0 ×16, dual slot',
     powerWatts: 300,
     status: 'stable',
+    gpuWidth: 'double',
   },
   'gpu-h100-pcie': {
     id: 'gpu-h100-pcie',
@@ -419,6 +428,7 @@ export const catalog: Record<string, CatalogItem> = {
     description: '80 GB HBM2e, PCIe 5.0 ×16, dual slot',
     powerWatts: 350,
     status: 'stable',
+    gpuWidth: 'double',
   },
 
   // --- Transceivers ---
@@ -646,11 +656,20 @@ export function serversForUsage(usage: ServerUsage): CatalogItem[] {
     .sort(byAvailability)
 }
 
-/** GPU models offerable for a server model — empty unless it accepts any. */
+/** Whether a GPU card physically fits a server node's GPU bay: a
+ *  double-width bay takes any card, a single-width bay only single-width
+ *  cards. */
+export function gpuFits(gpu: CatalogItem, server: CatalogItem): boolean {
+  return server.gpuWidth === 'double' || gpu.gpuWidth === 'single'
+}
+
+/** GPU models offerable for a server model — empty unless it accepts any,
+ *  and only cards that fit its GPU bay width. */
 export function gpusForServer(serverModelId: string): CatalogItem[] {
-  if (!catalog[serverModelId]?.gpuCapable) return []
+  const server = catalog[serverModelId]
+  if (!server?.gpuCapable) return []
   return Object.values(catalog)
-    .filter((i) => i.category === 'gpu')
+    .filter((i) => i.category === 'gpu' && gpuFits(i, server))
     .sort(byAvailability)
 }
 

src/views/plan/RackSection.tsx

@@ -55,6 +55,12 @@ function ServerGroupRow({
     value: i.id,
     label: optionLabel(i),
   }))
+  if (group.gpu && !gpuOptions.some((option) => option.value === group.gpu?.modelId)) {
+    gpuOptions.push({
+      value: group.gpu.modelId,
+      label: `${itemLabel(group.gpu.modelId)} (incompatible)`,
+    })
+  }
   const perNode = Math.min(group.gpu?.perNode ?? 1, catalog[group.modelId]?.gpuCapable ?? 1)
 
   return (
@@ -137,6 +143,16 @@ function ServerGroupRow({
           onChange={(v) => patch({ gpu: v ? { modelId: v, perNode: perNode } : undefined })}
         />
       )}
+      {group.gpu && (catalog[group.modelId]?.gpuCapable ?? 0) > 1 && (
+        <NumberField
+          label="GPUs per node"
+          value={group.gpu.perNode}
+          min={1}
+          onChange={(v) =>
+            group.gpu && patch({ gpu: { modelId: group.gpu.modelId, perNode: v || 1 } })
+          }
+        />
+      )}
       <button
         onClick={() => removeServerGroup(partition.id, rack.id, group.id)}
         className="btn-secondary justify-self-start"